Paul-Antoine Tual: AI insights for business leaders

AI's Speculative Bubble: the Microsoft-Mistral Sovereign Pivot

Thu, 23 Jul 2026 00:00:00 GMT

By Paul-Antoine TUAL, AI Transformation Leader, Croissance et Transitions. July 2026.

Framing. A speculative cycle is never proven at the moment it is described: it is only confirmed after the fact. What can be documented, on the other hand, are its signals: index concentration, the gap between what is invested and what is earned, and what those who manage other people's money think, month after month. We already asked the question, differently, in Where is AI's value heading?, from the technology value-chain side. In an earlier article we followed AI's money through its four floors and its closed loop; in another, we assessed the real AI maturity of the CAC 40 and Nasdaq Top 10. Here, we take a higher view of the full cycle, anchoring it in the first major European industrial deal of the summer: the pivot signed on 21 July 2026 between Microsoft and French champion Mistral AI, with Nvidia chips behind it. This article is not investment advice; several figures are press estimates, non-audited projections, or talks unconfirmed by the parties, flagged as such.

1. The framework: where are we in the Minsky cycle?

The framework known as “Minsky's” takes its name from economist Hyman Minsky's financial-instability hypothesis, from which financial historian Charles Kindleberger drew a five-stage anatomy of bubbles (Manias, Panics, and Crashes, 1978). It describes the recurring mechanics of speculative cycles: displacement (a technological or economic shock that opens a new field of profit), boom (money flows in, credit loosens), euphoria (valuations decouple from fundamentals, everyone wants in), profit-taking (insiders start selling), then revulsion, sometimes brutal. Generative AI's displacement is dated without ambiguity: the launch of ChatGPT, at the end of 2022. The boom filled 2023 and 2024. The question stirring markets in July 2026 is not whether the cycle exists, it always does for any breakthrough technology, but which stage we are in, and whether the current stage is still boom or already euphoria.

Three families of signal help decide, provisionally: concentration and valuation in the markets, the sentiment of professional fund managers, and the gap between capital poured in and proof of profitability. All three point, in July 2026, in the same direction.

<figure style="margin:2rem 0;"> <svg viewBox="0 0 720 200" width="100%" height="auto" role="img" aria-labelledby="f1t f1d" style="max-width:100%;height:auto;font-family:system-ui,-apple-system,'Segoe UI',Roboto,sans-serif;"> <title id="f1t">The Minsky cycle in five phases</title> <desc id="f1d">Five boxes linked by arrows: displacement, boom, euphoria, profit-taking, revulsion; generative AI is placed in advanced euphoria as of mid-2026, ahead of any profit-taking.</desc> <text x="8" y="20" font-size="14" font-weight="700" fill="#0B2545">The Minsky cycle applied to generative AI</text> <rect x="8" y="40" width="128" height="130" rx="10" fill="#13315C"/> <rect x="152" y="40" width="128" height="130" rx="10" fill="#13315C"/> <rect x="296" y="40" width="128" height="130" rx="10" fill="#FF7A59"/> <rect x="440" y="40" width="128" height="130" rx="10" fill="#0B2545"/> <rect x="584" y="40" width="128" height="130" rx="10" fill="#FFFFFF" stroke="#E55F3B" stroke-width="2" stroke-dasharray="5,3"/> <text x="72" y="90" font-size="14" font-weight="700" fill="#FFFFFF" text-anchor="middle">Displacement</text> <text x="72" y="112" font-size="10.5" fill="#E2E8F0" text-anchor="middle">ChatGPT, 2022</text> <text x="216" y="90" font-size="14" font-weight="700" fill="#FFFFFF" text-anchor="middle">Boom</text> <text x="216" y="112" font-size="10.5" fill="#E2E8F0" text-anchor="middle">2023-2024</text> <text x="360" y="90" font-size="14" font-weight="700" fill="#0B2545" text-anchor="middle">Euphoria</text> <text x="360" y="112" font-size="10.5" fill="#0B2545" text-anchor="middle">2025, mid-2026</text> <text x="504" y="90" font-size="14" font-weight="700" fill="#FFFFFF" text-anchor="middle">Profit-taking</text> <text x="504" y="112" font-size="10.5" fill="#E2E8F0" text-anchor="middle">insiders selling</text> <text x="648" y="90" font-size="14" font-weight="700" fill="#475569" text-anchor="middle">Revulsion</text> <text x="648" y="112" font-size="10.5" fill="#475569" text-anchor="middle">not yet reached</text> <line x1="138" y1="105" x2="150" y2="105" stroke="#475569" stroke-width="2"/> <polygon points="146,100 146,110 152,105" fill="#475569"/> <line x1="282" y1="105" x2="294" y2="105" stroke="#475569" stroke-width="2"/> <polygon points="290,100 290,110 296,105" fill="#475569"/> <line x1="426" y1="105" x2="438" y2="105" stroke="#475569" stroke-width="2"/> <polygon points="434,100 434,110 440,105" fill="#475569"/> <line x1="570" y1="105" x2="582" y2="105" stroke="#475569" stroke-width="2"/> <polygon points="578,100 578,110 584,105" fill="#475569"/> </svg> <figcaption style="font-size:0.8rem;color:#475569;margin-top:0.5rem;line-height:1.5;">Figure 1. Five classic phases of a speculative cycle: the signals of July 2026 place generative AI in advanced euphoria, ahead of any profit-taking or revulsion. Source: Minsky, Kindleberger (1978), see section 1.</figcaption> </figure>

2. Overheating signals: concentration, valuation, sentiment

Concentration first. Stocks directly tied to the AI ecosystem now account for roughly 44 to 45% of the S&P 500's total market capitalisation, a basket tracked by JPMorgan that stood at 26% in 2022, and are estimated to have driven between 75% and 80% of the index's cumulative gain since ChatGPT's launch, depending on methodology [1]. This is not an isolated record: the Nasdaq shows the same signature, with a Top 10 that alone accounts for close to half the index, Nvidia leading at around $5 trillion. We detailed this ranking in our CAC 40 / Nasdaq study [2].

Valuation next. The S&P 500's Shiller CAPE ratio (cyclically adjusted price/earnings) has stayed above 40 continuously since May 2026, at 41.6 in May, 40.96 in June, 41.4 in July [3]. Only one precedent exists in the whole of modern stock-market history: the peak of 44.19 reached in December 1999, months before the dot-com bubble burst, followed by a 49% drop in the S&P 500 between March 2000 and October 2002 [3]. Being close to the 1999-2000 record does not mean 2026 will follow the same trajectory. Current leader multiples (around 25 times earnings) remain, in fact, below the 58 times seen in March 2000 [4], as we detail in our article on the closed loop. The absolute level of the CAPE, however, is unprecedented in a quarter of a century. The two measures reconcile in one sentence: if the whole index is as expensive as in 2000 while the leaders' multiples are not, it is because their current earnings have exploded. The debate is therefore no longer about the price paid for today's profits, but about the durability of those profits, part of which rests on circular flows internal to the sector (more on this below).

<figure style="margin:2rem 0;"> <svg viewBox="0 0 720 260" width="100%" height="auto" role="img" aria-labelledby="f2t f2d" style="max-width:100%;height:auto;font-family:system-ui,-apple-system,'Segoe UI',Roboto,sans-serif;"> <title id="f2t">S&P 500 Shiller CAPE ratio, May to July 2026</title> <desc id="f2d">Line at 41.6 in May, 40.96 in June and 41.4 in July 2026, staying below the historic peak of 44.19 reached in December 1999.</desc> <text x="8" y="20" font-size="14" font-weight="700" fill="#0B2545">Shiller CAPE: three months above 40, well below the 1999 peak</text> <line x1="60" y1="89" x2="680" y2="89" stroke="#E55F3B" stroke-width="2" stroke-dasharray="6,4"/> <text x="676" y="83" font-size="11" fill="#E55F3B" text-anchor="end">Dec. 1999 peak: 44.19</text> <path d="M180,130 L380,141 L580,134" fill="none" stroke="#0B2545" stroke-width="3"/> <circle cx="180" cy="130" r="5" fill="#0B2545"/> <circle cx="380" cy="141" r="5" fill="#0B2545"/> <circle cx="580" cy="134" r="5" fill="#0B2545"/> <text x="180" y="115" font-size="12" font-weight="700" fill="#0F172A" text-anchor="middle">41.6</text> <text x="380" y="126" font-size="12" font-weight="700" fill="#0F172A" text-anchor="middle">40.96</text> <text x="580" y="119" font-size="12" font-weight="700" fill="#0F172A" text-anchor="middle">41.4</text> <text x="180" y="238" font-size="11" fill="#475569" text-anchor="middle">May</text> <text x="380" y="238" font-size="11" fill="#475569" text-anchor="middle">June</text> <text x="580" y="238" font-size="11" fill="#475569" text-anchor="middle">July</text> </svg> <figcaption style="font-size:0.8rem;color:#475569;margin-top:0.5rem;line-height:1.5;">Figure 2. The CAPE has stayed stable above 40 since May 2026, still below the peak of 44.19 reached in December 1999. Source: see note [3].</figcaption> </figure>

OpenAI illustrates the tension. The company was valued at $852 billion following a $122 billion funding round closed on 31 March 2026, led by SoftBank alongside Amazon, Nvidia and a16z [5]. Nvidia, whose chips run OpenAI, thereby subscribes to the equity of its own customer: the circular-financing pattern described in our previous article, at work even within this record round. At the same time, internal financial documents leaked to the press and verified by the Financial Times show, for full-year 2025, $13.07 billion in revenue against a $20.92 billion operating loss, or roughly $1.60 lost for every dollar of revenue booked: the company spends about $2.60 to earn one [6]. This is a different figure from the $25 billion ARR and the projected $14 billion loss for 2026 that we cited in our article on the value chain [7]: the two do not contradict each other, they measure different periods, 2025 actuals against a 2026 projection, but the shared trajectory is clear, spending runs well ahead of revenue.

Sentiment, finally: it is the freshest signal. Bank of America's monthly Global Fund Manager Survey, conducted 2-9 July 2026 among international fund managers and published on 14 July, marks a clear shift [8] (see Figure 3):

Indicator (BofA survey)	July 2026	Comparison
AI bubble named top “tail risk”	45%	28% in June, ahead of an inflation resurgence (26%)
Hyperscaler capex = most likely source of a systemic credit event	48%	ahead of private credit (34%)
Long position on semiconductors = “most crowded trade”	82%	record: 80% in June, 73% in May
Net tactical allocation to technology	18% overweight	versus 26% the previous month

<figure style="margin:2rem 0;"> <svg viewBox="0 0 720 250" width="100%" height="auto" role="img" aria-labelledby="f3t f3d" style="max-width:100%;height:auto;font-family:system-ui,-apple-system,'Segoe UI',Roboto,sans-serif;"> <title id="f3t">Four indicators from the BofA survey, July 2026</title> <desc id="f3d">Four indicators tightening within a month: AI bubble named top tail risk (45% versus 28%), hyperscaler capex as a systemic credit-shock risk (48% versus 34%), semiconductors as the most crowded trade (82% versus 80%), net tactical tech allocation pulling back (18% versus 26%).</desc> <text x="8" y="20" font-size="14" font-weight="700" fill="#0B2545">BofA survey, July 2026: four signals tightening</text> <rect x="70" y="130" width="45" height="90" fill="#0B2545"/> <rect x="135" y="164" width="45" height="56" fill="#FF7A59"/> <rect x="235" y="124" width="45" height="96" fill="#0B2545"/> <rect x="300" y="152" width="45" height="68" fill="#FF7A59"/> <rect x="400" y="56" width="45" height="164" fill="#0B2545"/> <rect x="465" y="60" width="45" height="160" fill="#FF7A59"/> <rect x="565" y="184" width="45" height="36" fill="#0B2545"/> <rect x="630" y="168" width="45" height="52" fill="#FF7A59"/> <text x="92" y="124" font-size="12" font-weight="700" fill="#0F172A" text-anchor="middle">45%</text> <text x="157" y="158" font-size="12" font-weight="700" fill="#0F172A" text-anchor="middle">28%</text> <text x="257" y="118" font-size="12" font-weight="700" fill="#0F172A" text-anchor="middle">48%</text> <text x="322" y="146" font-size="12" font-weight="700" fill="#0F172A" text-anchor="middle">34%</text> <text x="422" y="50" font-size="12" font-weight="700" fill="#0F172A" text-anchor="middle">82%</text> <text x="487" y="54" font-size="12" font-weight="700" fill="#0F172A" text-anchor="middle">80%</text> <text x="587" y="178" font-size="12" font-weight="700" fill="#0F172A" text-anchor="middle">18%</text> <text x="652" y="162" font-size="12" font-weight="700" fill="#0F172A" text-anchor="middle">26%</text> <text x="125" y="236" font-size="11" fill="#475569" text-anchor="middle">AI bubble</text> <text x="290" y="236" font-size="11" fill="#475569" text-anchor="middle">CapEx</text> <text x="455" y="236" font-size="11" fill="#475569" text-anchor="middle">Semiconductors</text> <text x="620" y="236" font-size="11" fill="#475569" text-anchor="middle">Tech allocation</text> </svg> <figcaption style="font-size:0.8rem;color:#475569;margin-top:0.5rem;line-height:1.5;">Figure 3. Within a month, the four indicators in the BofA survey tightened: AI bubble = risk No. 1, hyperscaler capex approaching private credit, semiconductor trade at its most crowded on record, tactical tech allocation pulling back. Navy = July 2026, coral = prior month. Source: see note [8].</figcaption> </figure>

This is not (yet) the revulsion phase in the Minsky sense: fund managers remain invested, they are not fleeing, but it is the typical signature of an advanced euphoria phase that knows itself to be euphoria, where consensus starts to turn on itself.

3. The CapEx/ROI mismatch: the bet keeps growing faster than the proof

The heart of the debate is not stock-market concentration, it is what funds it. In 2026 alone, the four large American hyperscalers (Microsoft, Amazon, Alphabet, Meta) are investing on the order of $660-725 billion in AI infrastructure (more still if Oracle is added), up from $410 billion in 2025 and $226 billion in 2024 on the same basis: spending that now absorbs roughly 94% of their net operating cash flow, against about three-quarters two years earlier [9]. Goldman Sachs, in its Tracking Trillions analysis, puts the global trajectory at $765 billion in 2026, rising to $1.6 trillion a year by 2031, for a cumulative $7.6 trillion over the period, split between compute ($5.1 trillion), data centres ($2.1 trillion) and energy ($358 billion) [10]. Morgan Stanley, in notes revised repeatedly throughout 2026, gives a comparable order of magnitude, around $800 billion for 2026 and $1.2-1.4 trillion for 2027-2028 depending on the version [11].

On the other side, proof of profitability is struggling to keep up, and the gap, far from closing, is widening. The most closely watched analyst on this question, Sequoia Capital's David Cahn, put a “revenue gap” of $200 billion in 2023 between the revenue needed to justify AI investment and actual revenue; he revised it to $600 billion in June 2024. By July 2026, his new estimate, reported by TechCrunch, puts that figure at roughly $3 trillion [12]. The gap has therefore not narrowed as the sector matured: it has been multiplied by five in two years.

On the corporate-user side, surveys converge on a modest profitability picture:

Source	Indicator	Result
Gartner (Jan., revised May 2026)	Expected global AI spending, 2026	≈ $2.5-2.59 trillion [13]
McKinsey, State of AI (2025, 1,993 respondents, 105 countries)	Share of organisations that are “AI high performers” (≥5% of EBIT attributed to AI)	6% [14]
Morgan Stanley, AI Market Trends 2026	S&P 500 companies citing an AI-linked benefit	21% [15]
Forrester, Predictions 2026	AI decision-makers reporting an EBITDA gain over 12 months	15% [16]
PwC, 29th Global CEO Survey (4,454 leaders, 95 countries)	CEOs with both higher revenue AND lower cost from AI	12% [17]
S&P Global	Companies that scrapped most of their AI initiatives in 2025	42% (versus 17% in 2024) [18]
Gartner (via Fiddler AI)	Agentic AI projects to be abandoned by 2027	> 40% [19]

The imbalance also shows up on the balance sheet. Bond issuance earmarked for financing data centres nearly doubled in 2025, to $182 billion against $92 billion in 2024 [20]. An analysis by Japanese business daily Nikkei, published in July 2026, puts the five largest hyperscalers' off-balance-sheet commitments (long-term leases, dedicated financing vehicles) at $1.65 trillion, some 22% more than the $1.35 trillion of debt they openly declare [21]. That funding is shifting this way, from cash flow toward debt and off-balance-sheet vehicles, is in itself a late-cycle signal: it is the point at which the boom stops being self-funded.

We have already examined this mismatch from another angle: our article on AI's value chain and closed loop details how part of this capital circulates among a small number of players (chipmakers, hyperscalers, labs) to the point that the same dollar can be counted as revenue at several links of the chain. This circular-financing mechanism is not repeated here in detail; it does, however, help explain why the CapEx/ROI gap described above can coexist, for a time, with strong revenue growth at infrastructure suppliers.

4. 21 July 2026: Microsoft and Mistral, with Nvidia's chip at the centre

It is against this backdrop that, on 21 July 2026, Microsoft and Mistral AI announced their deal [22][23]. The official statement, signed for Microsoft by Brad Smith and for Mistral by Arthur Mensch, stayed deliberately vague on amounts: “several billion dollars” over several years, with no precise figure or duration published, as Brad Smith explicitly confirmed to Reuters. Three strands structure the deal:

Compute capacity. Microsoft commits to reserving a significant share of the capacity of Mistral's future European data centres, where Mistral is deploying thousands of Nvidia chips from the brand-new Vera Rubin generation (a “Vera” CPU plus a “Rubin” GPU, in NVL72 configuration, that is 72 GPUs and 36 CPUs per rack).
Software integration. Microsoft integrates the Mistral Medium 3.5 and Mistral OCR 4 models into Microsoft Foundry and Copilot Studio, and makes them available on Azure, including in a hybrid environment (Azure Local) or fully disconnected from the network (Foundry Local), for regulated sectors (defence, healthcare, finance) that cannot send their data to a remote public cloud.
No dilution. The deal includes no new equity stake by Microsoft in Mistral: it is a capacity-purchase agreement and a joint go-to-market plan, not an equity investment.

This deal sits against a revealing geopolitical backdrop. In June 2026, the US administration temporarily suspended, on national-security grounds, international access to Claude Mythos and Claude Fable 5, Anthropic's most advanced models, an export-control measure lifted three weeks later, at the end of June, after a new safety mechanism was validated [24]. The episode was brief and already closed by the time this article was published; it nonetheless reminded European companies and governments that access to frontier AI conditioned on a foreign authorisation remains, by construction, reversible. An open-weight model, deployable offline, such as Mistral via Foundry Local, answers that risk directly: once the weights are downloaded, no remote authorisation can be revoked. The nuance to keep in mind is that the execution software layer (Azure Local, Foundry Local) is still published by Microsoft, a company under United States law subject to the CLOUD Act: the sovereignty gained applies to the model and the data, less to the stack that runs them. It is real progress, not full independence.

July's deal builds on a foundation already laid: as early as 30 March 2026, Mistral had secured a $830 million debt facility, led by Bpifrance alongside six other banks (BNP Paribas, Crédit Agricole CIB, HSBC, La Banque Postale, MUFG, Natixis CIB), to fund a sovereign data centre at Bruyères-le-Châtel: 13,800 Nvidia GB300 GPUs across 44MW, operated by Eclairion. Company-wide, Mistral states an ambition of 200MW by 2027 and 1GW by 2030, a trajectory, not capacity already in service: the power under contract today remains well short of that target [25].

5. A secondary angle: Mistral's shifting cap table. How far does sovereignty hold?

The 21 July deal is not an isolated event in Mistral's capital trajectory: it coincides with an ongoing negotiation that, if it closes, will meaningfully redraw its cap table. Two methodological caveats before the numbers: Mistral does not publish the detail of its shareholding structure, and the finest breakdown available comes from a single secondary source (24pm Academy's estimates), to be treated as an estimate, not as an established fact.

The starting point. Mistral's Series C, led by Dutch equipment maker ASML, closed on 9 September 2025: €1.7 billion raised at a post-money valuation of €11.7 billion (not €12 billion, as is sometimes reported); ASML invested €1.3 billion for around 11% of the capital, becoming the largest individual private shareholder in the company [26].

Today's snapshot (estimate). According to 24pm Academy, the capital would be split as follows:

Bloc	Estimated share	Composition
France	≈ 62.6%	Founders (Arthur Mensch, Guillaume Lample, Timothée Lacroix) ≈ 35.5% economic; French seed pool (Bpifrance, Xavier Niel and other early-stage investors) ≈ 27%
United States	≈ 22.8%	a16z (≈ 7-8%), Lightspeed, General Catalyst, Nvidia, Salesforce, Cisco
Non-French Europe	≈ 11.1%	Mainly ASML
Other international	≈ 3.5%	DST Global, Belfius, Exor, Mubadala

One point deserves to be highlighted, more solid than the plain economic snapshot: Mistral is structured as a French SAS with preferred shares carrying enhanced voting rights for the founders, meaning their real weight in decisions exceeds their 35.5% economic stake and secures them a controlling majority, independent of successive funding rounds [27]. It is this legal lock, more than the composition of the capital itself, that constitutes the true guarantee of sovereignty.

What is under discussion, and what is not yet: Mistral is reportedly negotiating a new round of several billion euros targeting a valuation of €20 billion, a re-rating of around 70% in ten months, with South Korea's Samsung in advanced talks to invest up to €1 billion, alongside the EQT Scaleup Europe fund, Danish investor Novo Holdings and Spanish bank Santander. This round is not signed: it was reported on 22 and 23 July 2026 by Axios and the Financial Times, and neither Mistral nor Samsung has confirmed the terms at this stage [28]. If this round closed at the reported amounts, simple arithmetic would put the non-French share in a range of roughly 45 to 50% of the capital, but this projection is not corroborated by any named source; it should be read as a plausible order of magnitude, not as a settled figure.

The reading that matters for a European business leader is therefore neither “Mistral is selling out” nor “Mistral remains 100% French”: both readings are wrong. It is more nuanced: raising tens of billions to keep pace with American hyperscalers requires international capital, and the guarantee of sovereignty now plays out on legal and governmental ground (voting rights, French law, the location of data and teams), not on the nationality of each euro invested.

6. Two scenarios for the next 12 to 24 months

Returning to the Minsky framework set out at the opening, two trajectories emerge for the AI ecosystem:

A Minsky-style adjustment: if the gap between infrastructure depreciation and usage revenue persists, hyperscalers could revise their spending plans downward, compressing valuation multiples across the whole semiconductor chain and forcing a consolidation of the startups most dependent on venture capital. The sector would then be left with a fleet of under-used data centres, whose book value would have to be written down, like the “dark fibre” laid in excess by the telecom industry in 2000, which sat unused for years. The difference that makes AI's case worse: fibre did not age, whereas a GPU becomes technologically obsolete within a few years, and the depreciated hardware could be obsolete before it is ever redeployed. This is the scenario feared by the 45% of managers in the BofA survey.

A productive absorption: if the rollout of next-generation chips such as Vera Rubin lowers the cost per token enough, the adoption of agentic AI in enterprise processes could accelerate to the point of gradually closing the revenue gap, stabilising the market on real cash flows rather than promises. This scenario has one precise condition, often left unsaid: the fall in unit price must trigger a more-than-proportional rise in volume (the “Jevons paradox”: cheaper, hence so much more used that total spending rises anyway). Failing that, the deflation in the cost of intelligence remains, as we wrote in the article on the value chain, a tax on those who sell tokens, and the gap widens instead of closing. This is the implicit bet behind the Microsoft-Mistral deal: that usage grows faster than the price falls.

The two scenarios do not entirely exclude each other: they may follow one another, or coexist floor by floor. We had already shown, in our article on the value chain, that the bottom of the chain (silicon, infrastructure) pockets very real cash while the top (models, applications) is still selling a promise. The question that will settle this is not ideological; it will show up, quarter after quarter, in the earnings releases of the technology giants.

What this means for a business leader (not an investor)

A leader of an SME or a mid-cap does not have to choose between buying or selling AI stocks. But this reading offers three concrete markers. The falling cost of intelligence (each new generation of chips, from Blackwell to Vera Rubin, mechanically lowers the cost per token) keeps working in favour of those who use AI rather than those who sell it: this is the moment to industrialise use cases, not to speculate on the cycle's peak. The Mistral episode, for its part, is a reminder that AI infrastructure that is deployable offline and under verifiable governance is no longer a theoretical luxury but an answer to a real risk, that of foreign access being reversed overnight. The caution shown by professional fund managers, finally (the BofA survey), is not a signal to ignore: it calls for budgeting AI as an investment that must prove its worth, not as a bet on the cycle continuing indefinitely. That is precisely the point of our white paper on AI industrialisation and the MATIA Method™.

Disclaimer. This article is educational and informational. It constitutes neither investment advice nor a recommendation to buy or sell any security whatsoever. Several figures cited are press estimates, non-audited projections, or talks unconfirmed by the parties, flagged as such in the text. Past performance is no guarantee of future performance. For any investment decision, consult a licensed investment adviser.

Want your company on the right side of this cycle, the one that puts AI to work while it becomes abundant, not the one betting on its peak? Start with a free AI Express Audit & Roadmap.

Sources

Figures as of the third week of July 2026. Several amounts are press estimates, non-audited projections, or reported-but-unconfirmed talks with the parties concerned, flagged as such in the text.

[1] Sherwood News (JPMorgan, AI basket ≈ 44% of the S&P 500); financial press (Moneycontrol, MSN) on AI's contribution to the index's cumulative gain since November 2022 (75-80% range depending on methodology).

[2] "CAC 40 and Nasdaq Top 10: size doesn't tell you AI maturity", market caps as of 20 July 2026 (CompaniesMarketCap).

[3] The Motley Fool (22 July 2026, Shiller CAPE at 41.4); GuruFocus, historical S&P 500 Shiller CAPE series (peak of 44.19 in December 1999).

[4] "The cash of AI: who pockets it, and why it may be circular", 2026 multiples (~25×) versus March 2000 (~58×).

[5] Bloomberg, CNBC, Forbes (31 March 2026), OpenAI funding round at $852 billion post-money, $122 billion raised, led by SoftBank with Amazon, Nvidia, a16z, D.E. Shaw Ventures.

[6] Fortune, MLQ.ai (June 2026), leaked OpenAI 2025 financial documents (via Ed Zitron), independently verified by the Financial Times: $13.07 billion in revenue, $20.92 billion operating loss (excluding a one-off non-cash charge tied to the October 2025 conversion to a public-benefit corporation).

[7] "The cash of AI" (source [4]), OpenAI ARR ≈ $25 billion (mid-2026 run-rate) and projected 2026 loss ≈ $14 billion (non-audited internal projection, The Information).

[8] Reuters (14 July 2026, via Yahoo Finance); Benzinga; detailed summary by Atrani Capital (Substack), Bank of America Global Fund Manager Survey, fieldwork 2-9 July 2026.

[9] Introl, CNBC (6 February 2026), ValueAddVC, 2026 hyperscaler capex (four-company scope excluding Oracle: $410 billion in 2025, $226 billion in 2024) and share of operating cash flow absorbed (Bank of America, ≈94%).

[10] Goldman Sachs, Tracking Trillions: The Assumptions Shaping the Scale of the AI Build-Out (goldmansachs.com/insights).

[11] Morgan Stanley notes on data-centre capex, revised repeatedly through 2026 (~$765-805 billion for 2026 per the May note; $1.1-1.4 trillion for 2027-2028 depending on the version), cite as an evolving order of magnitude, not a fixed figure.

[12] Sequoia Capital, "AI's $600B Question" (June 2024, historical value); TechCrunch, "Can AI answer the $3 trillion question?" (9 July 2026), trajectory of David Cahn's estimate: $200 billion (2023) → $600 billion (2024) → ≈$3 trillion (2026).

[13] Gartner, press release of 15 January 2026 (revised +47% on 19 May 2026), global AI spending 2026 ≈$2.5-2.59 trillion. ⚠️ This figure is sometimes misattributed to McKinsey elsewhere, it is a Gartner figure.

[14] McKinsey, The State of AI (survey fielded 25 June-29 July 2025, 1,993 respondents, 105 countries), 6% “AI high performers”.

[15] Morgan Stanley, AI Market Trends Institute 2026, 21% of S&P 500 companies cite an AI-linked benefit.

[16] Forrester, Predictions 2026 (28 October 2025), 15% of AI decision-makers report a 12-month EBITDA gain.

[17] PwC, 29th Global CEO Survey (January 2026, 4,454 leaders, 95 countries), 12% with both higher revenue and lower cost simultaneously.

[18] S&P Global, via CIO Dive, 42% of companies scrapped most of their AI projects during 2025 (versus 17% in 2024).

[19] Gartner via Fiddler AI, more than 40% of agentic AI projects will be abandoned by 2027 (a figure already cited in our white papers).

[20] CNBC (19 December 2025, S&P Global data), data-centre-linked bond issuance: $182 billion in 2025 versus $92 billion in 2024.

[21] Nikkei analysis (July 2026), picked up by Tom's Hardware and 247wallst, $1.65 trillion in off-balance-sheet commitments across the five largest hyperscalers (Alphabet, Amazon, Meta, Microsoft, Oracle), around 22% more than the $1.35 trillion of debt reported on their balance sheet.

[22] Official Microsoft release (news.microsoft.com, 21 July 2026), Microsoft-Mistral AI deal.

[23] France24, SiliconANGLE, Reuters (21-22 July 2026), press coverage of the deal, with Brad Smith's confirmation that no figure was publicly disclosed.

[24] Anthropic (anthropic.com/news); CNBC, CNN Business, Forbes, CSIS (June-July 2026), suspension and subsequent restoration of international access to Claude Mythos and Claude Fable 5 (export controls, 12-30 June 2026).

[25] Bloomberg, CNBC, SiliconANGLE, DataCenterDynamics (30 March 2026), Mistral's $830 million debt facility (consortium led by Bpifrance, 7 banks), Bruyères-le-Châtel data centre (13,800 Nvidia GB300 GPUs, 44MW, site operated by Eclairion).

[26] CNBC, ASML (official release), Sifted, Built In (9 September 2025), Mistral's Series C led by ASML, €11.7 billion post-money valuation, ASML investment of €1.3 billion (≈11% of capital).

[27] 24pm Academy, estimate of Mistral's shareholding structure (unofficial source, the only one to publish this level of detail); Clubic, preferred-share structure and founders' enhanced voting rights.

[28] Axios (22 July 2026); Financial Times via Silicon Republic; TechFundingNews; TechStartups (22-23 July 2026), ongoing discussions for a new Mistral round at a €20 billion valuation, Samsung/EQT Scaleup Europe/Novo Holdings/Santander; unconfirmed by the parties as of publication.

CAC 40 and Nasdaq Top 10: size does not tell you AI maturity

Mon, 20 Jul 2026 00:00:00 GMT

By Paul-Antoine TUAL, AI Transformation Leader, Croissance et Transitions, July 2026.

Framing. There is only one leaderboard for large companies: market capitalisation. Yet it measures one thing only, what the market is willing to pay today. We wanted to build a second one, measuring what those same companies really know how to do with artificial intelligence internally: their AI maturity. So we took the ten largest CAC 40 and Nasdaq companies and scored them from 1 to 5 across eight dimensions, on public evidence only. Then we set the two readings side by side, line by line. They have almost nothing in common: NVIDIA, the world's largest company and king of AI chips, is only 9th out of 10 on the Nasdaq's AI maturity; L'Oréal, only 2nd in the CAC 40 by size, is 1st by maturity; and not one of the twenty companies analysed reaches the grid's highest level. Market-cap figures are as of 20 July 2026; the maturity study is dated at its source. This article is not an investment note: it is a reading lesson.

First, AI maturity: the question market cap never asks

Let us start with the less familiar of the two criteria, because it carries this article's thesis.

A company's AI maturity is its ability to make artificial intelligence work internally, end to end: a funded and steered strategy, governance that answers for its decisions, data whose provenance is known, a technical platform that holds up in production, employees trained and involved, use cases genuinely deployed on core processes, mapped risks and compliance that holds. It is an internal execution capability, not a balance-sheet size.

Two confusions to clear up straight away, because they explain most of the surprises that follow:

Selling AI ≠ knowing how to govern it. A company can make the chips the planet runs on and still have, internally, no AI committee, no training plan and no governance of its own training data. That is exactly NVIDIA's case.
Communicating about AI ≠ being mature on AI. A claim is not proof: in the grid, an unsupported statement does not lift the score, and a proven gap between promise and product is paid for (Apple, a $250 million settlement over Siri in May 2026).

The result fits in one table. Here, for each index, is the podium by size and the podium by AI maturity, both criteria applied to the same ten companies.

Index	Podium by size	Podium by AI maturity
CAC 40	LVMH · L'Oréal · Hermès	L'Oréal · Sanofi · Schneider Electric
Nasdaq	NVIDIA · Apple · Alphabet	Alphabet · Meta · Amazon

On each side, only one name survives the change of criterion: L'Oréal in Paris, Alphabet in New York. Hermès, France's third-largest company, drops to 10th and last on maturity; NVIDIA, the world's largest, to 9th. Three figures complete the picture: the CAC 40 Top 10 averages 3.05/5, the Nasdaq 3.16/5 (both panels at the same "Structured" level), and not one of the twenty companies reaches the Transformative level [12].

This deserves to be stated plainly for an SME or mid-cap leader. The twenty largest companies in both indices are not, on AI, light-years ahead. Most are in the same place as many mid-sized companies: use cases running, governance that exists on paper, and proof that is missing. The gap is not where you think it is. The lead is won on proof.

The rest reads in three steps: the maturity grid, detailed dimension by dimension; the two indices put through that grid, each company read on both its score and its market cap; and finally what overlaying the two says about where the value is going.

The analysis grid, dimension by dimension

Eight dimensions, each with a weight. The weighting reflects a methodological conviction: a company with neither strategy nor governance is not doing AI, it is running experiments. That is why D1 and D2 together carry 30% of the score. At the other end, communication carries only 10%: it is a symptom, not a capability.

Dimension (weight)	The question asked	What lifts the score	What caps it
D1 · Strategy and value creation (15%)	Is AI a steered, funded, quantified strategic priority?	A published roadmap, an identified budget, a dated value target (BNP Paribas targets €750 million a year; Alphabet guides to $180-190 billion of capex)	AI everywhere in corporate communication but absent from the investor presentation (Safran, Airbus)
D2 · Governance and accountability (15%)	Who answers for AI, and how far up does the board go?	A dedicated committee that actually meets, a Chief AI Officer, a framework aligned to a recognised standard (NIST AI RMF, ISO/IEC 42001), board reporting; Schneider Electric and its Digital Committee (7 meetings a year)	A charter announced but never published, no anchoring at board level, or "trustworthy AI" principles aimed at customers rather than at oneself (NVIDIA)
D3 · Data (12%)	Do you know where the data comes from, and can you prove it?	A group-level data catalogue, quality indicators, traceability and lawfulness of training data	A platform claimed without a single public indicator (the general case), let alone active litigation over data provenance (NVIDIA, Meta)
D4 · Technology and industrialisation (12%)	Can you run AI in production, not just in a demo?	A documented MLOps platform, robustness testing, red teaming, incidents disclosed and fixed (Microsoft, Amazon)	No documented drift monitoring or rollback procedure; pilots that never scale
D5 · Talent, culture and organisation (12%)	Are employees trained and involved, and does labour dialogue exist?	A quantified training plan and a signed labour agreement: Safran (AI agreement of 22 January 2026 with CFDT, CFE-CGC and FO; 43,716 employees trained)	Training on display but labour dialogue absent or contested (Airbus/CFE-CGC, Schneider/CFDT), job cuts left undiscussed
D6 · Deployment and realised value (12%)	How many use cases in production on core processes, for what measured value?	Usage at scale on the core business, with published indicators (Alphabet: 900 million monthly users on Gemini)	Value announced but never third-party audited: the case for almost the entire panel
D7 · Risk, compliance and internal control (12%)	Are AI risks mapped, systems classified, the AI Act prepared for?	An AI risk factor formalised in regulated filings, an assumed public position on the AI Act, a system inventory	Generic risk management in which AI is not a standalone category; no risk map published three weeks before the 2 August 2026 deadline
D8 · Financial, extra-financial and stakeholder communication (10%)	Is AI in the equity story, with figures that hold?	AI indicators published every quarter (Alphabet, Amazon), including the costs owned up to (electricity consumption)	A gap between promise and product, penalised (Apple, $250 million over Siri), or total silence (Hermès: zero mention of AI in its FY2025 results)

<figure style="margin:2rem 0;"> <svg viewBox="0 0 720 360" width="100%" height="auto" role="img" aria-labelledby="f1t f1d" style="max-width:100%;height:auto;font-family:system-ui,-apple-system,'Segoe UI',Roboto,sans-serif;"> <title id="f1t">Weight of the eight dimensions in the AI maturity grid</title> <desc id="f1d">Horizontal bar chart: strategy and governance each carry 15% of the score, 30% combined, against only 10% for communication, the lowest-weighted dimension.</desc> <text x="8" y="22" font-size="14" font-weight="700" fill="#0B2545">The weight of the eight dimensions in the grid</text> <line x1="190" y1="40" x2="190" y2="336" stroke="#E2E8F0" stroke-width="1"/> <g fill="#FF7A59"> <rect x="190" y="48" width="480" height="22"/> <rect x="190" y="86" width="480" height="22"/> </g> <g fill="#0B2545"> <rect x="190" y="124" width="384" height="22"/> <rect x="190" y="162" width="384" height="22"/> <rect x="190" y="200" width="384" height="22"/> <rect x="190" y="238" width="384" height="22"/> <rect x="190" y="276" width="384" height="22"/> <rect x="190" y="314" width="320" height="22"/> </g> <g font-size="12" fill="#0F172A"> <text x="8" y="63">D1 · Strategy</text> <text x="8" y="101">D2 · Governance</text> <text x="8" y="139">D3 · Data</text> <text x="8" y="177">D4 · Technology</text> <text x="8" y="215">D5 · Talent</text> <text x="8" y="253">D6 · Deployment</text> <text x="8" y="291">D7 · Risk</text> <text x="8" y="329">D8 · Communication</text> </g> <g font-size="12" font-weight="700" fill="#0B2545"> <text x="678" y="63">15%</text> <text x="678" y="101">15%</text> </g> <g font-size="12" fill="#475569"> <text x="582" y="139">12%</text> <text x="582" y="177">12%</text> <text x="582" y="215">12%</text> <text x="582" y="253">12%</text> <text x="582" y="291">12%</text> <text x="518" y="329">10%</text> </g> </svg> <figcaption style="font-size:0.8rem;color:#475569;margin-top:0.5rem;line-height:1.5;">Figure 1. Strategy and governance together account for 30% of the score, against only 10% for communication. Source: MATIA AI maturity grid, data as of 20 July 2026.</figcaption> </figure>

The five levels, and the ceiling nobody has broken

The overall score is the weighted average of the eight scores. It reads on five levels:

Level	Score	What it describes	Count (out of 20)
Emerging	< 1.8	AI is still a watching brief, with no identified use	0
Exploratory	1.8-2.5	Early uses, nothing steered or governed	1 (Hermès, 2.15)
Structured	2.6-3.4	Governance exists, use cases are running, the proof is missing	13
Integrated	3.5-4.2	AI is in the core processes and in the accounts	6
Transformative	> 4.2	AI redefines the business model, governance keeps pace	0

This may be the single most important number in the study: not one of the twenty largest companies in either index reaches the level the grid describes as the end state. The panel's best score, Alphabet, tops out at 3.91, Integrated, not Transformative. Thirteen of the twenty remain at the Structured level, where a great deal has already been done but cannot yet be proven.

<figure style="margin:2rem 0;"> <svg viewBox="0 0 720 280" width="100%" height="auto" role="img" aria-labelledby="f2t f2d" style="max-width:100%;height:auto;font-family:system-ui,-apple-system,'Segoe UI',Roboto,sans-serif;"> <title id="f2t">Distribution of the twenty panel companies by AI maturity level</title> <desc id="f2d">Horizontal bar chart: thirteen out of twenty companies are at the Structured level, six at Integrated, one alone (Hermès) at Exploratory, and none reaches the Transformative level.</desc> <text x="8" y="22" font-size="14" font-weight="700" fill="#0B2545">Twenty companies, five levels, a ceiling never broken</text> <line x1="190" y1="40" x2="190" y2="256" stroke="#E2E8F0" stroke-width="1"/> <g fill="#E2E8F0"> <rect x="188" y="48" width="4" height="24"/> <rect x="188" y="216" width="4" height="24"/> </g> <rect x="190" y="90" width="34" height="24" fill="#FF7A59"/> <rect x="190" y="132" width="442" height="24" fill="#0B2545"/> <rect x="190" y="174" width="204" height="24" fill="#13315C"/> <g font-size="12" fill="#0F172A"> <text x="8" y="64">Emerging</text> <text x="8" y="106">Exploratory</text> <text x="8" y="148">Structured</text> <text x="8" y="190">Integrated</text> <text x="8" y="232">Transformative</text> </g> <text x="200" y="64" font-size="12" fill="#475569">0</text> <text x="232" y="106" font-size="12" font-weight="700" fill="#0B2545">1 (Hermès)</text> <text x="640" y="148" font-size="12" font-weight="700" fill="#0B2545">13</text> <text x="402" y="190" font-size="12" fill="#475569">6</text> <text x="200" y="232" font-size="12" font-weight="700" fill="#FF7A59">0</text> </svg> <figcaption style="font-size:0.8rem;color:#475569;margin-top:0.5rem;line-height:1.5;">Figure 2. Thirteen out of twenty companies remain at the Structured level, and none reaches the Transformative level: the ceiling nobody has broken. Source: MATIA AI maturity grid, data as of 20 July 2026.</figcaption> </figure>

The mechanism that makes the grid defensible: the evidence level

Beyond the number, every score carries an evidence level: documented (a cited public item establishes it), declarative (the company asserts it, nothing attests to it), mixed, or absent. Three rules follow from this, applied without exception to all twenty companies:

Any score ≥ 3 requires cited public documentary proof, with its source. No source, no 3.
Where the narrative and the proof diverge, the lower score is kept. That is what holds TotalEnergies at 2 on data despite very real technology partnerships, its CEO having himself described progress, in September 2025, as "bits and pieces".
Without proof, the score stays low rather than being guessed (1 or 2, evidence level "absent"). Never an estimate in place of a fact.

It is this mechanism, not an opinion, that produces the ranking reversals. It has one assumed side effect: it penalises discretion. A low score there means "no public proof", which is not quite "no initiative", a nuance worth keeping in mind for the two tables that follow.

What the grid does not measure

Method note. To be read before the rankings: the full methodology calls for a 4-to-6-week campaign combining internal document review, 12 to 20 targeted interviews (executive management, finance, IT/data, HR, legal, internal audit, business lines, board secretariat) and tool demonstrations. This edition mobilised no interviews and no internal documents: it relies solely on public sources (universal registration documents, annual and sustainability reports, 10-K/20-F filings, proxy statements, investor transcripts, specialist press), consulted on 13-14 July 2026 [12]. An assumed consequence: dimensions with high public exposure (D1, D2, D7, D8) are mechanically better documented than internal dimensions (D3 Data, D4 Technology, D5 Talent), regardless of these groups' real maturity. These results therefore measure AI maturity as publicly documented: a starting point for prioritising checks, not a verdict.

Why eight dimensions here, and nine in the MATIA Method™? The MATIA Scale™ we use in SME and mid-cap diagnostics has nine dimensions and five levels (from Spectateur to Pionnier); it feeds on interviews and internal access. Four of those dimensions, model sovereignty, resilience, technical debt and security score, simply cannot be read in a universal registration document: this "listed company" edition absorbs them into technology and risk. In return, it isolates two dimensions specific to the listed world, data and financial communication, precisely because the obligation to publish creates verifiable proof there. Same backbone, a lens adapted to what a listed company must make public.

The CAC 40 through the grid: what size does not predict

Here are the ten largest companies in the Paris index, ranked by their AI maturity, no longer by market cap, which sits alongside so the two can be compared line by line.

Two words to read the table. An index is a basket of stocks representative of a market: the CAC 40 groups 40 of the largest companies listed in Paris (Euronext). Market capitalisation is the share price multiplied by the number of shares, the "value" the market assigns at a given moment. Neither says anything about AI maturity: that is exactly what this table puts to the test.

Maturity rank	Company	AI maturity	Market cap	(Rank by size)	Sector
1	L'Oréal	3.64 · Integrated	≈ €202 billion	(2)	Cosmetics
2	Sanofi	3.51 · Integrated	≈ €92 billion	(10)	Health
3	Schneider Electric	3.40 · Structured	≈ €150 billion	(6)	Electrical equipment
4	Safran	3.12 · Structured	≈ €135 billion	(7)	Aerospace & defence
5	Air Liquide	3.10 · Structured	≈ €113 billion	(9)	Industrial gases
6	BNP Paribas	3.03 · Structured	≈ €113 billion	(8)	Banking
7=	LVMH	2.88 · Structured	≈ €240 billion	(1)	Luxury
7=	TotalEnergies	2.88 · Structured	≈ €158 billion	(4)	Energy
9	Airbus	2.76 · Structured	≈ €151 billion	(5)	Aerospace
10	Hermès	2.15 · Exploratory	≈ €176 billion	(3)	Luxury

Maturity scores: Croissance & Transitions in-house study "AI Maturity: CAC 40 Top 10", 13 July 2026, on public sources [12]. Market caps: CompaniesMarketCap as of 20 July 2026, converted from US dollars at the day's rate (€1 ≈ $1.14) and rounded [1][11]. Prices move daily; the L'Oréal-Hermès contest for 2nd place by size is tight and can flip depending on source and day.

Figure 3 below links each company to its two ranks, read in the same direction.

<figure style="margin:2rem 0;"> <svg viewBox="0 0 720 420" width="100%" height="auto" role="img" aria-labelledby="f3t f3d" style="max-width:100%;height:auto;font-family:system-ui,-apple-system,'Segoe UI',Roboto,sans-serif;"> <title id="f3t">Rank by size compared with rank by AI maturity, CAC 40 Top 7 by maturity</title> <desc id="f3d">Slope chart linking, for the CAC 40's seven largest companies by AI maturity, their rank by market cap to their rank by maturity: Sanofi moves from 10th to 2nd place, LVMH falls from 1st to 7th.</desc> <text x="8" y="20" font-size="14" font-weight="700" fill="#0B2545">Rank by size vs rank by AI maturity (CAC 40 Top 7)</text> <text x="80" y="42" font-size="11" fill="#475569">Rank by size</text> <text x="460" y="42" font-size="11" fill="#475569">Rank by AI maturity</text> <g stroke="#E2E8F0" stroke-width="2"> <line x1="200" y1="58" x2="200" y2="390"/> <line x1="520" y1="58" x2="520" y2="284"/> </g> <g font-size="11" fill="#475569" text-anchor="middle"> <text x="200" y="52">1</text> <text x="200" y="404">10</text> <text x="520" y="52">1</text> <text x="520" y="298">7</text> </g> <g stroke="#0B2545" stroke-width="2.5"> <line x1="200" y1="94" x2="520" y2="58"/> <line x1="200" y1="238" x2="520" y2="130"/> <line x1="200" y1="274" x2="520" y2="166"/> <line x1="200" y1="346" x2="520" y2="202"/> <line x1="200" y1="310" x2="520" y2="238"/> </g> <g stroke="#FF7A59" stroke-width="2.5"> <line x1="200" y1="382" x2="520" y2="94"/> <line x1="200" y1="58" x2="520" y2="274"/> </g> <g fill="#0B2545"> <circle cx="200" cy="94" r="5"/> <circle cx="520" cy="58" r="5"/> <circle cx="200" cy="238" r="5"/> <circle cx="520" cy="130" r="5"/> <circle cx="200" cy="274" r="5"/> <circle cx="520" cy="166" r="5"/> <circle cx="200" cy="346" r="5"/> <circle cx="520" cy="202" r="5"/> <circle cx="200" cy="310" r="5"/> <circle cx="520" cy="238" r="5"/> </g> <g fill="#FF7A59"> <circle cx="200" cy="382" r="5"/> <circle cx="520" cy="94" r="5"/> <circle cx="200" cy="58" r="5"/> <circle cx="520" cy="274" r="5"/> </g> <text x="532" y="62" font-size="12" fill="#0F172A" text-anchor="start">L'Oréal (1)</text> <text x="532" y="98" font-size="12" font-weight="700" fill="#0B2545" text-anchor="start">Sanofi (2)</text> <text x="532" y="134" font-size="12" fill="#0F172A" text-anchor="start">Schneider Electric (3)</text> <text x="532" y="170" font-size="12" fill="#0F172A" text-anchor="start">Safran (4)</text> <text x="532" y="206" font-size="12" fill="#0F172A" text-anchor="start">Air Liquide (5)</text> <text x="532" y="242" font-size="12" fill="#0F172A" text-anchor="start">BNP Paribas (6)</text> <text x="532" y="278" font-size="12" font-weight="700" fill="#0B2545" text-anchor="start">LVMH (7)</text> </svg> <figcaption style="font-size:0.8rem;color:#475569;margin-top:0.5rem;line-height:1.5;">Figure 3. Sanofi leaps from 10th to 2nd place, LVMH falls from 1st to 7th: size does not predict the maturity rank. Source: MATIA AI maturity grid, data as of 20 July 2026.</figcaption> </figure>

Read the two columns together: they contradict each other almost everywhere. L'Oréal, only 2nd by size, is 1st by AI maturity: a dedicated AI presentation at its AGM, a tech/AI budget above its R&D budget, a responsible-AI framework since 2021. Sanofi is the clearest demonstration: last of the Top 10 by market cap, at €92 billion against LVMH's €240 billion, it is second by maturity, calls itself an "AI-powered biopharma" and runs an AI tool inside a phase-3 clinical trial. Conversely, LVMH, the index's largest company, falls to 7th on maturity, and Hermès, third by size, comes dead last: at the maker of the silk scarf, AI is nearly absent from financial communication, a discretion consistent with its brand culture, but which, absent public proof, is paid for in the grid.

Sector predicts no better than size. Luxury and beauty (LVMH, L'Oréal, Hermès) together account for nearly 40% of the Top 10 by market cap, and occupy ranks 1, 7 and 10 on AI maturity. Three houses from the same universe, at opposite ends of the ranking. As for the index's most "technological" company, Schneider Electric, it comes only third, and that is electrification and automation, not software: no software, cloud or AI champion appears at the top of the CAC 40. Keep that in mind; it takes on its full meaning against the Nasdaq.

The LVMH lesson: three rankings, three different answers

LVMH deserves a pause, because it occupies three distinct ranks depending on what is measured: 1st by size, 3rd by index weight, 7th by AI maturity.

The second rank is explained by the free float, that is, the share of stock actually available to buy once the blocks durably held by a family, a founder or a state have been stripped out. In February 2026, the Arnault family crossed the 50.01% threshold of LVMH's capital (and nearly 66% of voting rights), according to a filing with the French financial markets authority (AMF) dated 25 February 2026 [4]. Barely half of LVMH's shares therefore "float" on the market. Since an index weighting is calculated on free-float market cap, LVMH weighs far less in the index than its raw size: as of 12 March 2026, the CAC 40 was led by Schneider Electric (8.25%) and TotalEnergies (8.11%), with LVMH coming only in third position (6.94%) [3].

Keep the triple distinction, it is the thread running through this whole article: size measures what the company is worth, weight measures how much of it the market can actually trade, and maturity measures what it can do with AI. None of the three follows from the other two.

The Nasdaq through the grid: the AI sellers are not the most mature

Same exercise on the other side of the Atlantic, and the same layout: AI maturity first, market cap alongside.

Maturity rank	Company	AI maturity	Market cap	(Rank by size)	Sector
1	Alphabet	3.91 · Integrated	≈ $4.3 trillion	(3)	Internet, cloud & AI
2	Meta	3.79 · Integrated	≈ $1.6 trillion	(7)	Social media & AI
3	Amazon	3.64 · Integrated	≈ $2.7 trillion	(5)	E-commerce & cloud
4	Microsoft	3.54 · Integrated	≈ $3.0 trillion	(4)	Software & cloud
5	Netflix	2.88 · Structured	out of the Top 10	(10 in Jan.)	Streaming
6	ASML	2.86 · Structured	out of the Top 10	(9 in Jan.)	Semiconductor equipment
7	Apple	2.78 · Structured	≈ $4.8 trillion	(2)	Consumer hardware
8	Tesla	2.76 · Structured	≈ $1.4 trillion	(8)	Automotive & robotics
9	NVIDIA	2.73 · Structured	≈ $5.0 trillion	(1)	Semiconductors / AI
10	Broadcom	2.71 · Structured	≈ $1.8 trillion	(6)	Semiconductors

Maturity scores: Croissance & Transitions in-house study "AI Maturity: Nasdaq Top 10", 14 July 2026, on public sources, January 2026 scope [12]. Market caps: CompaniesMarketCap as of 20 July 2026 [2]. Micron (≈ $1.0 trillion, having joined the trillion-dollar club in Q2 2026 [9]) and AMD (≈ $840 billion) have since overtaken ASML and Netflix in the Top 10 by size: as new entrants, they have not yet been scored. Also worth noting, SpaceX, listed on the Nasdaq on 12 June 2026 and valued at around $1.6 trillion, is listed without (yet) being a Nasdaq-100 constituent [2][10]: being listed on a venue is not the same as being a member of an index.

The first four lines tell a coherent story: these are AI users (Alphabet, Meta, Amazon, Microsoft), the only ones in the panel to reach the Integrated level. Then the ranking drops away, exactly where the market pays the most.

Apple and NVIDIA are fighting over the world's largest market cap, at around $5 trillion. NVIDIA was the first company in history to cross that mark, even passing $5.5 trillion in mid-May 2026 [5], and Apple briefly took back the top spot on 17 July 2026 [6]. These two sit at the 7th and 9th ranks of their own index's AI maturity. Perhaps the most telling comparison is this one: the ten largest Nasdaq stocks weigh about nine times the entire CAC 40 [2], for an average maturity score of 3.16 against 3.05. Nine times the market value, eleven hundredths of a point of maturity.

The finding: the pick-and-shovel sellers are the dunces of the class

Here is the most counter-intuitive result of the whole study. NVIDIA, the world's largest company and supplier of silicon to the entire industry, is only 9th out of 10 on the Nasdaq's AI maturity. Broadcom brings up the rear.

This is not an anomaly: it is exactly what the grid is built to reveal. It measures a company's ability to govern AI at home, not the quality of the AI it sells to its customers. Yet NVIDIA, at the date of the analysis, has no dedicated AI committee, no structured AI training plan for its tens of thousands of employees, and faces active litigation over training data [12]. Selling AI and knowing how to govern it are two different jobs, and the two are not correlated.

The mirror of this finding: Alphabet, only 3rd by size, is 1st by maturity (Gemini at 900 million monthly users, a top score in both strategy and deployment, a double full mark it shares only with Meta); Apple, 2nd by size, is only 7th by maturity, dragged down by a proven case of AI-washing, a $250 million settlement in May 2026 for misleading communication about Siri's capabilities [12].

One last case is worth the detour, because it bridges the table's two columns. Microsoft has fallen below $3 trillion, behind Alphabet, which crossed $4 trillion as early as January 2026 [7]. In 2026 the market stopped crediting AI spending blindly: it now demands to see it turn into revenue. The same capex billions read as an asset at Alphabet, its in-house TPU chips, its Gemini model, and as a reason for caution at Microsoft, whose AI architecture rests largely on its partner OpenAI, currently renegotiating [8]. The market is starting, in its own way and with its own instruments, to score maturity.

Box. The Top 10 moves. Our maturity study froze the Nasdaq Top 10 in January 2026: ASML (a semiconductor-equipment maker) and Netflix then held the 9th and 10th places by size. Six months later, carried by the Q2 2026 chip rally, Micron and AMD overtook them. It is a live demonstration that a "Top 10" is a moving target: between two snapshots, two names out of ten have already changed. The maturity scores above apply to the January scope.

Dimension by dimension: what the grid really reveals

The overall score hides what matters most. It is by going down to the eight dimensions that the two panels diverge, and that the surprises appear. Here is each dimension's average across the ten companies of each index (computed from the studies' detailed scores [12]).

Dimension	CAC 40 average	Nasdaq average	Gap
D1 · Strategy	3.3	3.6	Nasdaq +0.3
D2 · Governance	3.2	3.2	tie
D3 · Data	2.4	2.1	CAC 40 +0.3
D4 · Technology	3.1	3.4	Nasdaq +0.3
D5 · Talent	3.1	2.4	CAC 40 +0.7
D6 · Deployment	3.3	3.8	Nasdaq +0.5
D7 · Risk / compliance	2.7	3.3	Nasdaq +0.6
D8 · Communication	3.2	3.4	Nasdaq +0.2

Four lessons come out of it.

1. Data is the general blind spot, and the two indices' only common ground. D3 is the weakest dimension in both panels, the only one below 2.5 on both sides (2.4 for the CAC 40, 2.1 for the Nasdaq, the lowest score in the whole study). All claim a unified data platform, none publishes a quality or traceability indicator for its training data. This is no longer just an artefact of under-communication: at NVIDIA as at Meta, it is an active dispute over data provenance.

The Nasdaq pulls ahead where AI is the product. D6 Deployment (3.8 against 3.3) and D1 Strategy (3.6 against 3.3): when generative AI is directly your growth engine, the value created is visible, quantified and published every quarter. For an industrial or luxury group, AI remains an internal transformation lever, harder to document, and often absent from the equity story.

3. The French surprise: labour dialogue. D5 Talent is the only dimension where the CAC 40 clearly beats the Nasdaq, 3.1 against 2.4, the widest gap in the study (its only other lead is on D3, by 0.3 point, and from the bottom). On one side, signed labour agreements (Safran, January 2026) and quantified training plans (L'Oréal, Sanofi); on the other, an aggressive AI narrative coexisting almost everywhere with documented waves of job cuts (Microsoft, Meta, Tesla, Amazon). French labour law and the presence of employee representative bodies in AI transformation are not just a constraint: in this grid, they produce proof, and therefore score.

The compliance paradox. One would expect the opposite, but it is the Nasdaq that leads the CAC 40 on D7 Risk and compliance (3.3 against 2.7), even though the European AI Act applies first and foremost to Europeans. The explanation is mechanical: American companies detail their AI risks in their 10-K filings and have taken clear public positions on the EU's GPAI Code of Practice (Alphabet, Microsoft and Amazon signed; Meta alone refused). On the CAC 40 side, not one of the ten has published an AI risk map or a classification plan for its systems, while the obligations applicable to high-risk systems take effect on 2 August 2026 [13], less than three weeks after the study's date. Being subject to a rule and being able to prove you are preparing for it are two different things.

Where is the value going?

Overlay the two readings and a thesis emerges, consistent with what we argue in Where is AI's value going?. Three takeaways for a business leader.

Market value has gone to AI infrastructure; maturity lies elsewhere. Those who top the market-cap ranking are the sellers of the underlying layer: chips, cloud, models. Yet under the lens of internal governance, it is the AI users, Alphabet, Meta, Amazon, L'Oréal, Sanofi, who lead, not the sellers of silicon. The lesson for an SME is therefore not "buy this or that stock": that would be a misreading, and it is not our job. It comes down instead to one sentence: "become a company that uses AI with maturity": strategy, governance, clean data, use cases in production, compliance. That is exactly the path described in our white paper on industrialising AI.

The market no longer rewards the announcement; it rewards the proof. Microsoft back below $3 trillion for failing to turn its capex into revenue, Apple fined $250 million for an unkept AI promise: the penalty, financial or legal, now falls on unproven claims. The grid says the same thing at the micro scale: a score ≥ 3 without public proof does not exist. Proof is read in the facts, not in the slogans. This is true for a Nasdaq giant as much as for a mid-cap in the Drôme.

Sovereignty is not a slogan, it is a reading of these tables. That Europe has no chip or cloud champion in the world Top 10, the only European in the Nasdaq panel, ASML, being an equipment maker, not an AI-software player, is a strategic fact. Depending on an AI infrastructure listed 6,000 kilometres away, billed in dollars and subject to other rules, is a supply-chain risk that can be managed: data hosting in Europe, open models that run locally, vendor reversibility. We developed this point in Digital sovereignty.

How to read these two rankings without falling into traps

A market cap is a snapshot, not a truth. NVIDIA and Apple swapped the world's number-one spot several times in 2026; ASML and Netflix left the Nasdaq Top 10 within six months. Always date the figure.
A big market cap ≠ a good investment, and size ≠ AI maturity. Market cap measures what the market already pays; maturity, the ability to govern AI; neither predicts a share price.
Weight in the index ≠ size (the LVMH reminder), and index membership ≠ world ranking (the SpaceX and ASML reminders).
The maturity measured here is public, not audited. It mechanically under-scores internal dimensions; a low score can reflect discreet communication (Hermès) as much as a genuine lag.
A score is always read together with its evidence level. A 2 can mean "does not do it" or "does not publish it": the grid explicitly distinguishes the two (documented, declarative, absent). That is what makes it arguable line by line, and therefore useful.

Disclaimer. This article is educational and informational. It is neither investment advice nor a recommendation to buy or sell any security whatsoever. An AI-maturity score is not a price forecast. Past performance does not predict future performance. For any investment decision, consult a licensed financial adviser.

Do you run an SME or a mid-cap and want to turn "where AI's value is going" into concrete decisions, with real AI maturity rather than slogans? That is what the MATIA Method™ is for. Start with a free AI Express Audit & Roadmap.

Sources

Market-cap figures as of 20 July 2026; maturity scores dated at their source. Market capitalisations move daily.

[1] CompaniesMarketCap, Largest CAC 40 companies by market capitalisation (total index capitalisation ≈ $2.859 trillion). https://companiesmarketcap.com/cac-40/largest-companies-by-market-cap/

[2] CompaniesMarketCap, Companies ranked by Market Cap (world ranking and Nasdaq stocks). https://companiesmarketcap.com/

[3] Tout sur mes finances, Bourse de Paris: the composition of the CAC 40 index (weightings as of 12 March 2026). https://www.toutsurmesfinances.com/bourse/a/bourse-de-paris-la-composition-de-l-indice-cac-40

[4] Boursorama, The Arnault family passes 50% of LVMH's capital (AMF filing of 25 February 2026; 50.01% of capital). https://www.boursorama.com/bourse/actualites/la-famille-arnault-accroit-sa-participation-dans-lvmh-et-depasse-les-50-du-capital-57f8b4e9830e31b1e6098b26a909a313 · Forbes France, The Arnault family secures majority control of LVMH. https://www.forbes.fr/business/la-famille-arnault-securise-la-majorite-du-capital-de-lvmh/

[5] Forbes, Nvidia Hits Record $5.5 Trillion Value, First Company To Ever Reach Mark (13 May 2026). https://www.forbes.com/sites/antoniopequenoiv/2026/05/13/nvidia-hits-record-55-trillion-value-first-company-to-ever-reach-mark/

[6] CNBC, Apple, Nvidia vie for title of world's most valuable company (17 July 2026). https://www.cnbc.com/2026/07/17/apple-nvidia-aapl-nvda-market-cap.html · Forbes, Apple Briefly Unseats Nvidia As World's Largest Company. https://www.forbes.com/sites/tylerroush/2026/07/17/apple-unseats-nvidia-as-worlds-largest-company/

[7] CompaniesMarketCap, Alphabet (Google), Market capitalization (crossed $4 trillion as early as January 2026; ≈ $4.345 trillion as of 20 July 2026). https://companiesmarketcap.com/alphabet-google/marketcap/

[8] The Motley Fool, The Glaring Reason Microsoft Is Falling Behind Alphabet and Amazon (3 May 2026). https://www.fool.com/investing/2026/05/03/microsoft-falling-behind-alphabet-amazon/

[9] The Motley Fool, 1 Unstoppable Stock to Buy Before It Joins Micron and Broadcom in the $1 Trillion Club (16 July 2026). https://www.fool.com/investing/2026/07/16/1-unstoppable-stock-to-buy-before-it-joins-micron/ · Dealroom, Micron, Intel, AMD add $2T market cap in Q2 2026 AI chip rally. https://app.dealroom.co/news/feed/micron-intel-amd-add-2t-market-cap-in-q2-2026-ai-chip-rally

[10] Invesco QQQ / Nasdaq, Index methodology & holdings (modified weighting with anti-concentration rules; the top 10 ≈ half the index). https://indexes.nasdaq.com/docs/Methodology_NDX.pdf

[11] Trading Economics, Euro US Dollar Exchange Rate (EUR/USD) (€1 ≈ $1.14 as of 20 July 2026). https://tradingeconomics.com/euro-area/currency

[12] Croissance & Transitions, in-house studies "AI Maturity, CAC 40 Top 10" (13 July 2026) and "AI Maturity, Nasdaq Top 10" (14 July 2026), the same methodology applied identically to both panels. Proprietary grid derived from the MATIA Method™: 8 weighted dimensions (D1 Strategy and value creation 15% · D2 Governance and accountability 15% · D3 Data 12% · D4 Technology and industrialisation 12% · D5 Talent, culture and organisation 12% · D6 Deployment and realised value 12% · D7 Risk, compliance and internal control 12% · D8 Financial, extra-financial and stakeholder communication 10%), scored from 1 to 5 with an evidence level attached (documented / declarative / mixed / absent), the weighted overall score read on 5 levels (Emerging < 1.8 · Exploratory 1.8-2.5 · Structured 2.6-3.4 · Integrated 3.5-4.2 · Transformative > 4.2). The per-dimension averages quoted in the article are calculated from both studies' detailed scores. Scoring built exclusively on public sources: universal registration documents, annual and sustainability reports, SEC 10-K / 20-F filings, DEF 14A proxy statements, investor transcripts, specialist press. Includes the documented AI-washing cases (Apple's $250 million Siri settlement, May 2026; Tesla's FSD statistics challenged by a Reuters investigation) and public positions on the AI Act (Alphabet, Microsoft, Amazon signatories of the GPAI Code of Practice; Meta's refusal).

[13] Regulation (EU) 2024/1689 on artificial intelligence (AI Act), obligations applicable to high-risk AI systems from 2 August 2026.

The cash of AI: who pockets it, and why it may be circular

Mon, 20 Jul 2026 00:00:00 GMT

By Paul-Antoine TUAL, AI Transformation Leader, Croissance et Transitions, July 2026.

Framing. “The technology can be very real and the valuation excessive, both at once.” That was true in 2000, when Cisco was worth more than $550 billion and it took nearly 25 years for its stock to reclaim that peak; it is true in 2026. As a young consultant, I lived through the telecoms crash, and I recognise mechanisms that are returning today. In an earlier article, we ranked the ten largest CAC 40 and Nasdaq companies by size and by AI maturity. Here, we follow the money, in three steps: who really pockets the cash (the value chain), why it may be circular (circular financing), and who could flip the table (the industrial challengers). Neither doom-mongering nor denial: a grid for sorting. This article is not investment advice; several figures are press or non-audited estimates, flagged as such.

1. The four floors: who really pockets the cash?

During a gold rush, it is not always the prospectors who get rich. It is often the sellers of picks and shovels. The AI value chain reads exactly the same way, in four floors, from closest to the silicon to closest to the user.

Floor 1 · Silicon (the picks). This is where the cash accumulates today. Nvidia posts a gross margin of about 75% [1], a dizzying level for hardware. Taiwan's TSMC, which etches these chips, reported in mid-July 2026 a gross margin of around 68% [2]; Korea's SK Hynix, king of the high-performance memory (HBM) that AI servers depend on, reached around 70% operating margin at its latest record quarter [3]. The risk on this floor is not demand, which remains insatiable, but cyclicality, and the fact that the biggest customers (Google, Amazon, OpenAI) are now designing their own chips to chip away at Nvidia's near-monopoly (see Part 3).

Floor 2 · Infrastructure (the shovels and the electricity). Profitable, but capital-hungry. The four big American hyperscalers (Amazon, Alphabet, Meta, Microsoft) are spending in 2026 on the order of $600-700 billion in capital expenditure (capex), even more if Oracle is included, mostly on data centres and chips [4]. The constraint is becoming physical: the International Energy Agency projects a rise of about 130% in the electricity consumption of American data centres by 2030 [5]. Cloud is profitable (Amazon Web Services exceeds 35% operating margin), but capex is swallowing a growing share of revenue, and financing is shifting towards debt.

Floor 3 · Models (the promise). Hypergrowth and massive losses. OpenAI reportedly shows around $25 billion of annualised revenue, but also on the order of $14 billion of projected losses in 2026 (a non-audited figure, to be treated as an estimate [6]). It is the most media-visible floor, and the most dependent on private capital.

Floor 4 · Applications (the smallest floor). The revenue actually captured by enterprise generative-AI applications is counted in the tens of billions, on the order of $20 billion according to Menlo Ventures' estimates [7].

<figure style="margin:2rem 0;"> <svg viewBox="0 0 720 424" width="100%" height="auto" role="img" aria-labelledby="f1t f1d" style="max-width:100%;height:auto;font-family:system-ui,-apple-system,'Segoe UI',Roboto,sans-serif;"> <title id="f1t">The four floors of the AI value chain</title> <desc id="f1d">Four boxes connected by arrows, from silicon to applications: about 75% gross margin at Nvidia, $600 to $700 billion of infrastructure capex, about $25 billion of revenue against $14 billion of projected losses for models, about $20 billion of revenue for applications.</desc> <text x="8" y="20" font-size="14" font-weight="700" fill="#0B2545">The four floors of the AI value chain</text> <rect x="8" y="40" width="704" height="72" rx="10" fill="#FFFFFF" stroke="#E2E8F0" stroke-width="1.5"/> <text x="24" y="70" font-size="13" font-weight="700" fill="#0F172A">Floor 1 · Silicon (the picks)</text> <rect x="466" y="58" width="230" height="36" rx="6" fill="#FF7A59"/> <text x="581" y="81" font-size="12.5" font-weight="700" fill="#FFFFFF" text-anchor="middle">≈ 75% gross margin (Nvidia)</text> <line x1="360" y1="112" x2="360" y2="128" stroke="#0B2545" stroke-width="2"/> <polygon points="354,128 366,128 360,140" fill="#0B2545"/> <rect x="8" y="140" width="704" height="72" rx="10" fill="#FFFFFF" stroke="#E2E8F0" stroke-width="1.5"/> <text x="24" y="170" font-size="13" font-weight="700" fill="#0F172A">Floor 2 · Infrastructure (the shovels)</text> <rect x="496" y="158" width="200" height="36" rx="6" fill="#FF7A59"/> <text x="596" y="181" font-size="12.5" font-weight="700" fill="#FFFFFF" text-anchor="middle">≈ $600-700 billion capex</text> <line x1="360" y1="212" x2="360" y2="228" stroke="#0B2545" stroke-width="2"/> <polygon points="354,228 366,228 360,240" fill="#0B2545"/> <rect x="8" y="240" width="704" height="72" rx="10" fill="#FFFFFF" stroke="#E2E8F0" stroke-width="1.5"/> <text x="24" y="270" font-size="13" font-weight="700" fill="#0F172A">Floor 3 · Models (the promise)</text> <rect x="416" y="258" width="280" height="36" rx="6" fill="#FF7A59"/> <text x="556" y="281" font-size="11.5" font-weight="700" fill="#FFFFFF" text-anchor="middle">≈ $25 billion revenue / ≈ $14 billion losses</text> <line x1="360" y1="312" x2="360" y2="328" stroke="#0B2545" stroke-width="2"/> <polygon points="354,328 366,328 360,340" fill="#0B2545"/> <rect x="8" y="340" width="704" height="72" rx="10" fill="#FFFFFF" stroke="#E2E8F0" stroke-width="1.5"/> <text x="24" y="370" font-size="13" font-weight="700" fill="#0F172A">Floor 4 · Applications (the smallest floor)</text> <rect x="516" y="358" width="180" height="36" rx="6" fill="#FF7A59"/> <text x="606" y="381" font-size="12.5" font-weight="700" fill="#FFFFFF" text-anchor="middle">≈ $20 billion revenue</text> </svg> <figcaption style="font-size:0.8rem;color:#475569;margin-top:0.5rem;line-height:1.5;">Figure 1. The closer a floor is to the physical component, the more real cash it already pockets; applications, the most visible floor to the end user, remain the smallest of the four in revenue. Sources: see notes [1][4][6][7] of the article.</figcaption> </figure>

Put the two ends side by side: about $700 billion of capex against about $20 billion of application revenue. This gap is not a detail, it is the point. It does not prove that AI is an illusion: the lower floors are, themselves, pocketing very real cash. But it says one simple thing: today, the bottom of the chain sells picks at a premium; the top is still selling a promise. A telling fact: for the first time in thirty years of software history, software has lower margins than hardware, with a gross margin estimated at around 33% at OpenAI, against 75% at Nvidia [1][6]. The entire investment playbook of the past thirty years is upside down.

<figure style="margin:2rem 0;"> <svg viewBox="0 0 720 200" width="100%" height="auto" role="img" aria-labelledby="f2t f2d" style="max-width:100%;height:auto;font-family:system-ui,-apple-system,'Segoe UI',Roboto,sans-serif;"> <title id="f2t">The gap between AI capex and application revenue</title> <desc id="f2d">Two bars to scale: about $700 billion of hyperscaler capex against about $20 billion of application revenue, one bar barely visible next to the other.</desc> <text x="8" y="20" font-size="14" font-weight="700" fill="#0B2545">The gap: AI capex against application revenue</text> <text x="8" y="60" font-size="12.5" font-weight="600" fill="#0F172A">Hyperscaler capex (2026)</text> <rect x="8" y="70" width="460" height="34" rx="4" fill="#0B2545"/> <text x="478" y="92" font-size="13" font-weight="700" fill="#0B2545">≈ $700 billion</text> <text x="8" y="150" font-size="12.5" font-weight="600" fill="#0F172A">Worldwide application revenue (2026)</text> <rect x="8" y="160" width="13" height="34" rx="2" fill="#FF7A59"/> <text x="31" y="182" font-size="13" font-weight="700" fill="#E55F3B">≈ $20 billion</text> </svg> <figcaption style="font-size:0.8rem;color:#475569;margin-top:0.5rem;line-height:1.5;">Figure 2. To scale, the application-revenue bar is barely visible next to the capex bar: about $700 billion invested against about $20 billion of application revenue captured. Sources: see notes [4][7] of the article.</figcaption> </figure>

2. The closed loop: when the same dollar is counted several times

Here is the mechanism that should make a business leader cautious. The loop, in one sentence: Nvidia invests in OpenAI → OpenAI pays a cloud provider (Oracle) for compute → that provider buys chips from Nvidia → Nvidia's revenue rises → its market capitalisation funds the next round. The same dollar can be counted as revenue at each link.

<figure style="margin:2rem 0;"> <svg viewBox="0 0 720 372" width="100%" height="auto" role="img" aria-labelledby="f3t f3d" style="max-width:100%;height:auto;font-family:system-ui,-apple-system,'Segoe UI',Roboto,sans-serif;"> <title id="f3t">The closed loop of AI financing</title> <desc id="f3d">Four boxes in a loop: Nvidia invests in OpenAI, which pays a cloud provider for compute, which buys chips from Nvidia, whose revenue rises and funds the next round.</desc> <text x="8" y="20" font-size="14" font-weight="700" fill="#0B2545">The closed loop: when the same dollar keeps looping back</text> <rect x="20" y="48" width="310" height="86" rx="10" fill="#FFFFFF" stroke="#E2E8F0" stroke-width="1.5"/> <text x="36" y="96" font-size="13" font-weight="700" fill="#0F172A">① Nvidia invests in OpenAI</text> <rect x="390" y="48" width="310" height="86" rx="10" fill="#FFFFFF" stroke="#E2E8F0" stroke-width="1.5"/> <text x="406" y="88" font-size="13" font-weight="600" fill="#0F172A">② OpenAI pays the cloud (Oracle)</text> <text x="406" y="108" font-size="13" font-weight="600" fill="#0F172A">for compute</text> <line x1="330" y1="91" x2="380" y2="91" stroke="#0B2545" stroke-width="2"/> <polygon points="380,85 380,97 392,91" fill="#0B2545"/> <rect x="390" y="266" width="310" height="86" rx="10" fill="#FFFFFF" stroke="#E2E8F0" stroke-width="1.5"/> <text x="406" y="306" font-size="13" font-weight="600" fill="#0F172A">③ The cloud provider buys</text> <text x="406" y="326" font-size="13" font-weight="600" fill="#0F172A">chips from Nvidia</text> <line x1="708" y1="144" x2="708" y2="256" stroke="#0B2545" stroke-width="2"/> <polygon points="702,256 714,256 708,268" fill="#0B2545"/> <rect x="20" y="266" width="310" height="86" rx="10" fill="#FFFFFF" stroke="#E2E8F0" stroke-width="1.5"/> <text x="36" y="306" font-size="13" font-weight="600" fill="#0F172A">④ Nvidia's revenue rises,</text> <text x="36" y="326" font-size="13" font-weight="600" fill="#0F172A">funds the next round</text> <line x1="390" y1="309" x2="340" y2="309" stroke="#0B2545" stroke-width="2"/> <polygon points="340,303 340,315 328,309" fill="#0B2545"/> <line x1="12" y1="256" x2="12" y2="144" stroke="#0B2545" stroke-width="2"/> <polygon points="6,144 18,144 12,132" fill="#0B2545"/> </svg> <figcaption style="font-size:0.8rem;color:#475569;margin-top:0.5rem;line-height:1.5;">Figure 3. The loop closes in on itself: each arrow corresponds to a real payment, but the starting point and the end point are the same company, Nvidia. Sources: see notes [8][9][11] of the article.</figcaption> </figure>

At the end of 2025, the financial press put these cross-deals at about $1 trillion, according to Bloomberg's “AI circular deals” tally [8]: a press estimate, not consolidated accounting data. The emblematic examples, with their real status, are worth more than the headline figure:

Nvidia → OpenAI: the “$100 billion” announced in 2025 was only a letter of intent. Jensen Huang himself acknowledged in early 2026 that it was “never a commitment”; it was replaced by a far smaller stake in the March funding round [9]. (Saying “announced, then unwound” is more accurate, and more instructive.)
OpenAI ↔ AMD, then Meta ↔ AMD: the customer is paid in shares (warrants on a slice of AMD's capital) to buy the alternative supplier's chips [10].
OpenAI ↔ Oracle: a compute purchase commitment on the order of $300 billion over five years, an order book, not revenue received [11].
Microsoft ↔ OpenAI: about a quarter of the capital against massive cloud purchase commitments (the October 2025 restructuring), recast again in spring 2026, ending exclusivity and revenue sharing [12].
Nvidia ↔ CoreWeave: Nvidia backstops several billion dollars of its own customer's unsold capacity, a customer that took on debt to buy… Nvidia chips [13].

The precedent I lived through. In the late 1990s, telecoms-equipment makers practised vendor financing: they lent their own customers the money to buy their equipment. Lucent carried billions of dollars in customer credit; Nortel lent in proportions that even exceeded the orders themselves. As long as the music played, the order book kept swelling. Then customers stopped paying: Lucent posted a colossal loss in 2001, its stock going from tens of dollars to less than one dollar; Nortel went from a market capitalisation of more than a hundred billion to bankruptcy [14]. And Cisco, the indestructible one: more than $550 billion of market capitalisation in March 2000, an earnings multiple of about 200, and nearly 25 years for its stock to reclaim that peak (late 2025) [15]. The lesson is not “the technology was fake”: the internet transformed everything. The lesson is that the shareholder who paid the peak lost anyway, because the valuation and the circular revenue did not survive it.

Who is ringing the alarm bell in 2026? The Bank of England (October 2025: stretched valuations, rising concentration, risk of a sharp correction), the IMF, investor Michael Burry (short positions on AI stocks), and above all the Bank for International Settlements (June 2026), which flags circular financing that is “poorly disclosed” and the risk of multiple pledging of the same asset [16]. The repricing has, in fact, already begun: in June 2026, a single trading session wiped out on the order of $1.3 trillion of market capitalisation across semiconductors [17].

Honesty, which is also credibility. This is not 2000 all over again, and saying so strengthens the analysis rather than weakening it: unlike back then, the hyperscalers finance most of this on real, massive cash flows; the leaders' valuation multiples run at around 25 times earnings, against nearly 58 times in March 2000 [18]; paying demand genuinely does exist; and AMD-style warrants cost no cash. The market read June's correction as a repricing, not a rupture. The conclusion, therefore, is not to “flee”, but to sort: floor by floor, promise against proof.

3. Who can flip the table: the 2027-2028 challengers

Record margins are not a shield: they are magnets for competition. To understand where the edifice can move, one concept is enough, and it is too rarely explained to the general public.

🎓 Training vs inference. Training a model means building it: a handful of mega-projects a year, tens of thousands of chips synchronised for months, zero tolerance for error; this is where the software ecosystem matters most. Inference means running the already-built model to answer each request: billions of times a day, where only the cost per request matters. Training is the prestige market; inference is the volume market, and volume is what is exploding. The entire competitive dynamic of 2027-2028 flows from this distinction: you do not dethrone Nvidia on training, you go around it on inference.

Nvidia (more than 80% of AI accelerators, down from ~92% in 2023) [19] owes its dominance as much to CUDA, the software layer with which fifteen years of AI code has been written, as to its chips. Moving off CUDA means rewriting and re-validating thousands of lines of critical code: a massive switching cost, exactly like changing a bank's core computing system. But this lock is asymmetric: enormous on training, weak on inference, where standardised layers run better and better on non-Nvidia hardware. Hence three attacks already real in 2026:

AMD, the frontal attack. Its new generation of chips (MI450) has landed giant commitments: a major lab has committed to several gigawatts, a major cloud player is deploying tens of thousands of them [20]. AMD no longer sells chips, but complete systems, like Nvidia.
The cloud giants' in-house chips, the integration attack. Google's TPUs are moving beyond its captive cloud (a leading lab is committing to them at scale) [21], and Amazon has deployed more than a million of its Trainium chips [22]. Why? At a hyperscaler's scale, saving 30 to 40% of inference cost justifies billions in research to design a dedicated chip of one's own.
Sovereign ASICs, finally: OpenAI is designing its own chip (with Broadcom), and China is pushing its alternatives under regulatory constraint.

🎓 The stock-market mechanics to understand. At more than 80% market share, Nvidia has almost no share left to gain, only to lose, and each point lost is paid for twice: in revenue and in valuation multiple. Yet multiple erosion almost always precedes revenue erosion. A sign already visible: by mid-July 2026, Nvidia's stock is up only about 9% on the year, while Apple's has risen by nearly 23% [23]. The market has begun to sort.

A word on the other floors: the fatter a floor's margins, the more concrete its 2027 alternatives. TSMC remains the only true bottleneck with no credible substitute within two years; its risk is geopolitical (Taiwan), not competitive. HBM memory (SK Hynix) shows 70% margins because we are at the peak of a notoriously violent cycle, one that the return of Samsung and the rise of Micron will shave down. And the models floor is the most substitutable of all: switching from one AI provider to another takes almost a single line of configuration, which pushes the price towards marginal cost. “75% margin at Nvidia created AMD-OpenAI; 70% at SK Hynix woke up Samsung and Micron. Record margins are not a moat, they are invitations to tender.”

What this changes for a business leader (not an investor)

An SME or mid-cap business leader does not have to bet on Nvidia's share price. But this reading gives them a compass, and it echoes the thesis of our articles Where is AI's value going? and the AI maturity of the Top 10.

The real question is not “should we buy AI?” but “who will AI re-rate, and who will it de-rate?” The deflation in the cost of intelligence (the price of a given AI capability is divided by around 10 every year) is a tax on those who sell tokens and a subsidy for those who consume them. In other words: value migrates mechanically towards the companies that put AI to work in their business, not towards those that sell it. For an SME, the stake is not to guess the top of the bubble; it is to stand on the right side of the counter: to industrialise concrete use cases while the technology layer becomes abundant and cheap. That is exactly the purpose of our white paper on industrialising AI and of the MATIA Method™.

And the reading rule, valid from the Nasdaq to the workshop floor: proof is read in margins, not in slogans. 75% gross margin at Nvidia is a fact; the productivity gains promised in keynote speeches are still a promise.

Disclaimer. This article is educational and informational. It is neither investment advice nor a recommendation to buy or sell any security whatsoever. Several figures cited are press or non-audited estimates, flagged as such. Past performance does not predict future performance. For any investment decision, consult a licensed financial investment adviser.

Would you like your company to be on the right side of this value migration, the one that puts AI to work, not the one that suffers it? Start with a free AI Express Audit & Roadmap.

Sources

Figures as of mid-July 2026. Several amounts are press or non-audited estimates, indicated as such in the text.

[1] NVIDIA official results (quarter ended April 2026), gross margin ≈ 74.9%, quarterly revenue ≈ $81.6 billion of which ≈ $75.2 billion data centre.

[2] TSMC Q2 2026 results (published 16 July 2026), gross margin ≈ 67.7% (rounded to ≈ 68% in the text).

[3] SK Hynix results (Q1 2026, a record), operating margin ≈ 72% (HBM memory).

[4] CNBC / market roundups (February 2026), combined 2026 capex of the four hyperscalers ≈ $640-700 billion.

[5] International Energy Agency, Electricity 2026 / Energy and AI: US data-centre electricity consumption ≈ +130% by 2030.

[6] Press estimates (The Information), OpenAI ≈ $25 billion annualised revenue, ≈ $14 billion projected 2026 losses, gross margin ≈ 33%; non-audited figures.

[7] Menlo Ventures, State of Generative AI in the Enterprise (Dec. 2025): enterprise-side GenAI application revenue ≈ $19-20 billion (private, non-audited estimate).

[8] Bloomberg, AI circular deals: cross-deals tallied on the order of $1 trillion (October 2025; press estimate, not consolidated accounting data).

[9] Fortune, Jensen Huang, “never a commitment” (February 2026): the “$100 billion” Nvidia→OpenAI, an unexecuted letter of intent.

[10] OpenAI/AMD and Meta/AMD press releases (2025-2026), warrants on a slice of AMD's capital.

[11] Bloomberg / CNBC, OpenAI↔Oracle commitment ≈ $300 billion over 5 years (purchase commitment, not cash received).

[12] CNBC, recast of the Microsoft-OpenAI agreement (April 2026).

[13] Financial press, Nvidia's guarantee on CoreWeave's unsold capacity (≈ $6.3 billion).

[14] American Affairs, Who Lost Lucent?; Lucent SEC filings (loss ≈ $16.2 billion in 2001, stock from ~$65 in 1999 to <$1 in 2002) and Nortel (2009 bankruptcy, after a peak >$100 billion).

[15] CNBC, Cisco's stock reclaims its March 2000 peak on 10 December 2025: >$550 billion market capitalisation at the peak, P/E ≈ 200 (only the share price recovered its level; market capitalisation remains lower, owing to share buybacks).

[16] Bank for International Settlements, Annual Report (28 June 2026); Bank of England (FPC, 8 October 2025); IMF (October 2025).

[17] Market roundups, a June 2026 session: ≈ −$1.3 trillion of market capitalisation across semiconductors (sector sell-off; press estimates).

[18] Multiple comparison: leaders ≈ 25× earnings (2026) vs ≈ 58× at the March 2000 peak.

[19] Sector estimates (IDC / TrendForce, second-hand), Nvidia ≈ 80-88% of data-centre AI accelerators in 2026 (down from ~92% in 2023).

[20] AMD / OpenAI / Oracle press releases, commitments on the MI450 generation (2026).

[21] Google Cloud / Anthropic, TPUs opened beyond the captive cloud; commitment from a major lab.

[22] Amazon Web Services, deployment of more than a million Trainium chips.

[23] CNBC, Apple, Nvidia vie for title of world's most valuable company (17 July 2026): year-to-date, Apple ≈ +23%, Nvidia ≈ +9%.

We removed Git from a production product: roadmap-driven development

Wed, 15 Jul 2026 00:00:00 GMT

On 11 July, we archived .git

On 11 July 2026, we compressed the .git directory of Junyr (our sovereign e-mail and business-management platform, in production, with more than 90 operational features), dropped the archive onto a backup server, and removed Git from the working tree. Ever since, any git status there answers exactly what it should: fatal: not a git repository.

This is a documented, reversible decision: the archive exists, and so does the restore procedure. It follows from an observation many teams will make in the years ahead. When code is no longer written by human hands but by orchestrated AI agents, Git still versions something; it is simply no longer the thing that matters.

Four days later, as we publish this: fifty plans recorded in the registry that replaced Git, twenty already delivered, and a batch of twelve plans implemented in a single day then promoted to production the day after: a cutover gated on service health, zero rollback.

This article tells the story of what we put in Git's place, why it works and, above all, the cases in which you should not imitate us.

What Git versions, what it no longer versions

Git was created in 2005 by Linus Torvalds for the development of the Linux kernel, to solve one precise problem: hundreds of humans editing the same files in parallel, with no central authority. All of its primitives (the branch, the merge, the diff, the blame) stem from that problem: reconciling the simultaneous work of people typing text. For that problem, Git remains unmatched, and nothing here says otherwise.

But when AI agents write most of the code, three silent shifts take place.

The diff stops being the unit of review. Nobody seriously reads a 3,000-line diff generated in twenty minutes. Code review by reading the diff (the social function of the pull request) was already dead in practice in teams that had adopted agents; it survived as a ritual. Figures published in May 2026 by The Register give the measure of the problem: AI-generated projects grew 206% year-on-year on GitHub, and generated code piles up 10.83 issues per pull request, against 6.45 for human code [2]. More volume, less readability, more defects per batch: control can no longer be the re-reading of the produced text. It has to move: upstream towards the decision, downstream towards tooled verification.

<figure style="margin:2rem 0;"> <svg viewBox="0 0 720 260" width="100%" height="auto" role="img" aria-labelledby="f1t f1d" style="max-width:100%;height:auto;font-family:system-ui,-apple-system,'Segoe UI',Roboto,sans-serif;"> <title id="f1t">Issues detected per pull request, AI-generated code versus human code</title> <desc id="f1d">Vertical bar chart: 10.83 issues on average per pull request for AI-generated code, against 6.45 for code written by humans.</desc> <text x="8" y="20" font-size="14" font-weight="700" fill="#0B2545">More volume, more defects per batch</text> <line x1="60" y1="200" x2="660" y2="200" stroke="#E2E8F0"/> <rect x="220" y="40" width="110" height="160" fill="#FF7A59"/> <text x="275" y="30" font-size="15" font-weight="700" fill="#E55F3B" text-anchor="middle">10.83</text> <text x="275" y="222" font-size="13" fill="#0F172A" text-anchor="middle">AI-generated code</text> <rect x="410" y="105" width="110" height="95" fill="#13315C"/> <text x="465" y="95" font-size="15" font-weight="700" fill="#0B2545" text-anchor="middle">6.45</text> <text x="465" y="222" font-size="13" fill="#0F172A" text-anchor="middle">Human code</text> </svg> <figcaption style="font-size:0.8rem;color:#475569;margin-top:0.5rem;line-height:1.5;">Figure 1. AI-generated code accumulates 10.83 issues per pull request, against 6.45 for human code. Source: The Register, 15 May 2026 [2].</figcaption> </figure>

The commit stops carrying the intent. A commit message written by an agent describes what changed, not what was decided. The real intent lives elsewhere: in the prompt, in the plan, in the orchestration session. Scott Chacon, co-founder of GitHub, acknowledges it himself: Git's interface has barely changed since 2005, and version control needs rethinking for humans and agents. His wager: "the best engineers of the future will be the best writers" [1].

The branch stops being the right level of parallelism. The merge conflict is the artefact of a choice: parallelising at the level of the code. If you parallelise at the level of decisions (several plans investigated at the same time) and serialise the writing (one implementer at a time), the merge conflict does not become easier to resolve: it disappears from the system.

The industry sees the same problem and mostly answers with "a better Git": agent-friendly interfaces, reworked parallel branches, dedicated tooling [1][2]. Even Cursor has gone this route: it announced Origin in June 2026, a Git forge "for the agentic era" built for dozens of agents running in parallel, where human review is explicitly redesigned as bulk approval, or even delegated to another agent [4]. That is the right answer for large human teams. In our context (one product, one human decision-maker, agents doing the writing), we chose the radical answer: remove Git and version something else.

The inversion: version the intent, not the text

This inversion has an academic name. In early 2026, a movement called spec-driven development took shape: the specification becomes the source of truth, the code a derived artefact, generated from it or verified against it. The literature distinguishes three degrees of rigour: spec-first (the spec precedes the code), spec-anchored (the spec remains the living reference), spec-as-source (the code is no longer maintained directly at all) [3]. GitHub, AWS and the main agent tools each have their own flavour.

Our variant moves the cursor slightly: we do not version full specifications, we version decisions. The nuance is practical. An exhaustive specification ages poorly and is hard to review; a decision (which mechanism to reuse, which approach to take, which contracts to honour, what is explicitly out of scope) ages well and can be reviewed in one screen. Code, meanwhile, has become what the casting is to a mould: regenerable, verifiable, replaceable. What is rare and precious is the mould.

Hence the name we give the method: roadmap-driven development. The roadmap becomes the operational registry through which every change to the product passes, updated continuously rather than reviewed quarterly.

In practice: two tables, one CLI, six states

The replacement infrastructure fits in one sentence: a dedicated database, two tables, a shell CLI of a few hundred lines.

The first table is the inbox, all the upstream material: ideas, backlog items, notes, and bug reports from beta users, surfaced automatically from production. Every item is triaged explicitly: dismissed, or promoted into a plan.

The second table is the plan registry. Each plan carries an identifier (T-001, T-002…), its dependencies, its risk level, and its full text. It moves through six states: draft → ready → approved → doing → done (plus abandoned). The full lifecycle:

roadmap next                 # next free identifier
roadmap submit T-042.md      # the plan enters the registry, the draft becomes disposable
roadmap approve T-042        # human gate: nothing gets implemented without it
roadmap claim T-042          # ONE implementer at a time
roadmap done T-042           # delivered, verified, recorded in the journal

Four rules hold the whole thing together.

1. Every agent session plans by default. None implements without an explicit order. N planning sessions run in parallel; each explores the code, investigates its subject and submits a plan. The plan must be "decision-complete", actionable without further exploration: exact files, existing mechanisms to reuse, contracts (signatures, schemas), tests to write, explicit out-of-scope. No full code: one to three screens. A plan that is too risky is split into plans chained by dependencies.

2. The human gate is absolute. An unapproved plan never gets implemented. This is code review moved to where it becomes possible again: reviewing a one-screen decision rather than a 3,000-line diff. Approval commits; it is the application to software development of the principle we defend for all agents: the agent proposes, the human commits.

3. A single writer. Implementation is carried by one orchestration session at a time, multi-agent inside, following the UltraCoding loop: parallel implementers, adversarial review, verification before delivery. First step after the claim, a freshness pass: cheaply checking that the files and mechanisms named by the plan still exist as described. If the code has moved since the plan was written, the plan goes back to submission and through the gate again. And any re-submission of an already-approved plan mechanically invalidates its approval: you cannot slip a change under an old signature. Research is exploring the opposite path: several agents writing in the same workspace, every write validated against the current state of the files and rejected if the agent's view is stale. The STORM framework (May 2026) gains 18.7 points on a coding benchmark that way against Git-worktree isolation [5]. That is our freshness pass, applied continuously, file by file; we preferred to serialise: a little less throughput, one entire class of conflicts fewer to govern.

4. The journal is authoritative. Whatever ships is recorded in an append-only journal: what shipped, when, why, with which follow-ups. The product's history reads as a narrative of dated decisions, not as a commit log.

The equivalence table

The honest question is "what performs each of Git's functions?" Here is the mapping, function by function:

Function performed by Git	What replaces it
Branches, worktrees: parallelise the work	Plans investigated in parallel; serialised writing (single-writer). The merge conflict disappears from the system, not merely resolved
Pull request: review before integrating	Human approval of the plan before implementation, plus tooled verification after (tests, lint, checks)
Log, blame: understand the past	Plan registry and append-only journal: the history of decisions, not of lines
Revert: go back	Deployment as complete images, previous version tagged `rollback`, cutover gated on service health
Bisect: find the origin of a regression	More than 200 numbered, executable anti-regression rules, a gate of nearly 700 checks, plus real-database test suites
Remote, push: back up	Daily off-site archive of the code and automatic export of the registry on every write (restore tested, not assumed)

The most important line is bisect. Git lets you find the offending commit after the fact; executable anti-regression rules forbid the known regression from ever returning, at every verification. It is the same shift as for review: from retrospective control over the text to permanent control over the behaviour.

Four days in: the numbers

The switch dates from 11 July 2026. As of 15 July: 50 plans in the registry (20 delivered, 10 in progress, 7 approved and awaiting implementation, 13 awaiting the gate), and around forty items in the inbox, including the beta bug reports surfaced from production. On 13 July, the orchestrator implemented a batch of 12 approved plans in one day; on the 14th, the batch went to production: a healthy cutover first time, zero rollback.

What the removed friction gave back: no more worktrees to set up, no more merges between agent sessions, no more rebasing. Planning sessions never tread on each other: they only ever write to the registry. And the system's bottleneck has become exactly the one you want: human decision time at the gate, a few minutes per plan.

Two safety nets stay permanently in place: the complete .git archives, meaning the decision is reversible in one command, and the registry export shipped every night in the off-site backup; the restore procedure was tested the very day of the switch.

<figure style="margin:2rem 0;"> <svg viewBox="0 0 720 260" width="100%" height="auto" role="img" aria-labelledby="f2t f2d" style="max-width:100%;height:auto;font-family:system-ui,-apple-system,'Segoe UI',Roboto,sans-serif;"> <title id="f2t">The lifecycle of a plan, from draft to delivery</title> <desc id="f2d">Boxes-and-arrows diagram: a plan moves through five sequential states, draft, ready, approved, doing, done, with a sixth, parallel terminal state, abandoned, reachable at any point before done.</desc> <text x="8" y="20" font-size="14" font-weight="700" fill="#0B2545">Six states, a single write path at a time</text> <rect x="10" y="50" width="112" height="46" rx="8" fill="#13315C"/> <text x="66" y="78" font-size="13" fill="#FFFFFF" text-anchor="middle">draft</text> <polygon points="126,68 126,78 138,73" fill="#475569"/> <rect x="140" y="50" width="112" height="46" rx="8" fill="#13315C"/> <text x="196" y="78" font-size="13" fill="#FFFFFF" text-anchor="middle">ready</text> <polygon points="256,68 256,78 268,73" fill="#475569"/> <rect x="270" y="50" width="112" height="46" rx="8" fill="#13315C"/> <text x="326" y="78" font-size="13" fill="#FFFFFF" text-anchor="middle">approved</text> <polygon points="386,68 386,78 398,73" fill="#475569"/> <rect x="400" y="50" width="112" height="46" rx="8" fill="#13315C"/> <text x="456" y="78" font-size="13" fill="#FFFFFF" text-anchor="middle">doing</text> <polygon points="516,68 516,78 528,73" fill="#475569"/> <rect x="530" y="50" width="112" height="46" rx="8" fill="#0B2545"/> <text x="586" y="78" font-size="13" fill="#FFFFFF" text-anchor="middle">done</text> <line x1="456" y1="96" x2="456" y2="150" stroke="#475569" stroke-dasharray="4,3"/> <polygon points="449,144 463,144 456,154" fill="#475569"/> <rect x="400" y="156" width="112" height="46" rx="8" fill="#E55F3B"/> <text x="456" y="184" font-size="13" fill="#FFFFFF" text-anchor="middle">abandoned</text> </svg> <figcaption style="font-size:0.8rem;color:#475569;margin-top:0.5rem;line-height:1.5;">Figure 2. A plan moves through five states towards delivery; abandonment remains a possible exit at any point before done. Source: Junyr development registry, cycle observed on 15 July 2026.</figcaption> </figure>

<figure style="margin:2rem 0;"> <svg viewBox="0 0 720 260" width="100%" height="auto" role="img" aria-labelledby="f3t f3d" style="max-width:100%;height:auto;font-family:system-ui,-apple-system,'Segoe UI',Roboto,sans-serif;"> <title id="f3t">Breakdown of the registry's 50 plans by state, as of 15 July 2026</title> <desc id="f3d">Horizontal bar chart: 20 plans delivered, 13 awaiting the human gate, 10 in progress, 7 approved and awaiting implementation, out of a total of 50 plans in the registry.</desc> <text x="8" y="24" font-size="14" font-weight="700" fill="#0B2545">The registry's 50 plans, by state</text> <text x="8" y="58" font-size="13" fill="#0F172A">Delivered</text> <rect x="180" y="44" width="480" height="22" fill="#0B2545"/> <text x="668" y="60" font-size="13" font-weight="700" fill="#0B2545" text-anchor="end">20</text> <text x="8" y="98" font-size="13" fill="#0F172A">Awaiting the gate</text> <rect x="180" y="84" width="312" height="22" fill="#13315C"/> <text x="500" y="100" font-size="13" font-weight="700" fill="#13315C" text-anchor="end">13</text> <text x="8" y="138" font-size="13" fill="#0F172A">In progress</text> <rect x="180" y="124" width="240" height="22" fill="#FF7A59"/> <text x="428" y="140" font-size="13" font-weight="700" fill="#E55F3B" text-anchor="end">10</text> <text x="8" y="178" font-size="13" fill="#0F172A">Approved, awaiting</text> <rect x="180" y="164" width="168" height="22" fill="#475569"/> <text x="356" y="180" font-size="13" font-weight="700" fill="#475569" text-anchor="end">7</text> <line x1="180" y1="200" x2="180" y2="206" stroke="#E2E8F0"/> </svg> <figcaption style="font-size:0.8rem;color:#475569;margin-top:0.5rem;line-height:1.5;">Figure 3. Four days after the switch, 20 of the registry's 50 plans were already delivered and 13 were still awaiting the human gate. Source: Junyr development registry, 15 July 2026.</figcaption> </figure>

The honest part: when you should absolutely not imitate us

This method works because five conditions hold. If even one is missing, removing Git is folly.

A single human decision-maker. Two humans editing the same code need merging: that is precisely the problem Git solves better than anything. A team of human developers should keep Git.
Diff review had already stopped being the real control. In our case, the code is written by agents and controlled by tests, executable rules and systematic verification. If your developers genuinely read the diffs, the pull request is your review tool. Keep it.
A promote-only production environment, never edited directly, with a proven rollback path (images, cutover gated on health).
A dense executable safety net: real-database tests, anti-regression rules, a gate of checks. Without it, removing Git means removing your last net.
Tested backups: the important word is "tested".

You also have to name what is genuinely lost: fine-grained bisect when a regression slips past the rules, line-by-line blame, and external-contribution tooling. An open-source project lives by pull requests and cannot do without them. If one of those uses is vital to you, the answer is not to imitate us, but to follow the path the industry is charting: a Git rethought for agents [1].

What this says to a business leader, beyond code

You will probably never remove Git from your systems. The real subject is the shift this story illustrates, and it extends beyond software.

When agents produce (code, but also sales proposals, campaigns, reports), the scarce resource that deserves a registry, states and a signature is no longer the artefact produced. It is the human decision. Roadmap-driven development is nothing more than that: a registry of intents in which every entry is explicit, approved, dated and traceable; the agent investigates, the human commits, execution is verified, the journal is authoritative.

It is the same loop we install at our clients with the MATIA Method™, function by function: generation becomes abundant everywhere; governance of the decision becomes the standard everywhere. The question to put to your teams (or to your provider) is less about "do you use AI?" than about the governance around it: where is the registry of what your agents are allowed to do, who signed each authorisation, and what verifies the execution? An organisation that answers by showing you a system has understood the era. An organisation that answers "it's in the commits", or worse "in the chat history", has a governance problem, not a tooling problem.

In one sentence

Git versioned the work of human hands; when the hands become agents, what needs versioning (and keeping under human signature) are the decisions.

To install this governance loop (a registry of intents, a human gate, verified execution) in your AI projects, see the MATIA Method™ and the AI Express Audit & Roadmap.

Going further

In the same series: Vibe coding is dead: enter UltraCoding, the production line that implements the plans, and Agentic systems engineering: the human in the loop, the "the agent proposes, the human commits" principle applied to every agent.

Paul-Antoine TUAL, AI Transformation Leader · Croissance et Transitions (SAS) · MATIA Method™ · Junyr Mail™ · Junyr Agents™

Sources: verified, July 2026

[1] a16z Podcast, "Rethinking Git for the Age of Coding Agents", interview with Scott Chacon (co-founder of GitHub, CEO of GitButler), 21 April 2026, Git's interface barely changed since 2005, version control to be rethought for humans and agents, "the best engineers of the future will be the best writers". https://a16z.com/podcast/rethinking-git-for-the-age-of-coding-agents-with-github-cofounder-scott-chacon/ [2] Joab Jackson, "Git is unprepared for the AI coding tsunami", The Register, 15 May 2026, +206% AI-generated projects on GitHub year-on-year (2025); 10.83 issues per pull request for generated code against 6.45 for human code; Scott Chacon: "run locally, mirrored globally". https://www.theregister.com/devops/2026/05/15/git-is-unprepared-for-the-ai-coding-tsunami/5241480 [3] Deepak Babu Piskala, "Spec-Driven Development: From Code to Contract in the Age of AI Coding Assistants", arXiv:2602.00180, 30 January 2026, the specification as the source of truth, code as a derived artefact; the three degrees spec-first / spec-anchored / spec-as-source. https://arxiv.org/abs/2602.00180 [4] Cursor (Anysphere), "Origin", a Git forge "for the agentic era", announced at the Compile conference on 17 June 2026, built for agents running in parallel (vendor demo: 22.6 commits per second on one repository), review designed as bulk approval; waitlist, release announced for autumn 2026. https://cursor.com/origin [5] Mengyang Liu, Taozhi Chen, Zhenhua Xu, Xue Jiang, Yihong Dong, "Multi-agent Collaboration with State Management" (the STORM framework, STate-ORiented Management), arXiv:2605.20563, 19 May 2026, concurrent agent writes validated against the current state of the shared workspace, stale writes rejected on the fly; +18.7 points on Commit0-Lite and +1.4 on PaperBench against Git-worktree isolation. https://arxiv.org/abs/2605.20563

The internal figures (50 plans, the batch of 12, lead times, checks) come from Junyr's development registry, read on 15 July 2026, four days after the switch described here.

Optimising context window size in LLMs: between cognitive dilution and hardware constraints

Sun, 12 Jul 2026 00:00:00 GMT

The context window size dilemma

Deploying large language models (LLMs) in production is governed by a fundamental trade-off between data-absorption capacity and computational efficiency. Contemporary architectures offer ever-larger context windows (from 128,000 tokens to more than a million for some frontier models), but simply expanding these windows quantitatively guarantees no proportional improvement in application-level performance.

Two opposing pitfalls await the engineer designing an LLM-based system:

Context insufficiency, which deprives the model of the factual anchors essential to answer accuracy and pushes it back onto its parametric knowledge, which is prone to hallucination.
Context dilution, a pathology where an excess of redundant or extraneous information saturates the transformer's attentional mechanisms and actively degrades reasoning quality, even when the sought-after information is genuinely present in the prompt.

This trade-off becomes critical in two very different operating contexts. For models run locally on resource-constrained hardware, managing the memory footprint of the Key-Value (KV) cache is a hard physical bottleneck: a GPU's VRAM is a finite resource, not a dial you can turn up by changing a setting. For frontier models billed by the token, ingesting superfluous data drives up cost and latency directly proportional to the injected noise. In both cases, optimising context window size is a matter of precision engineering, not empirical tuning.

The critical zone of context insufficiency

When a model is given too little information for the task at hand, it loses its ability to ground its answers in verifiable data. Absent explicit reference data in the prompt, the system falls back entirely on knowledge memorised in its parameters during pre-training, with three direct consequences for enterprise use:

Dependence on the training power law. A model's ability to reproduce a specific fact depends directly on how frequently documents associated with that fact appeared in its training corpus. "Long-tail" information (niche data, highly specialised facts, an SME's confidential technical documentation) is poorly represented there and so prone to outright retrieval failure.

Factual hallucination drift. Without grounding context, the model fills the gaps by generating statistically probable but factually wrong sequences of tokens. This is precisely the point of retrieval-augmented generation (RAG): reduce that dependency by injecting relevant documents into the prompt rather than relying on raw memorisation.

Failed semantic pivots. If the context window is too small to hold all the logical elements needed to solve a complex problem (the dependencies of a symbolic graph, the full history of a transaction), the model fails to establish the required connections, producing coherence breaks or incomplete answers.

Context level	Cognitive behaviour	Operational risks	Production usefulness
Meagre (< 1,000 tokens)	Exclusive reliance on parametric weights	High hallucination rate, stale facts, inability to handle private data	Low: limited to generic requests
Optimal (high density)	Precise alignment on re-injected sources, reasoning capabilities activated	Minimal risk of factual error, controlled latency	Maximal: precision RAG, specialised decision-making agents
Overloaded (> 32,000 tokens of noise)	Attentional dilution, "Lost in the Middle", reasoning shift	Latency explosion, accuracy degradation, prohibitive costs	Low: wasted hardware resources

Context dilution and the "Lost in the Middle" pathology

At the other end of the spectrum, massively and indiscriminately injecting data into the input window produces a cognitive pathology known as context dilution. Work by Nelson F. Liu and coauthors (Stanford; Liu et al., TACL 2024; originally published on arXiv in July 2023) revealed a U-shaped positional bias: language models are highly effective at exploiting data located right at the start or end of the context window, but their performance collapses when critical information sits in the middle of a lengthy prompt, even for models explicitly built for long context.

<figure style="margin:2rem 0;"> <svg viewBox="0 0 720 400" width="100%" height="auto" role="img" aria-labelledby="f1t-en f1d-en" style="max-width:100%;height:auto;font-family:system-ui,-apple-system,'Segoe UI',Roboto,sans-serif;"> <title id="f1t-en">Retrieval accuracy by position of relevant information in the context</title> <desc id="f1d-en">U-shaped curve illustrating the "Lost in the Middle" phenomenon: retrieval accuracy is high when the relevant information sits at the start or end of the context, and drops sharply in the middle.</desc> <text x="8" y="26" font-size="14" font-weight="700" fill="#0B2545">Retrieval accuracy by position of the relevant information</text> <text x="8" y="46" font-size="12" fill="#52514e">Schematic illustration of the U-shaped positional bias (Liu et al., TACL 2024)</text> <rect x="290" y="70" width="160" height="200" fill="#FF7A59" opacity="0.08"/> <line x1="60" y1="270" x2="680" y2="270" stroke="#c3c2b7" stroke-width="1"/> <line x1="60" y1="70" x2="60" y2="270" stroke="#c3c2b7" stroke-width="1"/> <path d="M 60,110 C 130,135 165,158 215,172 C 260,184 320,222 370,232 C 420,222 480,184 525,172 C 575,158 610,135 680,110" fill="none" stroke="#2563AC" stroke-width="2.5" stroke-linecap="round"/> <circle cx="60" cy="110" r="5" fill="#2563AC" stroke="#fcfcfb" stroke-width="2"/> <circle cx="370" cy="232" r="5" fill="#2563AC" stroke="#fcfcfb" stroke-width="2"/> <circle cx="680" cy="110" r="5" fill="#2563AC" stroke="#fcfcfb" stroke-width="2"/> <text x="60" y="98" font-size="12" font-weight="600" fill="#0b0b0b" text-anchor="start">High accuracy</text> <text x="370" y="255" font-size="12" font-weight="700" fill="#c1471f" text-anchor="middle">Low accuracy</text> <text x="370" y="270" font-size="11" fill="#c1471f" text-anchor="middle" dy="14">risk zone</text> <text x="680" y="98" font-size="12" font-weight="600" fill="#0b0b0b" text-anchor="end">High accuracy</text> <text x="60" y="292" font-size="11" fill="#898781" text-anchor="start">Start of context</text> <text x="370" y="292" font-size="11" fill="#898781" text-anchor="middle">Middle of context</text> <text x="680" y="292" font-size="11" fill="#898781" text-anchor="end">End of context</text> </svg> <figcaption style="font-size:0.8rem;color:#475569;margin-top:0.5rem;line-height:1.5;">Figure 1: A schematic curve (not measured values) illustrating the <em>Lost in the Middle</em> phenomenon: a model reliably retrieves information placed at the start or end of a prompt, but its accuracy degrades when that same information is buried in the centre of a long context. Source: Liu et al., <em>Lost in the Middle: How Language Models Use Long Contexts</em>, TACL 2024 (arXiv:2307.03172).</figcaption> </figure>

Several physical and structural factors explain this systematic degradation.

Zero-sum softmax normalisation. The self-attention mechanism computes similarity scores that are converted into probabilities via the softmax function, whose central property is that total attention allocated across the sequence must sum to exactly 1. Every irrelevant token injected into the prompt therefore "steals" a fraction of the attention that should have concentrated on the key information, degrading the signal-to-noise ratio of the internal representations.

Attention sinks. Xiao et al. ("Efficient Streaming Language Models with Attention Sinks", arXiv:2309.17453, ICLR 2024) show that models develop, during training, a tendency to over-allocate attention weights to the very first tokens of a sequence, regardless of their semantic relevance. These initial tokens act as a dumping ground for surplus attention, which artificially exacerbates the primacy bias at the expense of central data.

Semantic distraction. Introducing off-topic but thematically adjacent information (entities sharing similar roles, comparable ranges of numeric values) disrupts the selection of reasoning paths. The GSM-IC benchmark (Grade-School Math with Irrelevant Context, built on GSM8K by Shi et al., "Large Language Models Can Be Easily Distracted by Irrelevant Context", ICML 2023, arXiv:2302.00093) shows that arithmetic accuracy and the selection of solution trajectories drop sharply when facing this kind of interference.

Raw length-driven degradation. Even in the total absence of semantic distraction (and even when irrelevant tokens are masked or replaced with blanks), simply increasing input length imposes a cognitive tax on the model. A recent study (Du et al., "Context Length Alone Hurts LLM Performance Despite Perfect Retrieval", arXiv:2510.05381, Findings of EMNLP 2025; tested on Llama 3.1 8B Instruct, Mistral 7B v0.3 Instruct, GPT-4o, Claude 3.7 Sonnet and Gemini 2.0 across GSM8K, MMLU, HumanEval and a synthetic task) measures a performance degradation of 13.9% to 85% as prompt size grows, despite perfect retrieval.

Reasoning shift. Faced with large prompts containing long histories or dense background data, models trained for test-time-scaling reasoning sometimes shift behaviour in ways that are easy to miss. A 2026 preprint not yet peer-reviewed (Rodionov, Garipov and Yakushev, "Reasoning Shift: How Context Silently Shortens LLM Reasoning", arXiv:2604.01161) observes a reduction of up to 65% in the length of internal reasoning traces ("thinking traces") when the model is faced with context weighed down by irrelevant text, multi-turn conversations, or nested sub-tasks, without this necessarily degrading accuracy on easy tasks. This compression comes with a measurable drop in self-verification behaviour, which makes it more damaging on complex logical or mathematical problems. Treat this as an early, preliminary research signal rather than an established result.

<figure style="margin:2rem 0;"> <svg viewBox="0 0 720 400" width="100%" height="auto" role="img" aria-labelledby="f2t-en f2d-en" style="max-width:100%;height:auto;font-family:system-ui,-apple-system,'Segoe UI',Roboto,sans-serif;"> <title id="f2t-en">Response reliability by size of injected context</title> <desc id="f2d-en">Bell-shaped curve: reliability is low with insufficient context, reaches an optimal plateau with dense, tightly-scoped context, then degrades once context becomes overloaded and noisy.</desc> <text x="8" y="26" font-size="14" font-weight="700" fill="#0B2545">Response reliability by size of injected context</text> <text x="8" y="46" font-size="12" fill="#52514e">The "Goldilocks zone": neither meagre nor overloaded · dense and targeted</text> <rect x="60" y="70" width="200" height="200" fill="#898781" opacity="0.08"/> <rect x="260" y="70" width="220" height="200" fill="#2563AC" opacity="0.08"/> <rect x="480" y="70" width="200" height="200" fill="#FF7A59" opacity="0.08"/> <line x1="60" y1="270" x2="680" y2="270" stroke="#c3c2b7" stroke-width="1"/> <line x1="60" y1="70" x2="60" y2="270" stroke="#c3c2b7" stroke-width="1"/> <path d="M 60,225 C 130,210 175,145 230,120 C 280,100 320,92 370,90 C 420,92 470,102 520,125 C 570,150 610,185 680,215" fill="none" stroke="#2563AC" stroke-width="2.5" stroke-linecap="round"/> <circle cx="60" cy="225" r="5" fill="#898781" stroke="#fcfcfb" stroke-width="2"/> <circle cx="370" cy="90" r="5" fill="#2563AC" stroke="#fcfcfb" stroke-width="2"/> <circle cx="680" cy="215" r="5" fill="#FF7A59" stroke="#fcfcfb" stroke-width="2"/> <text x="160" y="290" font-size="12" font-weight="600" fill="#52514e" text-anchor="middle">Insufficient</text> <text x="160" y="305" font-size="10" fill="#898781" text-anchor="middle">< 1,000 tokens</text> <text x="370" y="290" font-size="12" font-weight="700" fill="#184f95" text-anchor="middle">Optimal</text> <text x="370" y="305" font-size="10" fill="#898781" text-anchor="middle">dense, targeted, reranked</text> <text x="580" y="290" font-size="12" font-weight="600" fill="#c1471f" text-anchor="middle">Overloaded</text> <text x="580" y="305" font-size="10" fill="#898781" text-anchor="middle">> 32,000 tokens of noise</text> </svg> <figcaption style="font-size:0.8rem;color:#475569;margin-top:0.5rem;line-height:1.5;">Figure 2: Response reliability is not monotonic with context size: it climbs as useful factual anchors are added, plateaus in a dense, targeted zone, then falls again once the volume of noise outweighs the volume of signal. A pedagogical illustration, not a single empirical measurement.</figcaption> </figure>

Hardware impact on local LLMs: hitting the VRAM wall and the physics of the KV cache

For practitioners running LLMs locally, context window size isn't just a logical tuning parameter: it's a hard hardware ceiling, set by the video memory (VRAM) available on the graphics card. During inference, a model's memory footprint has two components: a static load (the model's weights) and a highly volatile dynamic load, the Key-Value (KV) cache.

For every token processed, the transformer architecture computes and stores Key and Value vectors for each attention layer. This cache avoids recomputing the full set of attention relationships quadratically (O(n²)) on every new generated token, bringing the decoding phase down to linear complexity (O(n)). But its size grows strictly linearly with sequence length and batch size.

The memory footprint of the KV cache is calculated precisely via the following formula:

M_KV = 2 × L × H_kv × D_head × N_seq × B × P_bytes

where L is the model's number of attention layers, H_kv the number of attention heads allocated to keys and values (sharply reduced in modern architectures using Grouped-Query Attention, GQA, compared with classic multi-head attention), D_head the dimension of each head, N_seq the cumulative sequence length, B the batch size, and P_bytes the number of bytes per element depending on precision (2 bytes for FP16/BF16, 1 byte for INT8, 0.5 byte for INT4). The multiplier 2 corresponds to the separate storage of the Key and Value tensors.

As an illustration, Llama 3.1 8B Instruct exposes a GQA configuration of 32 layers, 8 KV heads and a head dimension of 128 [6]. For a single request (batch = 1) over a 32,768-token context in native BF16, the calculation gives:

M_KV = 2 × 32 × 8 × 128 × 32,768 × 1 × 2 bytes = 4,294,967,296 bytes = 4 GiB, exactly.

This calculation illustrates why the residual space available for the KV cache shrinks dramatically the moment you run a larger model on a consumer GPU with 12 or 16 GB of VRAM, or significantly extend the context. The moment the sum of model weights and KV cache exceeds physical VRAM capacity, execution spills over into the CPU's system memory, a crossover from the GPU's ultra-fast HBM to conventional DRAM access lanes, whose bandwidth is an order of magnitude lower. Decoding speed then collapses almost instantly.

Reference model	Weights (quantisation)	Context	KV cache (calculated)	Estimated total	Status on a consumer GPU (RTX 4090, 24 GB)
Llama 3.1 8B Instruct [6]	~16 GB (native BF16)	4,096 tokens	~0.5 GB	~16.5 GB	Optimal operation, comfortable margin
Llama 3.1 8B Instruct [6]	~16 GB (native BF16)	32,768 tokens	4 GB (exact calculation above)	~20 GB	Tight but stable
Phi-4, 14B [7]	~8-9 GB (INT4)	16,384 tokens (maximum native context)	~2.5 GB (BF16)	~11.5 GB	Comfortable
Qwen 3 32B [8]	~19.8 GB (GGUF Q4_K_M, weights only)	40,960 tokens (native context)	10 GiB ≈ 10.7 GB (calculation: 2×64×8×128×2 B = 256 KiB/token)	~30.5 GB	Already over budget on a 24 GB GPU, weights alone near the limit; requires more aggressive quantisation (Q3/Q2), at the cost of quality
Llama 3.3 70B [9]	~40-46 GB (INT4, method-dependent)	128,000 tokens	≈ 39 GiB ≈ 42 GB (BF16)	~75-90 GB	Impossible without enterprise-class hardware (multi-GPU or data-centre-class VRAM)

To work around this bottleneck, KV cache quantisation stands out as an essential technique: converting Keys and Values to INT8 or INT4 mechanically reduces memory usage by 50% to 75% (a direct consequence of the P_bytes term in the formula above), for a perplexity degradation generally described as minor across most tasks. The figure most commonly cited is 1% to 3%, though no single source is authoritative for this exact range across every architecture.

To optimise these savings without excessively degrading generation fidelity, one avenue explored by research is to quantise deeper layers of the network more aggressively than the earlier ones, a depth-based adaptation documented notably by PyramidKV [11], which shows that a reduced cache budget in the upper layers affects quality less than in the early layers. This is a distinct approach from the per-channel/per-token 2-bit quantisation proposed by KIVI [10], which targets precision rather than depth. The two techniques are complementary rather than competing.

<figure style="margin:2rem 0;"> <svg viewBox="0 0 720 340" width="100%" height="auto" role="img" aria-labelledby="f3t-en f3d-en" style="max-width:100%;height:auto;font-family:system-ui,-apple-system,'Segoe UI',Roboto,sans-serif;"> <title id="f3t-en">VRAM footprint: model weights and KV cache, by scenario</title> <desc id="f3d-en">Horizontal stacked bar chart showing, for five model/context scenarios, the share taken by model weights and by KV cache, compared against a 24 GB RTX 4090 ceiling.</desc> <text x="8" y="20" font-size="14" font-weight="700" fill="#0B2545">VRAM footprint: model weights + KV cache, by scenario</text> <text x="8" y="38" font-size="11" fill="#52514e">Cumulative bars, in gigabytes · compared against a consumer GPU ceiling (24 GB)</text> <line x1="350" y1="52" x2="350" y2="278" stroke="#0b0b0b" stroke-width="1.5"/> <text x="354" y="62" font-size="10.5" font-weight="600" fill="#0b0b0b">RTX 4090 · 24 GB</text> <line x1="230" y1="278" x2="680" y2="278" stroke="#c3c2b7" stroke-width="1"/> <text x="230" y="292" font-size="9.5" fill="#898781">0</text> <text x="350" y="292" font-size="9.5" fill="#898781" text-anchor="middle">24</text> <text x="455" y="292" font-size="9.5" fill="#898781" text-anchor="middle">50 GB</text> <text x="680" y="292" font-size="9.5" fill="#898781" text-anchor="end">90 GB</text> <text x="222" y="76" font-size="11" fill="#0b0b0b" text-anchor="end">Llama 3.1 8B · 4k</text> <rect x="230" y="64" width="80" height="22" fill="#2563AC"/> <rect x="311" y="64" width="3" height="22" fill="#FF7A59"/> <text x="318" y="79" font-size="10.5" font-weight="600" fill="#0b0b0b">16.5 GB</text> <text x="222" y="120" font-size="11" fill="#0b0b0b" text-anchor="end">Llama 3.1 8B · 32k</text> <rect x="230" y="108" width="80" height="22" fill="#2563AC"/> <rect x="311" y="108" width="20" height="22" fill="#FF7A59"/> <text x="336" y="123" font-size="10.5" font-weight="600" fill="#0b0b0b">20 GB</text> <text x="222" y="164" font-size="11" fill="#0b0b0b" text-anchor="end">Phi-4 14B · 16k (native)</text> <rect x="230" y="152" width="45" height="22" fill="#2563AC"/> <rect x="276" y="152" width="13" height="22" fill="#FF7A59"/> <text x="294" y="167" font-size="10.5" font-weight="600" fill="#0b0b0b">11.5 GB</text> <text x="222" y="208" font-size="11" fill="#0b0b0b" text-anchor="end">Qwen 3 32B · 40k (native)</text> <rect x="230" y="196" width="99" height="22" fill="#2563AC"/> <rect x="330" y="196" width="53" height="22" fill="#FF7A59"/> <text x="388" y="211" font-size="10.5" font-weight="600" fill="#0b0b0b">≈ 30.5 GB</text> <text x="222" y="252" font-size="11" fill="#0b0b0b" text-anchor="end">Llama 3.3 70B · 128k</text> <rect x="230" y="240" width="210" height="22" fill="#2563AC"/> <rect x="441" y="240" width="210" height="22" fill="#FF7A59"/> <text x="656" y="255" font-size="10.5" font-weight="700" fill="#0b0b0b" text-anchor="end">≈ 84 GB</text> <rect x="230" y="312" width="14" height="14" fill="#2563AC"/> <text x="248" y="323" font-size="11" fill="#52514e">Model weights</text> <rect x="380" y="312" width="14" height="14" fill="#FF7A59"/> <text x="398" y="323" font-size="11" fill="#52514e">KV cache</text> </svg> <figcaption style="font-size:0.8rem;color:#475569;margin-top:0.5rem;line-height:1.5;">Figure 3: The moment the cumulative bar crosses the 24 GB (RTX 4090) line, the model no longer fits entirely in VRAM without CPU offload or more aggressive quantisation. Detailed calculations in the text and sources [6]-[9].</figcaption> </figure>

Financial impact on frontier models: prompt caching and economies of scale

For teams using frontier-model APIs, every token sent in the input window represents a direct financial charge. The standard billing model imputes a cost for processing the prompt (the prefill phase) and a higher cost for generating output tokens (decode). In autonomous agent systems, enterprise RAG, or coding assistants, repeatedly ingesting long, identical context (typing rules, a reference codebase, system instructions) generates massive budget waste if nothing is done to prevent it.

Prompt caching consists of retaining, server-side, the KV cache corresponding to the pre-fill of a stable prompt prefix. When successive requests share that same prefix, the API bills processing those tokens at a heavily reduced rate. Access and pricing conditions vary considerably from one provider to the next.

Anthropic requires explicit cache_control markers to be inserted into the request structure to indicate which sections of the context should be cached. It is a manual approach that offers fine-grained control, with an eligibility threshold that varies by model (2,048 tokens minimum for Claude Sonnet 4.6, 4,096 tokens for the Opus family). A 90% discount applies to tokens read from cache, with a modest write penalty: a 25% surcharge on a prefix cached for the first time (5-minute TTL), amortised from the second identical request onward.

OpenAI automatically detects prefix matches on incoming requests, with no marker and no caching surcharge: if the first 1,024 tokens of the prompt match a sequence submitted recently, the system applies a 90% discount with no application-code changes required. This mechanism, confirmed on GPT-5.5 (April 2026), has since evolved: as of 9 July 2026, GPT-5.6 (which succeeded GPT-5.5 just days before this article's publication) introduces a 1.25× cache-write surcharge, ending fully free caching at OpenAI. A signal worth watching: providers' pricing terms shift almost as fast as the models themselves.

Google actually runs two distinct mechanisms, a nuance often overlooked. Automatic implicit caching applies by default from 4,096 tokens of shared prefix (no configuration required), while explicit caching, activated manually via the API, guarantees a fixed retention period (TTL) for large, stable RAG contexts, with a considerably higher eligibility threshold, starting at 32,768 tokens. On Gemini 3.1 Pro (still in Preview as of mid-2026), the discount reaches 90% in both cases for prompts up to 200,000 tokens.

DeepSeek offers the most aggressive mechanism: prefix detection is automatic and free to write, with a minimal eligibility threshold of just 64 tokens, far more permissive than the four-digit thresholds of other providers. On DeepSeek-V4-Flash, a cache read costs $0.0028 per million tokens against $0.14 for a standard write, a 98% discount.

Provider	Reference model	Standard input price (/1M tokens)	Cached input price (/1M tokens)	Discount	Activation	Minimum threshold
Anthropic	Claude Sonnet 4.6	$3.00	$0.30	-90%	Manual (`cache_control`)	2,048 tokens
OpenAI	GPT-5.5 (April 2026)	$5.00	$0.50	-90%	Automatic (prefix)	1,024 tokens
Google	Gemini 3.1 Pro (Preview)	$2.00 (≤ 200K)	$0.20	-90%	Automatic (implicit) or manual (explicit, guaranteed TTL)	4,096 tokens (implicit) / 32,768 tokens (explicit)
DeepSeek	DeepSeek-V4-Flash	$0.14	$0.0028	-98%	Automatic (prefix)	64 tokens

<figure style="margin:2rem 0;"> <svg viewBox="0 0 720 320" width="100%" height="auto" role="img" aria-labelledby="f4t-en f4d-en" style="max-width:100%;height:auto;font-family:system-ui,-apple-system,'Segoe UI',Roboto,sans-serif;"> <title id="f4t-en">Discount applied to tokens served from cache, by provider</title> <desc id="f4d-en">Column chart showing the percentage discount on input tokens served from cache for four providers: Anthropic, OpenAI, Google and DeepSeek.</desc> <text x="8" y="20" font-size="14" font-weight="700" fill="#0B2545">Discount on input tokens served from cache</text> <text x="8" y="38" font-size="11" fill="#52514e">By provider · reference model, mid-2026</text> <line x1="60" y1="260" x2="680" y2="260" stroke="#c3c2b7" stroke-width="1"/> <rect x="148" y="89" width="24" height="171" rx="4" fill="#2563AC"/> <text x="160" y="80" font-size="12" font-weight="700" fill="#0b0b0b" text-anchor="middle">-90%</text> <text x="160" y="278" font-size="11" font-weight="600" fill="#0b0b0b" text-anchor="middle">Anthropic</text> <text x="160" y="292" font-size="9.5" fill="#898781" text-anchor="middle">Claude Sonnet 4.6</text> <rect x="308" y="89" width="24" height="171" rx="4" fill="#2563AC"/> <text x="320" y="80" font-size="12" font-weight="700" fill="#0b0b0b" text-anchor="middle">-90%</text> <text x="320" y="278" font-size="11" font-weight="600" fill="#0b0b0b" text-anchor="middle">OpenAI</text> <text x="320" y="292" font-size="9.5" fill="#898781" text-anchor="middle">GPT-5.5</text> <rect x="468" y="89" width="24" height="171" rx="4" fill="#2563AC"/> <text x="480" y="80" font-size="12" font-weight="700" fill="#0b0b0b" text-anchor="middle">-90%</text> <text x="480" y="278" font-size="11" font-weight="600" fill="#0b0b0b" text-anchor="middle">Google</text> <text x="480" y="292" font-size="9.5" fill="#898781" text-anchor="middle">Gemini 3.1 Pro</text> <rect x="628" y="74" width="24" height="186" rx="4" fill="#FF7A59"/> <text x="640" y="65" font-size="12" font-weight="700" fill="#c1471f" text-anchor="middle">-98%</text> <text x="640" y="278" font-size="11" font-weight="600" fill="#0b0b0b" text-anchor="middle">DeepSeek</text> <text x="640" y="292" font-size="9.5" fill="#898781" text-anchor="middle">DeepSeek-V4-Flash</text> </svg> <figcaption style="font-size:0.8rem;color:#475569;margin-top:0.5rem;line-height:1.5;">Figure 4: The order of magnitude of the discount (90% and above) is comparable across providers; what varies sharply is the eligibility threshold and activation method (table above). DeepSeek stands out with a deeper discount (-98%) on a base price that is already far lower.</figcaption> </figure>

To make the most of these pricing structures, the semantic layout of requests must be rigorously ordered: global instructions, stable background data and output schemas at the top of the prompt; dynamic content (timestamps, user IDs, the user's question) at the end of the message. Any change to the prefix, however small, invalidates the downstream cache and triggers a costly cache miss. By combining carefully ordered requests with batch-processing pipelines (Batch APIs, which grant an additional flat discount at several providers), teams can meaningfully cut their inference bill.

Towards an optimal context architecture: reranking, compression and multi-agent systems

To resolve the tension between the need for factual data and the risk of attentional dilution or hardware overspend, modern architectures apply successive refinement passes to the incoming information flow, rather than saturating the context window with a brute-force approach.

Precision reranking. In a standard RAG system, vector embedding models perform a first, fast pass at selecting relevant documents, but these bi-encoder models don't analyse the fine-grained interaction between the question and each individual passage. A cross-encoder reranker model precisely scores the relative relevance of each excerpt against the question asked, keeps only the highest-scoring passages, then deliberately places the most critical excerpts at the very start and end of the prompt, leaving the centre of the window empty of any medium-importance content, directly working around the Lost in the Middle phenomenon documented in Figure 1.

Algorithmic prompt compression. Three tools in the LLMLingua family (Microsoft Research), often conflated, address distinct needs:

LLMLingua [16] (the original version) uses a small local language model to compute the perplexity of each segment of the prompt: highly predictable words (low perplexity) are dropped, while syntactic and semantic pivots carrying meaning (high perplexity) are preserved. Microsoft Research's benchmarks report compression of up to 20× with limited performance loss.
LLMLingua-2 [17] abandons the perplexity-based approach in favour of a binary (keep/drop) BERT-style classifier, trained via data distillation from GPT-4. Explicitly built to be task-agnostic (it compresses without ever seeing the final question), it targets a more modest compression range (roughly 2× to 5×) but is faster and more robust than the perplexity-based approach.
LongLLMLingua [18] is the variant built specifically for long-context and RAG scenarios: unlike LLMLingua-2, it is question-aware. It reorders and coarsely filters document segments by their relevance to the question asked, ahead of fine-grained compression, and includes an explicit document-reordering strategy to reduce positional bias. Alongside reranking, it is one of the few techniques designed to tackle the Lost in the Middle phenomenon head-on.

Multi-agent topologies. To work around the degradation of self-verification capabilities and the compression of reasoning traces induced by large contexts (see the previous section), a complex task can be split across a network of specialised agents. Instead of routing one global request to a single model equipped with a window spanning hundreds of thousands of tokens, the problem is broken into independent sub-goals handed to autonomous instances, each receiving an ultra-targeted, tightly filtered context, ideally under the 25,000-token mark, the "optimal" zone identified in Figure 2. Orchestrating these agents and synthesising their output aims for processing that is both complete and robust on logical accuracy.

<figure style="margin:2rem 0;"> <svg viewBox="0 0 720 360" width="100%" height="auto" role="img" aria-labelledby="f5t-en f5d-en" style="max-width:100%;height:auto;font-family:system-ui,-apple-system,'Segoe UI',Roboto,sans-serif;"> <title id="f5t-en">Multi-agent architecture: splitting context across specialised agents</title> <desc id="f5d-en">Architecture diagram: an orchestrator delegates a complex task to three specialised agents, each receiving a context under 25,000 tokens, ahead of a final synthesis step.</desc> <text x="360" y="20" font-size="14" font-weight="700" fill="#0B2545" text-anchor="middle">Splitting context rather than concentrating it</text> <rect x="290" y="34" width="140" height="46" rx="6" fill="#fcfcfb" stroke="#0B2545" stroke-width="1.5"/> <text x="360" y="62" font-size="12.5" font-weight="700" fill="#0B2545" text-anchor="middle">Orchestrator</text> <line x1="330" y1="80" x2="130" y2="140" stroke="#898781" stroke-width="1.5"/> <line x1="360" y1="80" x2="360" y2="140" stroke="#898781" stroke-width="1.5"/> <line x1="390" y1="80" x2="590" y2="140" stroke="#898781" stroke-width="1.5"/> <rect x="40" y="140" width="180" height="80" rx="6" fill="#fcfcfb" stroke="#0B2545" stroke-width="1.5"/> <text x="130" y="166" font-size="12" font-weight="600" fill="#0B2545" text-anchor="middle">Agent A</text> <text x="130" y="182" font-size="10.5" fill="#52514e" text-anchor="middle">targeted sub-task</text> <rect x="80" y="192" width="100" height="18" rx="9" fill="#FF7A59"/> <text x="130" y="205" font-size="9.5" font-weight="700" fill="#fcfcfb" text-anchor="middle">< 25,000 tokens</text> <rect x="270" y="140" width="180" height="80" rx="6" fill="#fcfcfb" stroke="#0B2545" stroke-width="1.5"/> <text x="360" y="166" font-size="12" font-weight="600" fill="#0B2545" text-anchor="middle">Agent B</text> <text x="360" y="182" font-size="10.5" fill="#52514e" text-anchor="middle">targeted sub-task</text> <rect x="310" y="192" width="100" height="18" rx="9" fill="#FF7A59"/> <text x="360" y="205" font-size="9.5" font-weight="700" fill="#fcfcfb" text-anchor="middle">< 25,000 tokens</text> <rect x="500" y="140" width="180" height="80" rx="6" fill="#fcfcfb" stroke="#0B2545" stroke-width="1.5"/> <text x="590" y="166" font-size="12" font-weight="600" fill="#0B2545" text-anchor="middle">Agent C</text> <text x="590" y="182" font-size="10.5" fill="#52514e" text-anchor="middle">targeted sub-task</text> <rect x="540" y="192" width="100" height="18" rx="9" fill="#FF7A59"/> <text x="590" y="205" font-size="9.5" font-weight="700" fill="#fcfcfb" text-anchor="middle">< 25,000 tokens</text> <line x1="130" y1="220" x2="330" y2="286" stroke="#898781" stroke-width="1.5"/> <line x1="360" y1="220" x2="360" y2="286" stroke="#898781" stroke-width="1.5"/> <line x1="590" y1="220" x2="390" y2="286" stroke="#898781" stroke-width="1.5"/> <rect x="270" y="286" width="180" height="50" rx="6" fill="#0B2545"/> <text x="360" y="316" font-size="12.5" font-weight="700" fill="#fcfcfb" text-anchor="middle">Final synthesis</text> </svg> <figcaption style="font-size:0.8rem;color:#475569;margin-top:0.5rem;line-height:1.5;">Figure 5: Each agent receives a tightly filtered subset of the overall problem, preserving its self-verification and reasoning capabilities rather than diluting them across a single, massive context.</figcaption> </figure>

Optimisation approach	Goal	Trade-off / implementation cost
Reranking	Eliminate noise, position the essentials at the edges of the prompt	Additional latency from the cross-encoder pass
Prompt compression (LLMLingua) [16][17]	Eliminate linguistic redundancy via perplexity analysis or token-by-token classification	Requires running a small compression model upstream; up to 20× for LLMLingua, 2-5× for LLMLingua-2
Question-aware compression (LongLLMLingua) [18]	Reorder and filter segments by their relevance to the question asked	Adds a per-request scoring step; specifically targets positional bias
Multi-agent architecture	Split a massive context across specialised agents with limited memory (ideally < 25,000 tokens)	Increased orchestration complexity and message-routing overhead

Conclusions and strategic recommendations

Managing context window size has become a pillar of AI systems engineering. Haphazardly saturating a transformer's input window damages its real cognitive capabilities: attention is a zero-sum resource, sensitive to positional bias and vulnerable to semantic interference. At the opposite extreme, excessively reducing context produces factual drift and hallucination.

Reconciling logical accuracy, the responsiveness of local systems, and the profitability of distributed architectures requires dynamic, qualitative context management: reranking to place critical information at the edges of the sequence, compression to eliminate unnecessary linguistic redundancy, and multi-agent architectures to break large tasks into low-memory-footprint sub-problems. By rigorously structuring input data to make the most of prompt-caching technologies, teams can build solutions that are robust and high-value, on both the hardware and the financial side.

That is exactly the logic behind the MATIA Method™: context window size isn't a dial you push to the maximum by reflex. It's an engineering variable you size to the task.

Going further

The AI Express Audit (a 60-minute video session, free, no obligation) includes a review of your context and inference-cost architecture: context window, KV cache, prompt caching and model choice.

On the FinOps side (token budgets, LLM gateways, dynamic routing), see Token budgets and AI APIs: the FinOps guide for SMEs (French original; English summary available on request).

The white paper "AI Maturity of French SMEs 2025-2026" is available at croissance-transitions.fr.

Sources

[1] Nelson F. Liu, Kevin Lin, John Hewitt, Ashwin Paranjape, Michele Bevilacqua, Fabio Petroni, Percy Liang, Lost in the Middle: How Language Models Use Long Contexts (TACL 2024, vol. 12, pp. 157-173; arXiv:2307.03172, July 2023), U-shaped positional bias in long-context information retrieval. https://arxiv.org/abs/2307.03172

[2] Guangxuan Xiao, Yuandong Tian, Beidi Chen, Song Han, Mike Lewis, Efficient Streaming Language Models with Attention Sinks (ICLR 2024; arXiv:2309.17453, Sept. 2023), over-allocation of attention to the first tokens of a sequence. https://arxiv.org/abs/2309.17453

[3] Freda Shi, Xinyun Chen, Kanishka Misra, Nathan Scales, David Dohan, Ed H. Chi, Nathanael Schärli, Denny Zhou, Large Language Models Can Be Easily Distracted by Irrelevant Context (ICML 2023; arXiv:2302.00093, Feb. 2023), GSM-IC benchmark, semantic distraction. https://arxiv.org/abs/2302.00093

[4] Yufeng Du, Minyang Tian, Srikanth Ronanki, Subendhu Rongali, Sravan Bodapati, Aram Galstyan, Azton Wells, Roy Schwartz, Eliu A. Huerta, Hao Peng, Context Length Alone Hurts LLM Performance Despite Perfect Retrieval (Findings of EMNLP 2025; arXiv:2510.05381, Oct. 2025), 13.9% to 85% degradation despite perfect retrieval, tested across 5 models. https://arxiv.org/abs/2510.05381

[5] Gleb Rodionov, Roman Garipov, George Yakushev, Reasoning Shift: How Context Silently Shortens LLM Reasoning (preprint, not yet peer-reviewed; arXiv:2604.01161, April-June 2026), reduction of up to 65% in internal reasoning traces under heavy context. Preliminary result, treat with caution. https://arxiv.org/abs/2604.01161

[6] Meta, Llama 3.1 8B Instruct, model configuration (32 layers, 8 GQA KV heads, head dimension 128); KV cache calculation derived from this configuration.

[7] Microsoft, Phi-4 Technical Report (arXiv:2412.08905, Dec. 2024), 14B parameters, 40 layers, 8 KV heads, native context of 16,384 tokens.

[8] Alibaba, Qwen3-32B, model card (April 2025, Apache 2.0 licence; 64 layers, 8 KV heads, head dimension 128, native context of 40,960 tokens).

[9] Meta, Llama 3.3 70B, model configuration (Dec. 2024; 80 layers, 8 KV heads, head dimension 128, 128,000-token context).

[10] Zirui Liu et al., KIVI: A Tuning-Free Asymmetric 2bit Quantization for KV Cache (arXiv:2402.02750, Feb. 2024), 2-bit quantisation, per-channel (keys) / per-token (values).

[11] Zefan Zhang et al., PyramidKV: Dynamic KV Cache Compression based on Pyramidal Information Funneling (arXiv:2406.02069, 2024), cache compression adapted to layer depth.

[12] Anthropic documentation, Prompt Caching, per-model thresholds, 90% read discount, 25% write surcharge (5-min TTL) or 100% (1h TTL).

[13] OpenAI documentation, Prompt Caching + GPT-5.6 announcement (9 July 2026), automatic prefix detection, introduction of a 1.25× write surcharge on GPT-5.6.

[14] Google DeepMind, Gemini 3.1 Pro, model card (19 February 2026, Preview status); implicit and explicit caching documentation.

[15] DeepSeek documentation, pricing and context caching, DeepSeek-V4-Flash (April 2026), minimum threshold of 64 tokens.

[16] Huiqiang Jiang, Qianhui Wu, Chin-Yew Lin, Yuqing Yang, Lili Qiu, LLMLingua: Compressing Prompts for Accelerated Inference of Large Language Models (EMNLP 2023; arXiv:2310.05736), perplexity-based compression, up to 20×. https://arxiv.org/abs/2310.05736

[17] Zhuoshi Pan, Qianhui Wu, Huiqiang Jiang et al., LLMLingua-2: Data Distillation for Efficient and Faithful Task-Agnostic Prompt Compression (ACL 2024 Findings; arXiv:2403.12968), GPT-4-distilled BERT classifier, task-agnostic. https://arxiv.org/abs/2403.12968

[18] Huiqiang Jiang, Qianhui Wu, Xufang Luo, Dongsheng Li, Chin-Yew Lin, Yuqing Yang, Lili Qiu, LongLLMLingua: Accelerating and Enhancing LLMs in Long Context Scenarios via Prompt Compression (arXiv:2310.06839, 2024), question-aware compression and document reordering. https://arxiv.org/abs/2310.06839

Written by Paul-Antoine TUAL, AI Transformation Leader, creator of the MATIA Method™.

Engineering agentic systems: golden rules, architecture and security of the “Human-in-the-Loop”

Sat, 04 Jul 2026 00:00:00 GMT

When a machine acts without a guardrail

On 1 August 2012, it took just forty-five minutes for the brokerage firm Knight Capital to lose approximately $440 million. A faulty software deployment had reactivated obsolete code on a stock-order router, which began flooding the market with erroneous transactions, without any human mechanism being able to stop it in time [1]. The company, one of the largest American market makers, never recovered. This catastrophe is not a story about artificial intelligence, but it recounts exactly what is at stake today: the cost of an automated execution devoid of real-time operational guardrails.

The evolution of AI has set in motion a profound structural transition: the shift from passive systems, geared towards generating responses, to autonomous agents oriented towards action. Unlike conventional language-processing interfaces, modern agentic architectures design and execute multi-step action plans, interact directly with digital and physical environments, and handle critical tools: production databases, financial transaction gateways, industrial systems. This deployment relies on emerging protocols: Anthropic's Model Context Protocol (MCP) [3] and agent-to-agent (A2A) communication protocols [4]. They allow a system to discover new capabilities dynamically at execution time.

This increased autonomy comes with a considerable widening of the attack surface and of operational risks. Yet language models have a documented flaw: they are systematically overconfident. Work on their calibration shows that a model commonly verbalises a confidence of 90% or more where its actual accuracy is markedly lower, and that hallucinations frequently occur with a high displayed level of certainty [2]. An agent can therefore generate an erroneous action plan, or even a destructive hallucination, while declaring itself perfectly confident. To underestimate this failure mode is to reproduce, at software scale, the Knight Capital scenario.

To remedy these vulnerabilities, the integration of human-control protocols, Human-in-the-Loop (HITL), must not be conceived as a mere layer of ergonomic validation, but as a fundamental engineering constraint. The level of control must, however, be matched to the criticality of the operations, the reversibility of the actions and the regulatory imperatives.

Three supervision regimes: HITL, HOTL, HOOTL

The first architectural decision consists in choosing, for each type of action, who decides and when. The literature on autonomous systems distinguishes three supervision regimes, a taxonomy inherited from the debate on weapons systems and transposable to software agents [5].

Regime	Temporality	Autonomy	Recommended use cases	Cognitive load	Operational & HR challenges
Human-in-the-Loop (HITL)	Synchronous interruption: the agent freezes its execution and awaits explicit human approval.	Low: the agent proposes, but the physical execution of the action is blocked.	Critical and irreversible operations: transfers, database modifications, medical diagnoses.	High: continuous assessment, context analysis, manual validation.	Requires dedicated teams of approvers, available to avoid bottlenecks.
Human-on-the-Loop (HOTL)	Asynchronous supervision: the agent runs on its own but under continuous monitoring.	Moderate to high: the agent acts within predefined limits.	Medium criticality, high volumes: booking modifications, scheduling, status updates.	Moderate: alerts triggered when a threshold is crossed or on abnormal behaviour.	A single supervisor can steer and audit several agents in parallel.
Human-out-of-the-Loop (HOOTL)	After-the-fact audit: no real-time intervention.	Maximum: the agent is sovereign within its scope.	Low risk, highly standardised: call routing, answers to frequently asked questions.	Low: periodic analysis of the logs and re-training.	Requires regular compliance auditors and offline annotators.

None of these regimes is “better” in absolute terms: the right system combines several of them, action by action. What remains is to determine how, concretely, to structure the agent's behaviour so as to preserve human authority.

The golden rules of Human-in-the-Loop

Rule 1. Separate the proposal of an action from its execution

An agent's architecture must never grant it the ability to execute, directly and unilaterally, an action with irreversible effects on its environment. The agent must produce a documented action proposal (intention, parameters, estimated consequences, source data used) and submit it to an isolated execution engine. Physical execution remains suspended until a qualified operator has approved the transaction. This separation protects the infrastructure against malicious request injections as well as against the model's erroneous inferences.

Rule 2. Structure the interruption points before writing the code

Human control points are not added after the fact, as a catch-up. Engineers must map high-risk transactions from the design stage. This structuring relies on quantitative intervention thresholds that automatically trigger the pausing of the system: spending ceilings, data sensitivity levels, regulatory criteria (GxP, HIPAA, GDPR).

Rule 3. Make the agent's reasoning entirely transparent

The operator's trust depends on the visibility of the steps leading to a recommendation. The agent must set out, in an intelligible manner, the data it relies on, the business rules applied and the alternatives discarded. This transparency is best delivered through accessible interfaces (of the low-code or no-code kind) so that non-technical teams instantly understand the reason for an escalation and correct the agent's trajectory without reading raw code.

Rule 4. Codify the non-negotiable limits deterministically

This is the most counter-intuitive rule, and the most important. An agent built on a probabilistic model cannot guarantee the observance of a security constraint formulated solely in natural language. Non-negotiable limits (restricted access to production databases, budget ceilings, the scope of authorised tools) must be encoded deterministically in the software harness surrounding the model. The system incorporates strict schema validation (typically via a validator such as Pydantic) and imperative rules that block a non-compliant request before it reaches the model's interface or an external API. The guardrail lives in the code, not in the prompt.

Rule 5. Design for progressive automation

The implementation must begin at the maximum level of human intervention. This conservative start makes it possible to record decisions, to analyse the divergences between the model's proposals and human corrections, and to measure the system's reliability precisely. As the metrics improve and the models become calibrated, the pause criteria are relaxed: the system evolves from synchronous control (HITL) towards supervision by exception (HOTL). Automation must be earned; it is never the starting point.

Rule 6. Anchor governance in five obligations: the “HEART” memo

To place human-machine collaboration within a responsible approach, it is useful to summarise the control obligations under a mnemonic acronym, HEART, which brings together five pillars, each independently recognised as a principle of AI governance:

Human safety & accountability: human safety and ultimate legal responsibility must rest with a natural person.
Explainability: the explainability of every structuring choice, in order to rule out the “black box” effect.
Alignment: the continuous alignment of the agent's operations with the business objectives and the company's ethical charter.
Review: the regular audit and reassessment of performance and emergent behaviours.
Trackability: the absolute technical traceability of the machine's actions and of human validation decisions.

(HEART is here a mnemonic grouping device, not an established normative standard: the five obligations, for their part, are classic governance requirements.)

Rule 7. Manage escalation levels dynamically

The interruption system must adapt to the nature of the request. When it freezes its execution, the agent must be able to distinguish a simple approval (binary validation), a structured request for clarification (when an essential variable is missing), or the delegation of execution to a third-party system. Operator feedback must be guided by structured feedback prompts (to avoid vague responses) and stored, so as to enrich the system's historical learning context.

Rule 8. Strictly bound the agent's iterations

To prevent resource exhaustion and to stop an agent from becoming trapped in fruitless correction loops in the event of repeated disagreement with the human, workflows must impose a strict iteration limit. Once crossed without acceptable validation, this limit must switch the system into a safe fallback mode: suspension of the task and alerting of a higher level of supervision.

Architecture frameworks and emerging standards

Moving from principle to production requires protocols capable of unifying the technical execution of tasks and compliance control. Five references structure the field today.

The 12 factors of agents (12-Factor Agents)

Inspired by the “Twelve-Factor App” methodology from SaaS development, the 12-Factor Agents guide, theorised by Dex Horthy (HumanLayer) in line with the lessons learned from orchestrating coding agents such as Claude Code and the CodeLayer IDE, defines the criteria for agentic applications that are reliable and operable in production [6].

Factor	Technical description	Role with respect to human supervision
F1. Natural Language to Tool Calls	Translating intention expressed in natural language into tool-call schemas.	Formalises, in an intelligible way, the action the agent is about to undertake.
F2. Own your prompts	Control, versioning and mastery of the instructions passed to the model.	Prevents semantic drift and guarantees behavioural stability.
F3. Own your context window	Engineering the size and structure of the data injected into the context.	Prevents overloading the model and limits decision-making hallucinations.
F4. Tools are just structured outputs	Modelling the agent's connectors as strict output schemas.	Allows deterministic validation before any external communication.
F5. Unify execution state and business state	Synchronising the machine's execution state and the state of the company's databases.	Avoids decisions based on obsolete business data.
F6. Launch/Pause/Resume with simple APIs	Starting, freezing or resuming a flow via standard APIs.	Indispensable for suspending the agent during asynchronous human interruptions.
F7. Contact humans with tool calls	Integrating the request for human help as a function executable by the agent.	Allows the agent to call on human expertise as a tool, in the event of uncertainty.
F8. Own your control flow	Retaining deterministic control over the state-transition graph.	Avoids entrusting the routing logic to the model's random interpretation.
F9. Compact Errors into Context Window	Simplifying and intelligently injecting system errors into the context.	Helps the agent correct its errors without systematic human intervention.
F10. Small, Focused Agents	Dividing the work into a constellation of specialised micro-agents.	Isolates tool access rights according to the principle of least privilege.
F11. Trigger from anywhere	Triggering and supervising the agent from any channel (Slack, SMS, email).	Brings approval requests closer to employees' workspaces.
F12. Make your agent a stateless reducer	Eliminating volatile internal states in favour of a pure, memoryless reducer.	Facilitates resumption after interruption and guarantees the reproducibility of audits.
F13. Pre-fetch all context (honourable mention)	Proactively loading the required data before the decision-making process.	Improves speed and avoids untimely data requests mid-flow.

Three factors directly carry human control: F7 (contacting the human as one calls a tool), F6 (suspending/resuming cleanly) and F10 (small agents with granular rights). These are the technical foundations of HITL.

AESP: economic sovereignty remains with the human

As soon as an agent carries out financial tasks (subscribing to a service, settling an invoice, managing third-party assets), a tension arises between the speed of the machine and human sovereign control over the assets. A research paper published in early 2026, the Agent Economic Sovereignty Protocol (AESP), formalises a strict invariant: the agent is economically capable, but never economically sovereign [7]. Any movement of funds must pass through a deterministic rules engine with eight verification levels before reaching the settlement layer.

Security rule	Control exercised	Consequence of a violation
Per-transaction limit	The absolute ceiling authorised for a single payment order.	Automatic rejection or escalation to human validation.
Time-window limit	Maximum cumulative spending over a sliding time window.	Immediate suspension of the agent's payment capabilities.
Address allowlist	Network addresses or target wallets certified in advance.	Blocking of any transaction outside the approved destinations.
Chain allowlist	Restriction to the authorised execution protocols.	Immediate rejection of the submission.
Method allowlist	Specifically permitted functions or smart-contract calls.	Blocking at the cryptographic layer.
First-payment review	Compliance check at the first interaction with a third party.	Requirement for reinforced human approval.
Minimum balance	Cash reserve to be preserved in the originating account.	Freezing of any commitment to new spending.
Budget limits	Overall envelope allocated to the agent's operating cycle.	Temporary deactivation of the financial tools.

These execution controls are reinforced, at the infrastructure level, by bilaterally signed cryptographic commitments (via the EIP-712 standard): no party can unilaterally modify the terms of a transaction. The protocol also preserves the confidentiality of flows thanks to ephemeral addresses derived by HKDF, which prevent an external observer from correlating the payments of a single agent. At this stage, AESP remains a recent academic proposal, not an adopted industry standard: its value lies in the clarity of the invariant it posits, more than in any maturity proven in production.

Capability-bound agent certificates

In multi-agent production, architectures face a risk of drift that 2026 research calls the capability-identity gap: an agent, having received human authorisation for a given scope, dynamically modifies its capabilities at execution time (by importing new tools or by delegating to subordinate agents) without the central system detecting this privilege escalation [8]. Conventional frameworks authenticate who the agent is, but not what it can do.

The proposed countermeasure: capability-bound identity certificates, which cryptographically bind the agent's identity to the exhaustive list of its declared tools, its underlying models and its limits (concretely, by extending an X.509 certificate with a fingerprint of the “capability manifest”). The mechanism rests on three security properties:

G1. Capability integrity: any modification of the tool configuration immediately invalidates the active rights and requires explicit human re-authorisation.
G2. Behavioural verifiability: all actions must be auditable after the fact to detect model substitutions or parameter tampering.
G3. Interaction auditability: exchanges between agents and with humans are recorded in a tamper-proof ledger, providing the forensic analysis trail in the event of an incident.

KnowNo: disturbing the human only when warranted

HITL has an enemy: approval fatigue. An agent that requests validation at every step wears out its supervisors and loses all value. The KnowNo method [9], presented in the work Robots That Ask For Help (Ren et al., CoRL 2023), addresses this problem by drawing on conformal prediction, which offers a statistical coverage guarantee (the correct option appears in the selected set with a probability of at least 1 − α, where α is the error rate tolerated by the organisation) while minimising the interruptions imposed on the operator.

The principle, without equations, comes down to one idea: instead of asking the agent “are you sure?” (a question it answers poorly, cf. the overconfidence of models), its uncertainty is calibrated on examples validated by experts, and escalation occurs only when several actions remain statistically plausible.

Concretely, faced with a state x, the agent considers a set of options 𝒴 = {y_1, …, y_m}. A calibration set S_cal = {(x_i, y_i)} of n historical situations is available, where each y_i is the action validated by an expert. For each example, a non-conformity score is computed, the inverse of the probability the model assigned to the correct answer:

s_i = 1 − f̂(y_i | x_i)

where f̂(y | x) is the probability assigned by the model to the option y. These scores are ordered and the quantile threshold corresponding to the targeted tolerance level is derived from them:

q̂ = Quantile( {s_1, …, s_n} ; ⌈(n+1)(1−α)⌉ ⁄ n )

At execution time, faced with a previously unseen situation x_test, the agent builds the set of acceptable candidate actions by retaining only those options whose non-conformity score remains below the threshold:

C(x_test) = { y ∈ 𝒴 : 1 − f̂(y | x_test) ≤ q̂ }

The decision whether or not to interrupt the human is then governed by the cardinality of this set:

Set `C(x_test)`	System decision
a single option	Autonomous execution of the sole action `y*`.
two options or more	Interruption: escalation to the operator in the form of a multiple-choice question.

If the set reduces to a single option, the model possesses a certainty mathematically aligned with the required level of safety, and it acts without disturbing the supervisor. If it contains several valid options, the agent identifies a zone of ambiguity, freezes its execution and presents the pre-selected choices. The guarantee of task completion then holds provided that the human answers correctly: KnowNo converts the statistical coverage guarantee into a guarantee of success, at the cost of a single question, asked only when it is necessary.

Fourteen principles for the human-agent lifecycle

Finally, the viability of an agentic system is not measured by the machine's performance alone: human-agent interaction is a design target in its own right. A 2026 academic synthesis [10] proposes fourteen principles spread across four stages of an agent's operational life, from commissioning to the repair of trust after an incident.

Stage 1: Initial scoping and expectations. Explicitly announcing the known capabilities, limits and failure modes (Set Accurate Expectations); calibrating anthropomorphism to the context (neutral in a professional environment); establishing a clear relational framework (triage assistant, co-pilot, adviser) before entrusting real responsibilities.
Stage 2: Interaction and shared control. Dynamically negotiating the level of initiative according to risk, with instant interruption and resumption (Negotiate Shared Control); making intention transparent; adjusting social cues without inducing dependence; regulating proactivity to protect the operator's concentration.
Stage 3: Long-term collaboration. Adapting to the user's state (fatigue, correction rate); consolidating in memory their objectives and values, not merely their past actions; actively preventing unhealthy dependence; favouring behavioural consistency (predictable errors) over erratic excellence.
Stage 4: Failure and repair of trust. Supporting a shared repair (direct-manipulation interfaces for correcting the agent); adapting the recovery strategy to the type of failure (state rollback and audit for serious errors); combining cognitive repair (explaining the error technically) and affective repair (expressing contextual regret and reaffirming the new safety instructions).

Synthesis: three directives for decision-makers

To deploy autonomous agents in industrial workflows without reproducing a Knight Capital-style scenario, IT managers and systems engineers can structure their architecture around three directives.

1. Isolate security control within the execution environment. Rule out security filters implemented solely as semantic instructions (system prompts). The validation of tool-call parameters, the verification of network rights and the management of transaction budgets must be handled by deterministic modules written in an imperative language, external to the model.

2. Establish dynamic control points and tamper-proof traceability. Structure complex flows into asynchronously managed state graphs. Synchronous interruption freezes execution without loss of information while awaiting approval. Every proposal, every operator response and every final validation are recorded in a decision log with cryptographic proofs, to simplify compliance auditing.

3. Use the human as a contextual corrector and learning guide. Direct the supervisor towards high-value-added cases. By statistically calibrating uncertainty (in the manner of KnowNo), the system automates standardised, high-confidence tasks and calls on the human, in the form of a choice, only for ambiguous situations. Every human correction is stored as an adaptive-learning knowledge base, gradually reducing the future need for manual interruptions.

These three directives trace one and the same conviction, which lies at the heart of the MATIA Method™: in the age of agents, scarcity does not shift towards generation (which becomes abundant) but towards verification. An agent has value in a company only if its autonomy is defensible before a client, an auditor or a regulator. Human-in-the-Loop is not the brake on automation: it is what makes automation industrialisable.

To turn these principles into a governed agentic architecture in your organisation, see the MATIA Method™ and the AI Express Audit & Roadmap.

Sources

[1] Knight Capital Group, Form 8-K (realised pre-tax loss of approximately $440 million), SEC EDGAR, 2 August 2012; and SEC, Order / press release 2013-222 (faulty code deployment reactivating obsolete code on an order router; ~45 min at market open), 16 October 2013. https://www.sec.gov/newsroom/press-releases/2013-222 [2] Mind the Confidence Gap: Overconfidence, Calibration, and Distractor Effects in LLMs (arXiv:2502.11028, 2025), verbalised overconfidence of models (displayed confidence ~88-99% vs markedly lower actual accuracy); corroborated by Can LLMs Express Their Uncertainty? (arXiv:2306.13063). The figure sometimes cited of “87% confidence” cannot be sourced to a precise benchmark and should be regarded as illustrative. https://arxiv.org/abs/2502.11028 [3] Anthropic, “Introducing the Model Context Protocol” (open protocol for connecting models to tools and data), 25 November 2024; dynamic capability discovery documented in the specification (tools/list, resources/list). https://www.anthropic.com/news/model-context-protocol · https://modelcontextprotocol.io [4] Google, “Announcing the Agent2Agent Protocol (A2A)” (agent-to-agent interoperability, complementary to MCP), 9 April 2025; protocol contributed to the Linux Foundation, June 2025. https://developers.googleblog.com/en/a2a-a-new-era-of-agent-interoperability/ [5] In-the-loop / on-the-loop / out-of-the-loop taxonomy, originating in the literature on autonomous systems: Human Rights Watch & Harvard IHRC, Losing Humanity: The Case Against Killer Robots, 19 November 2012 (adapted to software agents). https://www.hrw.org/report/2012/11/19/losing-humanity/case-against-killer-robots [6] Dex Horthy / HumanLayer, 12-Factor Agents, Patterns of reliable LLM applications (transposition of the “Twelve-Factor App” to agents; F11: “Trigger from anywhere, meet users where they are”; F13 as an honourable mention). https://github.com/humanlayer/12-factor-agents [7] Jian Sheng Wang, AESP: A Human-Sovereign Economic Protocol for AI Agents with Privacy-Preserving Settlement (arXiv:2603.00318, 27 February 2026), “economically capable but never economically sovereign” invariant, deterministic engine with eight checks, EIP-712, HKDF. Proposal preprint, not an adopted standard. https://arxiv.org/abs/2603.00318 [8] Ziling Zhou, Governing Dynamic Capabilities: Cryptographic Binding and Reproducibility Verification for AI Agent Tool Use (arXiv:2603.14332, March 2026), “capability-identity gap”, capability-bound certificates (X.509 extension), properties G1 (integrity), G2 (behavioural verifiability), G3 (auditability). Preprint. https://arxiv.org/abs/2603.14332 [9] Allen Z. Ren et al., Robots That Ask For Help: Uncertainty Alignment for Large Language Model Planners (“KnowNo”), CoRL 2023 (arXiv:2307.01928), conformal prediction, non-conformity score 1 − f̂(y|x), escalation to the human if the prediction set is not a singleton. https://arxiv.org/abs/2307.01928 [10] Haiyi Zhu, Canwen Wang, Qing Xiao, Hong Shen (Carnegie Mellon University), Design Principles for Human-Agent Interaction (arXiv:2606.20630, 2026), fourteen principles spread across four stages (Initially, During Interaction, Over Time, When Things Go Wrong). https://arxiv.org/abs/2606.20630

How to involve employee representative bodies (IRP) in your AI transformation: what recent case law says

Tue, 30 Jun 2026 00:00:00 GMT

The finding. Three judicial-court decisions in under two years, the latest in January 2026, confirm and refine one and the same rule: deploying AI without consulting the works council (CSE) exposes the project to court-ordered suspension, subject to a penalty payment. This is no longer a theoretical risk; it is a line of case law taking shape at a brisk pace. But the subject of this article is not fear of litigation. It is the method for doing the opposite: making it a social-dialogue milestone that strengthens an AI deployment rather than delaying it.

Too many companies that embark on AI still treat the CSE as a box to tick at the end of the project, when it is not a forgotten formality altogether. Judges no longer see it that way. Here is what the recent decisions say, what the law says, and how the MATIA Method™ turns this obligation into an advantage rather than a risk.

Three decisions in under two years: case law is taking shape fast

Decision	Facts	Outcome
TJ Nanterre, 14 Feb. 2025, no. 24/01457 [1]	Deployment of 5 software packages (including Finovox, Synthesia, Notify) incorporating AI, announced in January 2024; CSE consultation begun only in late September 2024	Suspension of the deployment subject to a penalty of €1,000/day for 90 days; €5,000 in damages; €2,000 (art. 700)
TJ Créteil, 15 Jul. 2025, no. 25/00851 [2]	Press group deploying generative-AI tools without prior consultation of the CSE	Suspension of the tools' use until the consultation is concluded
TJ Nanterre, 29 Jan. 2026, no. 25/02856 [3]	Replacement of a talent-management tool with two HR software packages featuring AI modules (annual reviews, assignment, identification of training needs); central CSE not consulted, the employer invoking a “simple technical evolution”	Characterisation rejected; suspension subject to a penalty of €500/day of delay

The common thread across the three decisions: the court holds that the introduction of AI constitutes a “new technology” within the meaning of Article L.2312-8 of the Code du travail, and that the pilot phase (once it involves real use by employees, even partial) is not a mere experiment exempt from consultation, but a first implementation that is, on the contrary, subject to it [1][4].

What the law says: threshold, time limits, and the right to expert assessment

Article L.2312-8 of the Code du travail requires the CSE to be informed and consulted prior to any decision affecting working conditions (including, now, the introduction of new AI technologies). Three points, often poorly understood, determine the real scope of this obligation:

50-employee threshold. The CSE exists from 11 employees onwards, but the full consultative competence of Article L.2312-8 (and the right to expert assessment that flows from it) presupposes a CSE with full economic powers, reserved for companies with at least 50 employees. Below that, the CSE has only powers over individual and collective grievances.
Right to expert assessment: Art. L.2315-94. For any “major” project within the meaning of L.2312-8 (including the introduction of new technologies), the CSE may appoint an accredited expert, funded 80% by the employer and 20% by the CSE.
Time limits for the opinion: Art. R.2312-6. One month by default; two months if the CSE appoints an expert (the most likely scenario on an AI project), and even three months where there is a central CSE and establishment-level committees. Counting on “a few weeks” is the leading cause of an unworkable timetable.

These three elements paint a simple message for business leaders: the obligation is neither universal (it presupposes 50 employees and internal AI, affecting the company's own employees, not AI sold to external customers) nor instantaneous (allow one to two months for the opinion). But where it applies, it is unavoidable, and ignoring it now comes at a steep cost in court.

The most common trap: “it is only an update”

The 29 January 2026 decision is worth dwelling on, because it targets the most widespread reflex among business leaders in a hurry: presenting an AI deployment as a simple technical improvement to an existing tool, in order to avoid the procedure. The court looked at the facts, not the label: the new HR software extended usage to all employees (rather than just two departments), incorporated decision-support algorithms into the evaluation and assignment of tasks, and exploited HR data in unprecedented ways to recommend training paths [3].

The criterion adopted is therefore not the internal project's title, but its real effect on working conditions and the content of tasks. A direct point of vigilance for any company upgrading an ERP, an HR tool or a line-of-business assistant by “discreetly” adding an AI layer: if the use changes in nature or scope, reclassification as a “new technology” looms. Ultimately, it carries the same penalty as a project announced as such.

How to do it properly: the method in four steps

The good news is that consultation, when well prepared, does not significantly lengthen a transformation project. It merely keeps it from being derailed in court. Four practical markers:

Start the timeline at scoping, not when the pilot begins. The MATIA Method™ places CSE consultation in Phase 2 (Scoping), when use cases are prioritised, so that the opinion is delivered before Phase 4 (Pilots). Given the real time limits (1 to 2 months, see above), this sequencing avoids the pitfall observed in the three decisions: a consultation started after the project has already begun to take on a life of its own.
Reuse the AI Act register as the information file. A company bringing itself into compliance with the AI Act already keeps a register of the AI systems in use, their risk level and the data processed. That is exactly the material the CSE must receive to deliver an informed opinion. A single documentary effort serves both obligations, rather than recreating an ad hoc file under pressure.
Share the AgentOps Plan → Execute → Verify grid. The triptych that already frames the technical steering of AI agents (see the MATIA Method™) also gives the CSE a shared reading grid: what the system will do, on what criteria its success is judged, who verifies. A shared language defuses much of the opacity that judges penalise.
Anticipate recourse to an expert. If the project is significant, the CSE will probably appoint an expert (Art. L.2315-94). This is not a sign of distrust, it is its most ordinary right on this kind of matter. Building it into the timeline from the outset (the limit extended to two months) avoids the surprise that derailed the three companies that were sanctioned.

Beyond consultation: when companies negotiate

Consultation is a legal floor; some companies go further and negotiate the use of AI with their trade unions. The movement remains a minority one but is accelerating: a study by the CEET (Centre d'études de l'emploi et du travail, Cnam) already counted, in October 2024, 242 company agreements mentioning AI signed between 2017 and 2024 by 160 organisations, with the share of agreements addressing AI multiplied by 2.5 between 2018 and 2023 [5].

The most recent and most comprehensive example is that of MAIF, whose company agreement on AI, signed on 7 May 2026, was ratified by all six representative trade unions (CAT, CFDT, CFE/CGC, CGT, FO, UNSA) [6][7]. It provides in particular for: no economic redundancy motivated by the deployment of AI alone, the priority reinvestment of efficiency gains into the service delivered and into enriching jobs, and above all the creation of an AI committee attached to the CSE (twelve members, three meetings a year) tasked with monitoring projects and their impacts. The CGT, a signatory, welcomed the fact that the CSE's prerogatives were not called into question, while regretting the absence of more ambitious commitments on sharing productivity gains [6]. A useful reminder: the agreement does not erase the balance of power, it gives it a framework.

Other companies have undertaken comparable steps, to varying degrees of ambition: AXA France signed a dedicated AI agreement on 13 June 2025; Prisma Media created, as early as February 2025, a committee to monitor AI uses within the CSE (for a limited term); BPCE integrated an unprecedented AI component into its three-year jobs and career-path management (GEPP) agreement of July 2025 [8]. The common motive, beyond the variable scope of each text, is rarely constraint: it is the predictability they offer an AI transformation project, by avoiding the timetable breakdowns to which the strictly litigious route is exposed.

The tools that already exist to help elected members

A recurring obstacle, documented by several firms specialising in CSE matters, is less legal than cognitive: elected members often discover AI tools “as they go”, sometimes after employees have begun using them, for lack of technical bearings to gauge their real impacts [9]. Two responses already exist:

DialIA (dial-ia.fr), launched in January 2025 on the initiative of the IRES and co-funded by the Anact, built with trade unions, employer organisations, companies and researchers. It is a cultural-adoption tool explicitly designed to give all actors in social dialogue (employers and employee representatives alike) an equivalent level of information on AI systems and their impacts on work [10].
CSE-accredited expert firms (Syndex, Secafi, and others) now offer analysis engagements dedicated to AI, which can be commissioned through the right to expert assessment (Art. L.2315-94): a concrete way of giving elected members technical expertise comparable to management's, rather than leaving them to judge on paper alone.

This is precisely the logic of Dimension 5 (AI Governance) of the MATIA Method™ Scale: at the Architecte tier, governance is no longer confined to an internal committee, it includes structured IRP social dialogue; at the Pionnier tier, this dialogue becomes permanent, the trajectory that the MAIF agreement already illustrates.

What the MATIA Method™ says

The three court decisions say nothing other than what the MATIA Method™ lays down as a principle: an AI deployment that ignores the company's intermediary bodies is not only a legal risk, it is a poorly governed deployment, just like a deployment without an AI Act register or without a Plan-Execute-Verify triptych. Consultation of the CSE is not an external obstacle to AI transformation; it is, structurally, one of its conditions for success, on a par with data quality or change management.

Going further

Croissance et Transitions supports the business leaders of SMEs and mid-cap companies in scoping their AI projects, social dialogue included: consultation timeline, information file aligned with the AI Act register, and a path towards a company agreement when maturity warrants it.

AI Express Audit & Roadmap, 60 minutes over video to place your SME on the MATIA Method™ Scale (Spectateur → Pionnier): croissance-transitions.fr.
In the same series: SaaS & AI: from technical debt to orchestration and The MATIA Method™ Scale: the 5 levels of AI maturity.

Paul-Antoine TUAL · AI Transformation Leader · Croissance et Transitions (SAS) · MATIA Method™

Sources

Tribunal judiciaire de Nanterre, 14 February 2025, no. 24/01457, suspension of the deployment of AI software, penalty €1,000/day (90 days), €5,000 in damages. https://www.doctrine.fr/d/TJ/Nanterre/2025/U0681F4C8B4DFC59B0391 ; commentary: Capstan Avocats, « Consultation du CSE et IA : un juge des référés ordonne la suspension du projet ! ». https://www.capstan.fr/articles/2643-consultation-du-cse-et-ia-un-juge-des-referes-ordonne-la-suspension-du-projet/
CMS Law, « L'IA suspendue : le juge exige la consultation du CSE avant tout déploiement » (covering in particular TJ Créteil, 15 July 2025). https://cms.law/fr/fra/legal-updates/l-ia-suspendue-le-juge-exige-la-consultation-du-cse-avant-tout-deploiement
L'Expertise Droit Social, « Intelligence artificielle : le CSE doit être consulté, même en phase pilote » (TJ Nanterre, 29 January 2026, no. 25/02856). https://www.lexpertise-droit-social.fr/veille/intelligence-artificielle-cse-consultation-nanterre-janvier-2026/
Voltaire Avocats, « Suspension du déploiement d'outils informatiques en phase pilote jusqu'à l'achèvement de la consultation du CSE ». https://www.voltaire-avocats.com/fr/suspension-du-deploiement-doutils-informatiques-en-phase-pilote-jusqua-lachevement-de-la-consultation-du-cse/
CEET (Centre d'études de l'emploi et du travail, Cnam), « L'IA dans les entreprises : que révèlent les accords négociés ? », October 2024, 242 agreements 2017-2024, 160 organisations. https://ceet.cnam.fr/publications/connaissance-de-l-emploi/l-ia-dans-les-entreprises-que-revelent-les-accords-negocies--1501114.kjsp
Argus de l'Assurance, « Maif : l'ensemble des organisations syndicales signent un accord sur les conditions de déploiement de l'intelligence artificielle dans l'entreprise », May 2026. https://www.argusdelassurance.com/mutuelles/maif/maif-lensemble-des-organisations-syndicales-signent-un-accord-sur-les-conditions-de-deploiement-de-lintelligence-artificielle-dans-lentreprise.IYJW2G6AZRE55LJODDOVAQ7IMI.html
MAIF, press release, « Maif adopte un accord d'entreprise IA », 7 May 2026. https://entreprise.maif.fr/actualites/presse/2026/maif-accord-entreprise-developpement-ia
Lexia Conseil, « Intelligence artificielle : ce que prévoient les premiers accords d'entreprise » (Prisma Media, AXA France, BPCE). https://www.lexia-conseil.fr/intelligence-artificielle-ce-que-prevoient-les-premiers-accords-dentreprise/
Officiel CE, « Les CSE face au déploiement de l'intelligence artificielle (IA) qui s'accélère… et des contentieux qui se multiplient ». https://www.officielce.com/dossier/fonctionnement-du-cse/jurisprudence/les-cse-face-au-deploiement-de-l-intelligence-artificielle-ia-qui-s-accelere-et-des-contentieux-qui-se-multiplient
Éditions Tissot, « Dial IA : un outil pour structurer le dialogue social autour de l'intelligence artificielle » ; DialIA, https://dial-ia.fr/. https://www.editions-tissot.fr/actualite/representants-du-personnel-ce/dial-ia-un-outil-pour-structurer-le-dialogue-social-autour-de-lintelligence-artificielle

Legislation cited: Code du travail (French Labour Code), Articles L.2312-8 (information and consultation of the CSE), L.2315-94 (right to expert assessment on major projects), L.2311-2 (11-employee threshold), R.2312-6 (consultation time limits).

The end of email and Office suites by 2030?

Fri, 26 Jun 2026 00:00:00 GMT

Our stance. Neither email nor the office suite “dies” in 2030. What comes apart is the dispersion: the stack of heterogeneous tools (email, telephone, WhatsApp, SMS, Slack or Teams, scattered Office files) that every worker must juggle today. This fragmentation is converging towards a single workspace, modular, AI-mediated, and increasingly contested on the terrain of sovereignty. The real question is not “which tool disappears?”, but “who governs the single surface that replaces the stack?”. This article first sets out the facts, historical, cognitive, economic and sociological, then the method.

The prophecy of an obsolescence of electronic mail and of monolithic office suites by 2030 returns with regularity in the management literature. Announced as imminent under the pressure of “disruptive technologies”, this twin disappearance is in reality more subtle than a substitution. The analysis of usage reveals not a death, but a deconstruction: integrated tools break down into specialised components, then recompose within unified and dynamic environments. This socio-technical shift redefines the management of attention, the power relations within dispersed teams, digital independence, and the very structure of work.

Email, from a tool of exchange to a heritage archive

Designed as the digital equivalent of the interdepartmental memo, electronic mail has become a multifunction command centre: calendar, file sharing, automations. Along the way, it has acquired a paradoxical status, at once backbone of organisational memory and receptacle of a contemporary written heritage.

Cultural institutions understood this before businesses did. The British Library holds the hybrid archive of the poet Wendy Cope, some 40,000 emails, acquired as early as 2011 and processed with the ePADD archiving tool [1]. The Harry Ransom Center at the University of Texas acquired in 2014 the archives of the novelist Ian McEwan, including his complete electronic correspondence since 1997 [2]. And the library of the University of Manchester holds the archive of the publisher Carcanet Press, some 215,000 emails and 65,000 attachments, explored in the Palladium project by means of visualisation and network-analysis tools (ePADD, Datawrapper, Gephi) [3]. Let us be clear, so as not to take the easy route: these collections are processed, at this stage, by visualisation tools, not by “artificial intelligence”. As for Salman Rushdie, often wrongly cited as being at Texas, his digital archive is in fact held at Emory University, which even emulated his old Macintosh to recreate his writing environment [4]. Email has become a historical source of the first order, provided one cites the right institution.

This heritagisation has a linguistic counterpart. A research tradition in internet linguistics (Crystal, Baron, Herring) describes email as a hybrid mode, halfway between the written and the spoken, shaped by a “principle of least effort” [5]. Lower case by default, omission of the subject or the auxiliary, consonantal abbreviations: so many general tendencies of computer-mediated communication (forms such as cn for can or wld for would are its most visible illustration, more on the SMS side than in professional email, as it happens). One must be wary of seeing in this a causal, measured “erosion” of professional language: it is a variation in register, not a proven degradation. But the direction is clear. The pursuit of efficiency has turned a format that was formal at first into a conversational channel.

The trial of email: the lesson of Atos's “Zero Email”

Contestation of the model intensifies as informational saturation worsens. The Atos experiment offers the most instructive precedent. On 7 February 2011, at the instigation of Thierry Breton, the group announced its intention to become a “zero email” company within three years, internal email being designated an “informational pollution” [6]. The target: to replace these exchanges with an enterprise social network (the blueKiwi platform, acquired in April 2012) organised into thematic communities. At its peak, the scheme brought together some 74,000 employees across nearly 7,500 communities [7].

The result is instructive on both counts. The volume of internal email per employee fell from around 100 to fewer than 40 per week, a drop of the order of 60% by the end of 2013 [8]. But total eradication failed: by 2013, the “zero” objective was tacitly abandoned. The underlying reason is not nostalgic attachment to email; it is that email fulfils irreducible functions (external communication, asynchronous exchanges, digital identity) that no closed internal network covers. To this one may add, with hindsight, an architectural argument: unlike the closed proprietary ecosystems that would come to dominate the following decade, email remains the only open, decentralised and universally interoperable protocol. That is precisely what makes it impossible to remove, and valuable to keep as a bedrock.

Characteristic	Electronic mail (SMTP/IMAP)	Team messaging (Slack, Teams)	Collaborative canvas (Miro, Mural)
Governance	Open standard, interoperable, decentralised	Proprietary ecosystem, centralised	Proprietary platform, API access
Temporal regime	Structured asynchronous, conducive to reflection	Near-synchronous, demanding immediacy	Hybrid, activity-centred
Register	Syntactic reduction, hybrid written/spoken style	Fragmented, emoticons, immediacy	Sticky notes, visual codes, syntheses
Preservation	Heritage integration (cultural archives)	Transactional archiving, silos	Visual persistence, version history

The moral of Atos is neither “email is eternal” nor “it must be killed”. It is subtler: one does not remove a channel, one puts it back in its place. The mistake of 2011 was to believe that a closed tool could absorb an open standard. The convergence now taking shape does the opposite: it keeps the open protocol as a bedrock, and removes from it only its catch-all function.

Information overload and the science of attention

The omnipresence of electronic communications has bred organisational pathologies: information overload, technostress. The attention economy at work constantly pits the protection of deep concentration (deep work) against devices designed to capture immediate responsiveness.

The immoderate use of carbon copy (CC) drowns employees under irrelevant flows. The cognitive cost of interruption, for its part, has long been documented: a study by Loughborough University (Jackson, Dawson & Wilson) measured that it takes on average 64 seconds, “a little over a minute”, to refocus after an email interruption [9]. Multiplied by the number of daily checks, this “interruption tax” runs into lost hours: an hour and a half a day is often cited, but that figure is an illustrative extrapolation, not a primary measurement. The solid anchor remains the 64 seconds per interruption. To counter the drift, some services even proposed pricing attention: Boxbe's “Attention Bond” mechanism imposed, in the mid-2000s, a cost of a few cents on unauthorised senders [10]. An archival curiosity (the feature has since disappeared), but a sound intuition: attention is a scarce resource that ought to be charged to whoever wastes it.

The remedies, for their part, belong to behavioural discipline. Mark Forster's Do It Tomorrow method proposes to handle messages in a consolidated batch the next day, rather than in a continuous flow [11]. At the scale of the organisation, the right to disconnect, enshrined in France in Article L2242-17 of the Labour Code (law of 8 August 2016, in force since 1 January 2017), aims to frame the culture of permanent availability, through annual negotiation or a charter, rather than to penalise it directly [12]. The movement is not only French: Ontario legislated it in 2021, and the Canadian federal legislator in 2024. Quebec, by contrast, has declined to do so, contrary to a widespread belief [13].

On the quantitative-benchmark side, the Observatoire de l'infobésité et de la collaboration numérique (OICN), launched in May 2023 by Mailoop and the firm Mazars and not by any one isolated expert, provides two valuable operational thresholds: keeping the share of emails sent in copy below 20% (against around 30% today), and switching to a synchronous channel beyond three emails addressed to the same person [14]. Two rules, simple, measurable, and immediately actionable.

The deconstruction of the office suite: sovereignty and modularity

The office-suite market, long locked up by the Microsoft 365 / Google Workspace duopoly (together nearly 96% of the global market [15]), is undergoing a deconstruction under the effect of two forces: price inflation and the demand for sovereignty.

On prices, the trajectory is documented. Microsoft raised its commercial rates by +8.6% to +25% depending on the offer in 2022 [16]. In January 2025, it was the consumer Personal and Family offers that jumped, up to +43% in euros, with Copilot forcibly bundled into the basket [17]; Google made the same move by integrating Gemini into Workspace, with an increase of around +17% including for those who do not use the AI [18]. And a new commercial increase takes effect on 1 July 2026 (announced on 4 December 2025): from +5% to +43% depending on the plan, up to +15% to +23% of actual effect for large accounts once discounts are pared back [19]. Embedded AI has become the pretext for increases that the customer does not choose.

To this inflation is added legal dependency. The American CLOUD Act compels providers under United States jurisdiction to hand over the data they control, whatever its place of storage, including within the Union [20]. That the decoupling of a simple video tool (Teams) from the suite required two years of European antitrust proceedings, settled by binding commitments from Microsoft in September 2025, speaks volumes about the reality of the lock-in [21].

Hence the rise of modular ecosystems, founded on the interoperability of specialised components and on open source. In France, the “Suites bureautiques collaboratives cloud” call for projects under France 2030 (April 2022), operated by Bpifrance with ANSSI and the DGE, backed three winners (Wimi, Jamespot, Interstis) and eighteen partners, for €23 million in aid [22]. Two sovereign suites emerged from it: CollabNext (led by Jamespot, built on OnlyOffice and Jitsi, at around €9/user/month) [23] and Hexagone (led by Interstis, assembling BlueMint, Linphone, XWiki…), marketed to local authorities since October 2023 [24]. To these are added La Suite numérique from the State (DINUM, including the Tchap messaging service on the Matrix protocol) and a fabric of vendors, Whaller, Talkspirit, Twake, Jalios, that position themselves explicitly as an alternative to Microsoft 365 [25][26]. These solutions are hosted on SecNumCloud-qualified infrastructure (for example 3DS Outscale). The qualification bears on the hosting, not on the software, and one of its criteria is precisely imperviousness to the extraterritoriality of non-European laws [27][28]. The modular model makes it possible to migrate gradually, without the brutal and risky transition of a big bang.

The subject is not national in the sense of designating a culprit. Croissance et Transitions does not reason in terms of the origin of a risk, but of applicable jurisdiction and infrastructure dependency. In March 2026, the Conseil d'État did, moreover, judge “sufficiently framed” the hosting of health data on a non-European infrastructure, but under strict conditions, and the body concerned had itself begun an exit [29]. The point is not anathema; it is the trade-off made in good conscience.

The end of the telephone, of WhatsApp and of scattered channels

It is here that the subject broadens. For the office suite is only one piece of a stack that every employee must juggle: email and telephone and WhatsApp and SMS and Slack or Teams and shared files. This dispersion is the true disease, and it links the question of attention directly to that of the tool.

The historical channels are emptying one by one, not by disappearance but by migration. Voice consumption recedes continuously (−3% in France in 2024, falling for ten years [30]) and generational preference tips towards the written word: the share of those who favour the telephone to reach their loved ones drops from 62% among baby boomers to 17% among Generation Z [31]. The voice has not disappeared: it has changed pipe. WhatsApp has passed 3 billion monthly active users [32], global business messaging traffic is estimated at 2,000 billion messages in 2025 (projected to 3,000 in 2030) [33], while the classic SMS collapses, −27% over the year 2025, in favour of RCS and enriched messaging [34]. Gartner anticipates that, as early as 2027, self-service and chat will overtake the telephone and email as the dominant customer-service channels [35].

The problem, then, is not one channel too many: it is that none disappears fast enough to lighten the stack. Each added tool is a new box to monitor, a new context to reload: exactly the 64-second tax described above, multiplied by the number of surfaces. The market's answer is not “yet another application”, but the reverse: convergence towards a single surface where these channels come together. That is the meaning of the “single tool”: not a resurrected proprietary monolith, but a unified and governed workspace, resting on open standards, where email, chat, video, documents and agents coexist under a single data policy.

The test of the figures: across 874 business conversations, two thirds of “useful” email are substitutable

It remains to verify the thesis on the evidence. We have done so on our own data. Most studies of email overload count all messages, and inflate the volumes by mixing human mail with automatic notifications. We took the opposite route: removing the whole of machine-email from the outset so as to judge only the mail involving a human at the other end, then to test whether, even then, email remains the right tool.

The field: the professional inbox of a French AI-consulting micro-business, April-June 2026. From the raw flow (1,400 threads), we set aside 54 threads of technical noise and 472 threads of automated supervision (monitoring, application statuses, reports), that is 34% of the raw flow, which concentrated the bulk of the “urgency” signals. There remains a strictly business corpus of 874 conversations (631 with a usable summary), sorted by rules into seven categories. It is a monographic study (N = 1), with non-generalisable proportions and estimated, non-experimental substitutability rates: we accept its limits [47].

The first result is clear. Estimating, category by category, the share replaceable by a non-email channel (shared project space, in-app business object, booking link, catalogue integration), around 68% of the volume (≈ 595 threads) falls to a substitute.

Category	Threads	%	Substitutable (est.)	Main substitute
C3 · Client relationship & project coordination	285	32.6%	~60%	Shared client space + booking link
C2 · Training, pedagogy & Qualiopi	147	16.8%	~70%	Shared training repository + e-signature
C6 · Invoicing, payments & accounting	141	16.1%	~75%	Invoicing register + portal + e-invoicing
C8 · Third-party notifications, monitoring & logistics	107	12.2%	~85%	Automatic filing / unsubscription
C7 · Prospecting & inbound solicitations	69	7.9%	~45%	CRM intake form + booking link
C4 · Suppliers, catalogue & procurement	66	7.6%	~70%	Catalogue integration + supplier space
C5 · Quotes, contracts & signatures	59	6.8%	~80%	In-app quotes + electronic signature

<figure style="margin:2rem 0;"> <svg viewBox="0 0 720 432" width="100%" height="auto" role="img" aria-labelledby="f1t f1d" style="max-width:100%;height:auto;font-family:system-ui,-apple-system,'Segoe UI',Roboto,sans-serif;"> <title id="f1t">Breakdown of business email and estimated substitutable share</title> <desc id="f1d">Across 874 business conversations sorted into seven categories, around 68% of the volume falls to a non-email substitute.</desc> <text x="6" y="26" font-size="14" font-weight="700" fill="#0B2545">Estimated substitutable share of business email, by category (n = 874 threads)</text> <line x1="337" y1="56" x2="337" y2="374" stroke="#E2E8F0" stroke-width="1"/> <line x1="463" y1="56" x2="463" y2="374" stroke="#E2E8F0" stroke-width="1"/> <line x1="590" y1="56" x2="590" y2="374" stroke="#E2E8F0" stroke-width="1"/> <text x="337" y="390" font-size="11" fill="#94A3B8" text-anchor="middle">100</text> <text x="463" y="390" font-size="11" fill="#94A3B8" text-anchor="middle">200</text> <text x="590" y="390" font-size="11" fill="#94A3B8" text-anchor="middle">300 threads</text> <text x="6" y="85" font-size="12.5" font-weight="600" fill="#0F172A">C3 · Client relationship & project</text> <rect x="210" y="70" width="217" height="22" rx="2" fill="#FF7A59"/> <rect x="427" y="70" width="144" height="22" rx="2" fill="#0B2545"/> <text x="579" y="85" font-size="11.5" fill="#475569">32.6% · ~60% subst.</text> <text x="6" y="132" font-size="12.5" font-weight="600" fill="#0F172A">C2 · Training & Qualiopi</text> <rect x="210" y="117" width="130" height="22" rx="2" fill="#FF7A59"/> <rect x="340" y="117" width="56" height="22" rx="2" fill="#0B2545"/> <text x="404" y="132" font-size="11.5" fill="#475569">16.8% · ~70%</text> <text x="6" y="179" font-size="12.5" font-weight="600" fill="#0F172A">C6 · Invoicing & accounting</text> <rect x="210" y="164" width="134" height="22" rx="2" fill="#FF7A59"/> <rect x="344" y="164" width="45" height="22" rx="2" fill="#0B2545"/> <text x="397" y="179" font-size="11.5" fill="#475569">16.1% · ~75%</text> <text x="6" y="226" font-size="12.5" font-weight="600" fill="#0F172A">C8 · Notifications & monitoring</text> <rect x="210" y="211" width="115" height="22" rx="2" fill="#FF7A59"/> <rect x="325" y="211" width="21" height="22" rx="2" fill="#0B2545"/> <text x="354" y="226" font-size="11.5" fill="#475569">12.2% · ~85%</text> <text x="6" y="273" font-size="12.5" font-weight="600" fill="#0F172A">C7 · Inbound prospecting</text> <rect x="210" y="258" width="39" height="22" rx="2" fill="#FF7A59"/> <rect x="249" y="258" width="48" height="22" rx="2" fill="#0B2545"/> <text x="305" y="273" font-size="11.5" fill="#475569">7.9% · ~45%</text> <text x="6" y="320" font-size="12.5" font-weight="600" fill="#0F172A">C4 · Suppliers & catalogue</text> <rect x="210" y="305" width="59" height="22" rx="2" fill="#FF7A59"/> <rect x="269" y="305" width="25" height="22" rx="2" fill="#0B2545"/> <text x="302" y="320" font-size="11.5" fill="#475569">7.6% · ~70%</text> <text x="6" y="367" font-size="12.5" font-weight="600" fill="#0F172A">C5 · Quotes, contracts, signatures</text> <rect x="210" y="352" width="60" height="22" rx="2" fill="#FF7A59"/> <rect x="270" y="352" width="15" height="22" rx="2" fill="#0B2545"/> <text x="293" y="367" font-size="11.5" fill="#475569">6.8% · ~80%</text> <rect x="210" y="406" width="13" height="13" rx="2" fill="#FF7A59"/> <text x="229" y="416" font-size="11.5" fill="#475569">Substitutable (non-email channel)</text> <rect x="455" y="406" width="13" height="13" rx="2" fill="#0B2545"/> <text x="474" y="416" font-size="11.5" fill="#475569">Conversational core</text> </svg> <figcaption style="font-size:0.8rem;color:#475569;margin-top:0.5rem;line-height:1.5;">Figure 1. Bar length proportional to the number of threads; the <span style="color:#E55F3B;font-weight:600;">coral</span> share is the volume judged substitutable by a non-email channel, the <span style="color:#0B2545;font-weight:600;">navy</span> share the irreducible conversational core. Source: in-house study [47].</figcaption> </figure>

The second result is more surprising. Once the bots are removed, only 14% of threads are “urgent” (against 26% when supervision was included) and 14% of high importance. The lesson is limpid: the urgency of an inbox is largely manufactured by automated systems. Rid of them, the human mail of a micro-business is rarely pressing: 46% calls for an action, but at short or medium effort.

The third rests on an optical illusion. At least 21% of threads carry a document, but the density is very uneven: 97% for quotes and contracts, 65% for suppliers, 35% for accounting, 25% for training, the rest under 10%. Four categories out of seven concentrate almost the entirety of the document flow. Above all, the native “contains an attachment” indicator misleads: in a representative thread, an appointment-setting spread over nine emails and twelve days, each reply was marked “attachment” when it was only the signature logos copied over at each turn. A fraction of the “weight” of attachments is not business information; it is duplicated signature chrome.

<figure style="margin:2rem 0;"> <svg viewBox="0 0 720 320" width="100%" height="auto" role="img" aria-labelledby="f2t f2d" style="max-width:100%;height:auto;font-family:system-ui,-apple-system,'Segoe UI',Roboto,sans-serif;"> <title id="f2t">Document density by category</title> <desc id="f2d">Almost all of the document flow is concentrated in four categories; quotes and contracts carry a document in 97% of cases.</desc> <text x="6" y="26" font-size="14" font-weight="700" fill="#0B2545">Share of threads carrying a document, by category</text> <line x1="230" y1="56" x2="230" y2="288" stroke="#E2E8F0" stroke-width="1"/> <line x1="435" y1="56" x2="435" y2="288" stroke="#E2E8F0" stroke-width="1"/> <line x1="640" y1="56" x2="640" y2="288" stroke="#E2E8F0" stroke-width="1"/> <line x1="316" y1="58" x2="316" y2="288" stroke="#475569" stroke-width="1" stroke-dasharray="4 3"/> <text x="316" y="50" font-size="10.5" fill="#475569" text-anchor="middle">≥ 21% (corpus average)</text> <text x="6" y="87" font-size="12.5" font-weight="600" fill="#0F172A">C5 · Quotes & contracts</text> <rect x="230" y="70" width="398" height="26" rx="2" fill="#E55F3B"/> <text x="636" y="88" font-size="12" font-weight="700" fill="#0B2545">97%</text> <text x="6" y="135" font-size="12.5" font-weight="600" fill="#0F172A">C4 · Suppliers & catalogue</text> <rect x="230" y="118" width="267" height="26" rx="2" fill="#FF7A59"/> <text x="505" y="136" font-size="12" font-weight="700" fill="#0B2545">65%</text> <text x="6" y="183" font-size="12.5" font-weight="600" fill="#0F172A">C6 · Accounting</text> <rect x="230" y="166" width="144" height="26" rx="2" fill="#FF7A59"/> <text x="382" y="184" font-size="12" font-weight="700" fill="#0B2545">35%</text> <text x="6" y="231" font-size="12.5" font-weight="600" fill="#0F172A">C2 · Training</text> <rect x="230" y="214" width="103" height="26" rx="2" fill="#FF7A59"/> <text x="341" y="232" font-size="12" font-weight="700" fill="#0B2545">25%</text> <text x="6" y="279" font-size="12.5" font-weight="600" fill="#0F172A">Other (C3 · C7 · C8)</text> <rect x="230" y="262" width="33" height="26" rx="2" fill="#CBD5E1"/> <text x="271" y="280" font-size="12" font-weight="700" fill="#475569">< 10%</text> <text x="230" y="306" font-size="11" fill="#94A3B8" text-anchor="middle">0</text> <text x="435" y="306" font-size="11" fill="#94A3B8" text-anchor="middle">50%</text> <text x="640" y="306" font-size="11" fill="#94A3B8" text-anchor="middle">100%</text> </svg> <figcaption style="font-size:0.8rem;color:#475569;margin-top:0.5rem;line-height:1.5;">Figure 2. Share of threads carrying at least one document, by category (measured by lexical markers in the body, hence a floor). Four categories out of seven concentrate the bulk of the document flow. Source: in-house study [47].</figcaption> </figure>

What to conclude? That the lever is not editorial but architectural: it is a matter of getting information out of email. The “nine-emails-for-one-appointment” is resolved by a booking link, already present, ironically, in the correspondent's signature. Quotes and invoices become transactional objects signed and paid online rather than PDFs shuttled about. Project documents live in a versioned shared space where the file is a reference, not a frozen copy. The catalogue synchronises instead of circulating in a spreadsheet. Once these substitutes are in place, the substantive human exchange (negotiation, advice, relationship) weighs no more than ~32% of business mail (and an even smaller fraction of the real inbox). That is exactly what email was designed for.

One recognises, feature for feature, the modules of a unified workspace: the shared project space, the in-app business object, the booking link, the synchronised catalogue. The substitutability measured here does not herald the disappearance of email; it draws the empirical map of convergence: the proof, on real data, that two thirds of an inbox are already seeking to become something other than an email.

From static writing to dynamic collaborative canvases

At the heart of this single surface, the static document (.docx, .pdf) yields ground to living, shared data: no longer rigid files, but fluid blocks and databases editable in real time. This shift is embodied in visual-collaboration canvases, long studied by research in computer-supported cooperative work (CSCW).

The work of Vita Hinze-Hoare (University of Southampton, 2006-2007) models these environments as functional “spaces”: the collaborative workspace (CSCW) articulates communication, planning, sharing and production; learning contexts (CSCL) add to it reflection, the social dimension, assessment; collaborative research (CSCR) further adds knowledge, confidentiality, negotiation and publication [36]. One should retain the intuition rather than the formula: the richer a collaboration, the more it demands distinct spaces, including, crucially, a private space to mature an idea before exposing it.

Canvases such as Miro (or Mural, FigJam) materialise these spaces, and now integrate AI into them. IBM's research presented at ACM's CHIWORK 2024 conference (He et al.) shows how a large language model inserts itself into a canvas for brainstorming and the generation of idea sticky notes, while surfacing two imperative user needs: transparency over the ownership of AI-generated content, and the guarantee of private spaces to think before publishing on the common canvas [37]. Another, separate work, on the role of LLMs in collaborative design, adds a third need: to be able to filter the contributions of autonomous agents so as to avoid early anchoring and contextual drift [38]. These three requirements, ownership, privacy and filtering, sketch the architecture of a tenable augmented canvas.

Layer	Role in the canvas	Human-AI interaction	Managerial stake
Scoping	Creative objectives, initial hypotheses	Prompts, workshop configuration	Preserve conceptual autonomy, avoid anchoring bias
Interaction	Exchange flow, feedback loops	Synchronous co-editing, dynamic boards	Master the attentional load
AI system	Idea generation, classification	Proactive agents, generated sticky notes	Source traceability, hallucination detection
Ethics	Security, compliance, transparency	Encryption, bias filtering	Privacy, shared responsibility for decisions

Cognitive automation and the restructuring of work by 2030

Generative AI at the heart of this single surface redefines the nature of office tasks, and introduces a cognitive trade-off. In outsourcing reasoning and writing to agents, the knowledge worker risks becoming, in the phrase of the researcher Advait Sarkar (Microsoft Research / Cambridge), a “middle manager of his own thoughts”, a shift that can erode critical thinking and memorisation [39]. The remedy is not refusal of the tool, but discipline: never to treat an AI output as a final answer.

The recomposition of employment follows a taxonomy that the BCG Henderson Institute specified in April 2026 (“AI Will Reshape More Jobs Than It Replaces”): across all occupations, 5% are “amplified”, 14% “rebalanced”, 12% “substituted”, the rest distributed among enabled, divergent or little-exposed roles [40]. Three dynamics emerge from it:

Substituted roles: structured, repetitive, codifiable tasks. Harvard's research (Chen, Srinivasan & Zakerinia, working paper 25-039) measures, for the occupations most exposed to automation, a drop of the order of 13% in job postings per quarter and per firm since the arrival of generative AI, affecting functions such as medical transcription, correspondence secretarial work or telemarketing [41]. BCG evokes, in time, a possible elimination of 10% to 15% of these positions [40].
Augmented roles: expertise, judgement, creativity, empathy. For the occupations most complementary to AI (financial analysis, clinical neuropsychology, cartography…), the same work notes a rise of the order of 20% in postings combining these skills with the use of AI [41].
Rebalanced roles, whose routine components are automated to free up time of higher relational or strategic value.

This polarisation strikes young graduates first. The World Economic Forum, with PwC, estimates that more than one in three young workers holds a position with medium-to-high exposure to the changes induced by AI [42]. And the velocity of skills races ahead: according to the PwC barometer of June 2025, the skills required were evolving 66% faster in the occupations most exposed to AI, and exposed entry-level positions are far more likely to demand a “senior” level from the outset [43]. Without continuous training, the risk is the rapid deskilling of those entering the market.

On the sociological plane, finally, these tools fracture collectives. The research of Schulz, Wiborg & Robinson (American Behavioral Scientist, 2023) distinguishes two profiles: intensive users of instant messaging, the Slackers, prefer remote work, while those subjected to long videoconferences, the Zoomers, aspire to return to the office [44]. This “texture of practices”, to take up Silvia Gherardi's concept [45], becomes heterogeneous. Collaborative platforms moreover offer managers an asymmetric visibility over transactional activity, while thinning out the informal interactions of co-presence, a displacement of visibility at work analysed by Benedetto-Meyer & Boboc [46]. The risk, by 2030, is isolation and the weakening of team cohesion.

Profile (BCG / Harvard taxonomy)	Effect of AI by 2030	Trend in postings	Most exposed	Training strategy
Substituted (transcription, data entry, tier-1 support)	Process automation	Drop ~13% (observed)	Administrative office functions	Proactive redeployment towards the non-automatable
Augmented / amplified (analysis, design, consulting)	Creative capacities multiplied tenfold	Rise ~20%	Skilled, tech-savvy profiles	Human-AI collaboration, critical thinking
Rebalanced (managers, project leaders)	Automated reporting, managerial refocusing	Stable, internal reconfiguration	Managers in a matrix environment	Emotional intelligence, direct communication
Entry-level (young graduates)	High exposure to task change	Rapid mutation of requirements	Generation Z	Mentoring, reduction of multitasking

What the MATIA Method™ says: from endured dispersion to the governed workspace

Let us put the threads together. Email does not die; it loses its catch-all function. The Office suite does not die; it breaks down into modules. The telephone, WhatsApp, SMS do not die; they migrate. What dies is the scattered stack, and with it the attentional tax, the captive inflation and the legal dependency it drags along. Convergence towards a single tool is the logical answer. But everything depends on which single tool.

The trap would be to reconstitute the monolith: an integrated, proprietary, closed suite, outside jurisdiction, that is, replacing an endured dispersion with an endured lock-in. The MATIA Method™ posits the opposite: the single surface has value only if it is governed, that is, founded on three non-negotiable safeguards: open standards (email remains the bedrock of interoperability), modularity (each component replaceable one by one), reversibility (a real right of exit, with data portability). Without them, one merely changes owner.

It is this trajectory that the five-level AI-maturity scale describes: Spectateur, Artisan, Orchestre, Architecte, Pionnier. To move from the endured stack to the governed surface is not a technological leap; it is the accession to the Orchestre level, the one where one ceases to pile up isolated tools in order to coordinate a whole under a single policy. Concretely, this is what the Junyr™ suite embodies: a sovereign messaging service (Junyr Mail™, eIDAS, hosted in France) as an open and auditable bedrock; operable AI agents (Junyr Agents™) that execute within business processes under written mandate and audit log; a sovereign meeting assistant (Junyr Visio-IA™). All on a single surface, where each agent action remains traceable, supervised and revocable. The agents there apply a triptych at each step: Plan → Execute → Verify. The agent announces what it is about to do and its success criteria, executes, then a distinct step checks before continuing. Verification becomes native, and it is this that makes convergence tenable rather than opaque.

The counter-argument, taken seriously

Three objections deserve a frank response.

“Email will never die, you are exaggerating.” That is fair, and it is the whole point of the interrogative title. Email is the open standard par excellence; it persists, and that is good news: it becomes the interoperability bedrock of the single surface, just as the voice survived by migrating towards video and applications. What is extinguished is not the protocol, but the inbox as the place where everything gets done.

“Sovereign suites are less mature and less convenient.” The comfort of integrated suites is real, and the Conseil d'État itself has judged certain non-European uses acceptable under conditions [29]. But the equation tips: the inflation tied to embedded AI [17][18][19], the European antitrust proceedings [21] and the rising maturity of qualified alternatives [23][24][27] narrow the gap. Sovereignty is not an activist absolute; it is a cost/risk trade-off to be made in good conscience.

“A single tool: is that not a new lock-in?” This is the decisive objection, and it is healthy. The difference between a governed convergence and a new monopoly hangs entirely on the three safeguards: open standards, modularity, reversibility. A closed single surface would be worse than the dispersion it replaces. An open and reversible single surface is, on the contrary, the only way to recover attention, budget and sovereignty all at once. The devil is in the governance, not in the unification.

Three questions to ask this very week. 1. How many scattered communication tools must one of our employees monitor each day, and which are endured rather than chosen? 2. Have we set two measurable hygiene rules (fewer than 20% of emails in copy; synchronous switch beyond three back-and-forths)? 3. If we were to converge towards a single surface, would it be open, modular and reversible, or would we be swapping one lock-in for another?

Going further

Croissance et Transitions supports the leaders of mid-caps and SMEs in transforming an endured tool stack into a governed workspace: mapping of channels, measurable attentional hygiene, and a sovereign-convergence trajectory when the data demand it.

AI Express Audit & Roadmap: 60 minutes by video to situate your SME on the maturity scale (Spectateur → Pionnier), at croissance-transitions.fr.
In the same series: The end of the telephone: why the chat + video pairing is becoming the standard and The “All-Cloud” is dead: 5 risks that make strategic on-premise a must.

Paul-Antoine TUAL · AI Transformation Leader · Croissance et Transitions (SAS) · MATIA Method™ · Junyr Mail™ · Junyr Agents™

Sources: verifiable, June 2026

[1] Sophie Baldock, Process and progress: working with born-digital material in the Wendy Cope Archive at the British Library, Archives and Manuscripts (Taylor & Francis), 2017. https://www.tandfonline.com/doi/full/10.1080/01576895.2017.1408024

[2] Harry Ransom Center (UT Austin), Acclaimed Writer Ian McEwan's Archive Acquired by Harry Ransom Center (May 2014). https://www.hrc.utexas.edu/press/releases/2014/ian-mcewan-archive.html

[3] University of Manchester Library, Palladium: providing access to large literary archives in a digital medium (Carcanet Press project, 2019-2021). https://www.manchester.ac.uk/about/news/palladium-providing-access-to-large-literary-archives-in-a-digital-medium/

[4] Emory University, Rushdie: Digital archive at Emory "allowed me to write" memoir (March 2012). https://news.emory.edu/stories/2012/03/er_rushdie_digital_archives_memoris/campus.html

[5] David Crystal, Language and the Internet, Cambridge University Press, 2001 (and Naomi S. Baron, Always On, 2008; Susan C. Herring, Computer-Mediated Discourse). https://www.cambridge.org/core/books/language-and-the-internet/8F85B9421BA8C0435F67142D0DD14AD2

[6] Zonebourse / AOF, Atos Origin affiche son ambition de devenir une entreprise « zéro e-mail » d'ici trois ans (7 February 2011). https://www.zonebourse.com/cours/action/ATOS-SE-4612/actualite/ATOS-ORIGIN-affiche-son-ambition-de-devenir-une-entreprise-zero-e-mail-d-ici-trois-ans-13554015/

[7] N. Silic & M. Back, Atos, Towards Zero Email Company, ECIS 2015 (Univ. of St. Gallen). https://aisel.aisnet.org/ecis2015_cr/168/

[8] Zero Email initiative: a critical review of Change Management, Journal of Information Technology Teaching Cases (Springer), 2018. https://link.springer.com/article/10.1057/s41266-018-0033-y

[9] T. Jackson, R. Dawson, D. Wilson, The Cost of Email Interruption, Journal of Systems and Information Technology, 2001 (Loughborough University). https://interruptions.net/literature/Jackson-JOSIT-01.pdf

[10] Boxbe, Attention Bond Mechanism; see Getting Paid to Receive Spam, MIT Technology Review (2006). https://en.wikipedia.org/wiki/Boxbe

[11] Mark Forster, Do It Tomorrow and Other Secrets of Time Management, Hodder & Stoughton, 2006. http://markforster.squarespace.com/do-it-tomorrow/

[12] Légifrance, Article L2242-17 du Code du travail (law no. 2016-1088 of 8 August 2016, in force on 1 January 2017). https://code.travail.gouv.fr/code-du-travail/l2242-17

[13] Avantages, Québec ne réglementera pas le droit à la déconnexion (2025); cf. Ontario, Working for Workers Act (2021) and federal Canada, Bill C-69 (2024). https://www.avantages.ca/actualites/nouvelles/quebec-ne-reglementera-pas-le-droit-a-la-deconnexion/

[14] OICN (Mailoop & Forvis Mazars), Lancement de l'Observatoire de l'infobésité et de la collaboration numérique (15 May 2023); framework taken up by Talkspirit, Étude OICN : l'impact de nos pratiques collaboratives (2023). https://www.talkspirit.com/blog/etude-oicn-infobesite-quel-est-limpact-de-nos-pratiques-collaboratives

[15] Statista, Office productivity software, worldwide market share (2026). https://www.statista.com/statistics/983299/worldwide-market-share-of-office-productivity-software/

[16] Microsoft 365 Blog, New pricing for Microsoft 365 (19 August 2021, effective 1 March 2022; +8.6% to +25%, commercial offers). https://www.microsoft.com/en-us/microsoft-365/blog/2021/08/19/new-pricing-for-microsoft-365/

[17] Pureinfotech, Microsoft 365 Personal and Family now cost more, the 2025 price increase explained (January 2025, Copilot bundled in). https://pureinfotech.com/microsoft-365-personal-family-price-increase/

[18] Incentro, Google Workspace price increase 2025 (+17%, Gemini) (January 2025). https://www.incentro.com/en-EAF/news/google-workspace-price-increase-2025

[19] Microsoft 365 Blog, Advancing Microsoft 365: New capabilities and pricing update (4 December 2025, effective 1 July 2026). https://www.microsoft.com/en-us/microsoft-365/blog/2025/12/04/advancing-microsoft-365-new-capabilities-and-pricing-update/

[20] CLOUD Act (US, 23 March 2018, 18 U.S.C. §2713); see LeMagIT, CLOUD Act : entre le marteau et l'enclume (2025). https://www.lemagit.fr/tribune/CLOUD-Act-entre-le-marteau-et-lenclume

[21] European Commission, Commitments accepted, Microsoft Teams (IP/25/2048, 12 September 2025). https://ec.europa.eu/commission/presscorner/detail/en/ip_25_2048

[22] Bpifrance / DGE / ANSSI, France 2030 : vers un renforcement de l'offre cloud de confiance (call for projects "Suites bureautiques collaboratives cloud", April 2022-2023). https://presse.bpifrance.fr/france-2030-vers-un-renforcement-de-loffre-cloud-de-confiance

[23] LeMagIT, Sortie de CollabNext, suite collaborative souveraine issue de France 2030 (May 2024). https://www.lemagit.fr/actualites/366583175/Sortie-de-CollabNext-suite-collaborative-souveraine-issue-de-France-2030

[24] La Gazette des Communes, La France se dote d'une suite collaborative souveraine pour les collectivités (Hexagone) (2023). https://www.lagazettedescommunes.com/890712/

[25] Next.ink, Suites bureautiques européennes : la bataille de l'alternative à Microsoft et Google (2025). https://next.ink/232157/suites-bureautiques-europeennes-la-bataille-de-lalternative-a-microsoft-et-google/

[26] Blog Whaller, Cloud souverain : 8 acteurs français en alternative à Microsoft 365 (2024-2025). https://blog.whaller.com/presse/cloud-souverain-8-acteurs-francais-serigent-en-alternative-a-microsoft-365/

[27] Le Monde Informatique, Tixeo devient une suite collaborative souveraine (2025). https://www.lemondeinformatique.fr/actualites/lire-tixeo-devient-une-suite-collaborative-souveraine-99777.html

[28] ANSSI, SecNumCloud, enjeux technologiques / cloud (2025). https://cyber.gouv.fr/enjeux-technologiques/cloud/

[29] Usine Digitale, Le Conseil d'État valide Microsoft pour l'hébergement des données de santé et juge le risque lié au CLOUD Act acceptable (March 2026). https://www.usine-digitale.fr/cybersecurite/data-protection/donnees-de-sante/le-conseil-detat-valide-microsoft-pour-lhebergement-des-donnees-de-sante-et-juge-le-risque-lie-au-cloud-act-acceptable.T6ZKEMHAKRFAJKURNJOFS3WFEU.html

[30] Arcep, Observatoire des marchés des communications électroniques en France, Année 2024 (2025). https://www.arcep.fr/cartes-et-donnees/nos-publications-chiffrees/observatoire-des-marches-des-communications-electroniques-en-france/marche-communications-electroniques-france-2024-resultats-definitifs.html

[31] YouGov, Mythbusting claims about Gen Z and their phone habits (2 March 2026). https://yougov.com/en-gb/articles/54114-mythbusting-claims-about-gen-z-and-their-phone-habits

[32] Meta (via Rest of World), How WhatsApp for Business changed the world (3 billion monthly active users, 2024-2025). https://restofworld.org/2024/how-whatsapp-for-business-changed-the-world/

[33] Juniper Research, Conversational use cases fuel global messaging boom (1 September 2025). https://www.juniperresearch.com/press/conversational-use-cases-fuel-global-messaging-boom/

[34] Arcep, Observatoire des communications électroniques, T4 2025 (SMS drop −27%, rise of RCS, 9 April 2026). https://www.arcep.fr/cartes-et-donnees/nos-publications-chiffrees/observatoire-des-marches-des-communications-electroniques-en-france/t4-2025.html

[35] Gartner, Self-Service and Live Chat Will Surpass Traditional Channels by 2027 (27 August 2025). https://www.gartner.com/en/newsroom/press-releases/2025-08-27-gartner-survey-finds-self-service-and-live-chat-will-surpass-traditional-channels-as-top-customer-service-technologies-by-2027

[36] Vita Hinze-Hoare, CSCR: Computer Supported Collaborative Research (arXiv:cs/0611042, 2006) and CRESS (arXiv:0708.1624, 2007), University of Southampton. https://arxiv.org/abs/cs/0611042

[37] J. He, S. Houde, J. D. Weisz et al. (IBM Research), AI and the Future of Collaborative Work: Group Ideation with an LLM in a Virtual Canvas, CHIWORK '24 (ACM), 2024. https://dl.acm.org/doi/10.1145/3663384.3663398

[38] V. Jackson, Y. Cha, R. Prikladnicki, A. van der Hoek, The Role of LLMs in Collaborative Software Design (arXiv:2604.09120, April 2026). https://arxiv.org/abs/2604.09120

[39] Advait Sarkar (Microsoft Research / Cambridge), How to stop AI from killing your critical thinking, TED (2026); and The Impact of Generative AI on Critical Thinking, CHI 2025. https://www.ted.com/talks/advait_sarkar_how_to_stop_ai_from_killing_your_critical_thinking

[40] BCG Henderson Institute, AI Will Reshape More Jobs Than It Replaces (April 2026). https://www.bcg.com/publications/2026/ai-will-reshape-more-jobs-than-it-replaces

[41] R. Chen, K. Srinivasan, H. Zakerinia, Displacement or Complementarity? The Labor Market Impact of Generative AI, Harvard Business School Working Paper 25-039 (2025). https://www.hbs.edu/ris/Publication%20Files/25-039_05fbec84-1f23-459b-8410-e3cd7ab6c88a.pdf

[42] World Economic Forum (with PwC), Artificial Intelligence and the Future of Entry-Level Work (June 2026). https://www.weforum.org/publications/artificial-intelligence-and-the-future-of-entry-level-work-a-framework-for-safeguarding-and-reinventing-early-career-pathways/

[43] PwC, Global AI Jobs Barometer (June 2025; skills +66% faster in exposed occupations). https://www.pwc.com/gx/en/services/ai/ai-jobs-barometer.html

[44] J. Schulz, Ø. N. Wiborg, L. Robinson, Zooming Versus Slacking: Videoconferencing, Instant Messaging, and Work-from-Home Intentions, American Behavioral Scientist, 2023. https://journals.sagepub.com/doi/10.1177/00027642231155364

[45] Silvia Gherardi, Organizational Knowledge: The Texture of Workplace Learning, Blackwell, 2006. https://pmc.ncbi.nlm.nih.gov/articles/PMC7436565/

[46] M. Benedetto-Meyer & A. Boboc, Nouvelles formes de mise en visibilité et coopération au travail, in Sociologie du numérique au travail, Armand Colin, 2021. https://shs.cairn.info/sociologie-du-numerique-au-travail--9782200630096-page-123

[47] Croissance et Transitions / Junyr, Internal case study: substitutability of a micro-business's business email (corpus of 874 strictly business conversations, April-June 2026; rule-based classifier, 7 classes; importance/urgency signals from the email-intelligence pipeline). Monographic study (N = 1): non-generalisable proportions, estimated (non-experimental) substitutability rates, classifier precision ~85-90%, document measurement by lexical markers (a floor). Internal data, not web-verifiable.

SaaS & AI: from technical debt to orchestration

Tue, 23 Jun 2026 00:00:00 GMT

Writing code is becoming almost free. A “thin wrapper” around a language model can be replicated in a matter of days, when it is not absorbed by the model provider itself. For a software vendor, the conclusion is clear: the AI feature is no longer a moat.

But this observation conceals another, more actionable one. If generating code costs almost nothing, then the real battle shifts to two fronts that most vendors neglect: clearing decades of technical debt to become modernisable again, and turning developers into orchestrators rather than replacing them. That is where, not in the umpteenth “AI” button, the transformation is won.

The productivity paradox: coding faster ≠ delivering faster

The data from spring 2026 is unambiguous: individual productivity is soaring, while value for the business does not follow.

The individual gain does not convert into value for the business. The WRITER survey (April 2026, 2,400 respondents) shows individual productivity multiplied many times over, but only 29% of organisations draw significant ROI from generative AI (23% for agents) [1].
Very few scale up. The KPMG Global AI Pulse Q1 2026 (2,110 business leaders) classifies only 11% of organisations as “AI leaders”, capable of operating agents and capturing their advantage [2].
Failure comes from execution, not from the model. HCLTech (May 2026, 467 decision-makers at companies above $1 billion) anticipates that 43% of large AI initiatives will fail, for lack of ownership, trained users and integrated workflows [3].
And the cost shifts downstream. The Faros “Acceleration Whiplash” telemetry (April 2026, 22,000 developers, 4,000+ teams) confirms it: throughput rises (epics per developer +66%), but bugs per developer climb by 54%, code churn by 861%, and 31.3% of pull requests are merged without review [4].
Value follows a J-curve. The DORA report from Google Cloud (May 2026) models a “verification tax” and an “instability tax”: value falls before rising again, and a ROI of around 39% in the first year (with a return on investment in ~8 months for an organisation of 500 developers) arrives only with a redesign of the process, not with tooling alone [5].

The takeaway, anti-hype: the cost of writing code is collapsing; the cost of delivering reliable software is not. This is the 10-20-70 rule (BCG, 2026) applied to software: 10% technology, 20% data, 70% people and processes. Technology alone does not produce performance.

The real battle: technical debt, revealed and amplified by AI

AI is a merciless revealer. Plugged into a fragmented information system, it brings to light, at speed, the flaws in architecture and the lack of engineering discipline. And what is new in 2026 is that it manufactures debt as fast as it can clear it.

The empirical study “Debt Behind the AI Boom” (arXiv, 30 March 2026) analyses roughly 302,600 commits authenticated as AI-generated, across 6,299 repositories: the technical debt introduced by AI tends to accumulate rather than be corrected [6]. The +861% code churn measured by Faros is the operational trace of this. And the “Triple Debt” framework from Margaret-Anne Storey (March-April 2026) broadens the notion: to technical debt are now added a cognitive debt (no one any longer understands the code produced) and an intent debt (the gap between what the code does and what it was meant to do) [7].

For a vendor, six classic forms of debt directly block AI:

Type of debt	Blocking impact on AI
Architecture (coupled monoliths)	Impossible to insert AI services without destabilising the whole
Code (duplication, zero documentation)	Agents learn and propagate the bad practices
Process (no CI/CD)	Iteration and correction too slow
Infrastructure (legacy, no containers)	Neither the power nor the latency required
Security (archaic access rights)	Exfiltration risk, GDPR / AI Act non-compliance
Tests (low coverage)	Hallucinations and regressions shipped to production

To this is added the invisible twin, the data debt (dark data, duplicates, absence of a reference repository): an AI plugged into a contradictory database produces fragile and biased answers.

Ultracoding to the rescue of legacy, provided it is governed

Here is the decisive reversal: the AI that reveals the debt is also the best tool for eradicating it. Claude Opus 4.8 (Anthropic, 28 May 2026) reaches 88.6% on SWE-bench Verified and 69.2% on SWE-bench Pro; its maximum-effort mode, “ultracode”, pushes reasoning further and orchestrates a swarm of sub-agents in parallel. Above all, its central argument is not speed but reliability: around 4 times less likely to let a defect slip through [8], precisely the downstream cost (review, bugs, churn) that Faros quantifies.

In concrete terms, ultracoding and specialised agents finally make it possible to reverse-document a legacy estate (traversing millions of lines, reconstructing the business rules, mapping the dependencies), to isolate dead code, to accelerate refactoring, to generate non-regression tests and to secure the migration. Debt ceases to be an invisible ball and chain and becomes a measurable and steerable asset.

But this is the golden rule, and the whole lesson of the “Acceleration Whiplash”: this power has value only when governed. Without a control plan, ultracoding repays the old debt by creating a new one, faster. One never presumes the reliability of an agent: Plan → Execute → Verify, one measures it.

Modernising the foundations and prioritising by processes

The engine does not move without reliable fuel, nor without knowing where to apply it. On the data side, the target is an “AI-ready” platform: Data Fabric (integration and governance automated across silos) or Data Mesh (data treated as a product, under contracts, decentralised to the business domains). On the process side, one maps reality (because the declared version lies) through Process Mining on ERP/CRM traces, then prioritises through a value × effort matrix: two to four cases with high impact and low effort, and any project without ROI within less than 12 months is deferred. This is the discipline that separates the 11% of “AI leaders” from the 43% of initiatives that fail: not more AI, but AI where the ROI is incontestable.

The human pivot: from developer-producer to orchestrator

This is the heart of the matter, and the most poorly handled. When code generation becomes commoditised, the value of the developer does not disappear: it moves upward. The historical parallel is crystal clear: Dorothy Vaughan, at NASA at the turn of the 1960s, saw IBM computers destroy her task as a “human computer”; by learning FORTRAN, she became a pioneer and a leader.

Recent data validates the shift. The CodeSignal survey (April 2026, 450 engineers) shows that 91% already use agentic coding tools, and that 73% judge that non-adopters risk becoming uncompetitive: the profession explicitly moves “from coder to AI orchestrator” [9]. The BCG Henderson Institute (April 2026), across roughly 165 million jobs analysed, projects 50 to 55% of positions substantially transformed within two to three years, with a net effect: senior roles expand while junior roles contract, exactly the signature of an economy of orchestration and supervision [10].

The paradigm therefore shifts from producing code to critical supervision: scoping the intent, contextualising the outputs of the swarm of agents, validating reliability and compliance, arbitrating the architecture, keeping the stop rule. Three levers, in order:

Rewrite the narrative. Banish the “cost reduction” rhetoric (synonymous with fear of layoffs) in favour of augmentation: AI as an “exocortex” that absorbs the tedious mental load and gives back time for expertise.
Structure the cultural adoption (ADKAR model: Awareness, Desire, Knowledge, Ability, Reinforcement): involve developers from the scoping stage, rely on ambassadors, teach the limits (bias, hallucinations) as much as prompting, learn in a sandbox with no risk to production.
Move up a maturity level: from anecdotal use (summarisation, translation) to tactical use (semi-automation via dedicated assistants), then strategic (reinvention of entire swathes). This steering falls within strategic workforce planning; it is not left to the IT department alone.

Where the moat lies now

If functionality becomes commoditised, the defensible advantage shifts to five assets that your competitor's model does not reproduce:

Proprietary data and network effects.
Distribution, the brand, the channel.
Deep integration into the customer's workflow (switching costs).
Compliance and sovereignty: AI Act provider status, ISO/IEC 42001 certification.
Reliability in production: AgentOps, operated by developers turned orchestrators.

Business model and FinOps: do not bill at a loss

AI embedded in the product turns tokens into a cost of goods sold (COGS) that erodes the SaaS gross margin. The signal of June 2026 is clear: the phase of blind spending is over, value now takes precedence. Vendors are switching to usage-based pricing because the inference bill is becoming the line item that decides profitability (one major player consumed its annual AI budget in four months) [11]. And the order of magnitude of the problem is documented: the gross margin of AI products rises from 41% (2024) to 45% (2025), projected at 52% in 2026, that is, 23 to 33 points below the benchmark of a mature SaaS (75-85%), with inference already accounting for ~23% of revenue among B2B AI vendors in the scaling phase (ICONIQ, January 2026) [12].

To bill by usage without controlling the cost of inference is to bill at a loss. The lever: heterogeneous architecture (small specialised models by default, the frontier model reserved for difficult reasoning), aggressive caching, an LLM Gateway to measure the cost per request and per customer, and pricing aligned with cost.

Compliance as an advantage, not a constraint

Embedding AI makes the vendor a provider within the meaning of the AI Act (Article 50, applicable from 2 August 2026): informing the user that they are conversing with an AI, marking generated content in a machine-readable manner, disclosing deepfakes, with exceptions for standard assistive editing. Coupled with the ISO/IEC 42001 standard, compliance becomes a barrier with respect to large accounts, all the more so as the primary barrier cited by business leaders remains the security and confidentiality of data (75% in the KPMG Q1 2026) [2]. On sovereignty, EU law does not impose general localisation but rather safeguards (GDPR, Data Act); a sovereign framework (SecNumCloud) turns it into a moat of trust, decisive in the public sector, healthcare and finance.

The roadmap, in five phases

The MATIA Method™ frames the transformation over 12 months: 360° diagnostic (DORA metrics, audit of technical debt and data debt, AI Act exposure, team maturity) → Scoping (one internal case of clearing legacy through ultracoding + one product case, prioritised by value and effort) → Preparing the foundations (AI manifesto, Data Fabric/Mesh, LLM Gateway + FinOps, AgentOps control plan, AI Act register, sandboxes for developers) → Pilot deployments (measuring the net gain: coding time saved minus review time added, developers moving up towards orchestration) → Consolidation (industrialising AgentOps, aiming for ISO 42001, revising pricing, institutionalising the orchestrator role).

Every use case in production passes through the triptych Plan → Execute → Verify: one never presumes the reliability of an agent, one measures it.

In summary

When the feature no longer makes the moat, the winning vendor does not add “AI” buttons. It does two things that no one copies in a weekend: it clears its technical debt through ultracoding, without creating a new one, because it governs it; and it raises its developers to the rank of orchestrators. The figures of spring 2026 confirm it: individual productivity is soaring, but only those who rethink process, reliability and skills draw value from it. AI in the code is 10% of the matter. The 70% decide everything.

To build your roadmap and industrialise your AI use cases, see the MATIA Method™, the white paper “From POC to industrialisation” and the AI Express Audit & Roadmap.

Sources

WRITER (with Workplace Intelligence), “2026 AI Adoption in the Enterprise”, 7 April 2026, 2,400 respondents: 29% significant ROI (generative AI), 23% (agents).
KPMG International, “Global AI Pulse Survey, Q1 2026”, April 2026, 2,110 business leaders: 11% “AI leaders”; security and confidentiality = primary barrier (75%).
HCLTech, “AI Impact Imperatives 2026”, May 2026, 467 decision-makers (companies above $1 billion): 43% of AI initiatives judged doomed to fail.
Faros AI, “The AI Engineering Report 2026, The Acceleration Whiplash”, 12 April 2026, 22,000 developers / 4,000+ teams: bugs per developer +54%, code churn +861%, 31.3% of PRs without review.
Google Cloud (DORA team), “The ROI of AI-Assisted Software Development”, May 2026: J-curve, verification and instability tax; ~39% ROI in the first year, return in ~8 months.
Liu, Widyasari, Zhao, Lo et al. (SMU), “Debt Behind the AI Boom: A Large-Scale Empirical Study of AI-Generated Code in the Wild”, arXiv, 30 March 2026, ~302,600 AI commits across 6,299 repositories.
Margaret-Anne Storey, “From Technical Debt to Cognitive and Intent Debt”, arXiv / ACM Queue, March-April 2026: “Triple Debt” model.
Anthropic, “Claude Opus 4.8”, 28 May 2026: 88.6% SWE-bench Verified, 69.2% SWE-bench Pro, “ultracode” mode, reliability around 4×.
CodeSignal, survey on agentic coding, April 2026, 450 engineers: 91% use agentic tools, shift “coder → orchestrator”.
BCG Henderson Institute, “AI Will Reshape More Jobs Than It Replaces”, April 2026, ~165 million jobs: 50-55% transformed within 2-3 years; senior roles rising, junior roles falling.
CIO (Foundry), “The AI adoption spending spree is over. Time to focus on value”, 12 June 2026: tokens become a COGS, shift towards metered pricing.
ICONIQ Growth, “State of AI”, January 2026: gross margin of AI products (41%→45%→52%), inference ~23% of revenue among B2B AI vendors in scaling.

Regulation cited: Regulation (EU) 2024/1689 (AI Act), Article 50, transparency obligations applicable from 2 August 2026; ISO/IEC 42001; GDPR; Data Act (Reg. EU 2023/2854); SecNumCloud (ANSSI).

AI & promotional products: the anti-disintermediation playbook

Tue, 23 Jun 2026 00:00:00 GMT

The European market for promotional-products distributors is worth $14.24 billion in 2024 (+1.22% year on year). But behind this façade of stability, France (the 3rd-largest national market, at $1.67 billion) is declining by −1.7% in 2024, after −1.7% in 2023 [1]. Two consecutive years of contraction.

And cost pressure is intensifying: in the second quarter of 2025, 70% of the sector's suppliers raised their prices, driven by customs duties, while average sales declined (ASI, State of the Industry 2025) [6]. The context is North American, but the mechanics (sourcing under strain, margins eaten away) apply to the French industry too.

This figure changes everything. In a market that is no longer growing and where costs are rising, AI is not an engine of growth. It is a lever for margin, efficiency and retention. And the real story is not even there. It lies in a more structural threat: disintermediation.

The threat is not the technology, it is the loss of the upstream end of the relationship

The promotional-products buyer (a communications, marketing or HR manager for whom this purchase is secondary) is changing their starting point. Increasingly, they begin their search in a conversational assistant such as ChatGPT, Perplexity or Copilot, rather than on a search engine or by calling a supplier.

The direct consequence: the distributor that is no longer cited by these assistants loses the upstream end of the relationship. This is the “zero-click” risk, and it is precisely the challenge of AEO/GEO (Answer/Generative Engine Optimization): being findable and recommended by AI becomes a matter of commercial survival, not an SEO nicety.

A second force adds to this one: platform capture. The large sourcing hubs, coupled with AI matching, concentrate standardised catalogue data and audience. The “order-taker” distributor, with no added value of its own, risks becoming the mere front-end of a platform that owns the data and the algorithm. The margin migrates to whoever owns the data and the relationship.

The response can be summed up in a single sentence: shift value away from transactional intermediation (finding the product, which AI commoditises) towards what AI does not commoditise: creative advice, CSR compliance, marking-quality assurance, logistics, and proprietary data (proof history, supplier compliance, client preferences).

The paradox of choice: why an overabundant catalogue drives the buyer away

The industry is extremely fragmented and piles up catalogues of tens of thousands of references. Yet research in decision psychology is unambiguous: option overload paralyses the buyer, especially when the task is complex, preferences are uncertain, and they are seeking to minimise their effort. That is exactly the profile of the promotional-products buyer (meta-analysis by Chernev, Böckenholt & Goodman, 2015) [2].

AI, used well, acts as a cognitive filter: it turns a natural-language brief into a justified shortlist. But the underlying data must be clean. Hence the order of priorities.

The 4 priority levers (extract from the MATIA Method™ playbook)

1. The catalogue first: AI-augmented PIM. This is the non-negotiable foundation. Ingesting heterogeneous supplier catalogues, cleaning up descriptions, standardising attributes (material, MOQ, marking techniques, CSR labels), completing missing data, deduplication. Without this step, everything else hallucinates.

2. Self-service generation of proofs and visuals. Applying the client's logo to the product, colour consistency (avoiding the gap between the digital rendering and the real colour of the fabric), automated pre-flight checking (DPI, bleeds, marking area). The bottleneck of back-and-forth with designers eases.

3. Quote automation. A brief → product proposal + margin calculation + shipping + marking grid → quote. To be connected to the standardised catalogue.

4. Conversational sourcing and CSR compliance. An agent that produces explainable recommendations, and a per-order carbon calculation connected to the ERP, with automated auditing of supplier certifications (GOTS, Oeko-Tex, FSC).

These building blocks are not theoretical: the sector's business platforms are already deploying them. AI product search is in production at commonsku, mock-up generation is in beta there, and PIMs such as Afineo are integrating agents that write product sheets. Maturity varies, and most of the promised gains remain vendor figures to be verified against your own scope.

The trap to avoid: believing that “giving the sales team access to ChatGPT” is enough. That is precisely what produces zero return: the tool without the process redesign or the clean data.

Two regulatory points to address right now

The AI Act, Article 50, without overplaying it. The obligation to disclose “deepfake” content (applicable from 2 August 2026) has a narrow scope: a standard AI product visual, a background, an upscaling or a typographic banner are not deepfakes [3]. The real precaution is contractual: setting out in the terms and conditions of sale the use of generative-AI tools and liability for the elements supplied by the client. The real risk is to “sell rights that do not exist” on AI creations [4].

Greenwashing: enforcement is tightening. The DGCCRF inspected more than 3,000 establishments in 2023-2024, with more than 15% serious breaches, resulting in 430 injunctions, 500 warnings and 70 formal notices [5]. Environmental claims on promotional products are squarely in the crosshairs. AI must help to prove CSR performance (traceability, carbon calculation, anticipating the Digital Product Passport), not to embellish it.

The roadmap, in five phases

The MATIA Method™ structures the transformation over 12 months: 360° diagnostic (audit of the catalogue and the quote/proof cycle) → Scoping the use cases (2-3 priorities, business champions) → Preparing the foundations (catalogue standardisation, sovereign technology stack, governance) → Pilot deployments (restricted, measured scope) → Consolidation and governance (extension, AI committee, AI Act register).

The objective is not to leap to the “Pionnier” level in a single bound, but to move the organisation from the Artisan level (individual Shadow AI) to the Orchestre level: AI embedded in processes, steered and measured.

In summary

In a contracting market where AI threatens intermediation itself, the question is not “which tool to buy”. It is: how to own what AI cannot commoditise (the data, the relationship, compliance, advice). The distributor that structures its data and shifts its value towards advice and CSR proof does not suffer disintermediation: it becomes the trusted third party for it.

To position your organisation on the 5-level maturity Scale and build your roadmap, see the MATIA Method™, the SME AI Maturity white paper and the AI Express Audit & Roadmap.

Sources

ASI Research, “European Distributors' Annual Sales… topping $14.24 B” (June 2025) and “Europe's Distributors Top $14B” (August 2024). French estimates in euros: C-Mag/2FPCO (~€1.6 billion), PPAI (~€2.09 billion), different methodologies, same rank and trend.
A. Chernev, U. Böckenholt, J. Goodman, “Choice overload: A conceptual review and meta-analysis”, Journal of Consumer Psychology (2015).
Regulation (EU) 2024/1689 (AI Act), Article 50, transparency obligations applicable from 2 August 2026; “deepfake” scope restricted to realistic content that is falsely authentic.
Maddyness, “Création & IA : le piège de la vente de droits qui n'existent pas” (16 October 2025).
DGCCRF, press release “Lutte contre l'écoblanchiment” (1 October 2025).
ASI, “2025 Counselor State of the Industry” (7 August 2025), price increases linked to customs duties, changes in sales; PIWorld summary (12 August 2025). North American data.

Asynchronous work: in the age of AI agents, everyone becomes a manager

Tue, 23 Jun 2026 00:00:00 GMT

In-depth version, June 2026. This text extends and broadens my article “Junyr Agents™: delegating AI without losing control” (1 June 2026). Where the first described how to delegate to an agent, this one sets out the broader thesis that such delegation establishes: when the agent works in deferred time, work becomes asynchronous. Everyone, from the intern to the business leader, then inherits a job they did not choose: that of manager.

The idea, in one page

An AI agent does not work like a piece of software you drive from the keyboard, nor like a colleague you wait for in a meeting. You give it an objective, it goes off to work on its own for a few minutes, sometimes a few hours, and it comes back with a result to review. This simple change of tempo has a consequence that few organisations saw coming: work stops being synchronous. You no longer “do” the task in real time; you entrust it, you track it, you validate it. And entrusting, tracking, validating are not the acts of a doer. They are the acts of a manager.

This is the real upheaval of 2026, and it runs deeper than the badly framed question of job “replacement”. As agents absorb execution, every employee finds themselves at the head of a small team that does not sleep, does not unionise, and does not wait until Monday to deliver. Microsoft, which surveyed 20,000 workers across ten countries including France in 2026, puts it plainly: the human role shifts “from producing answers to evaluating, refining and owning them” [1]. Everyone becomes a manager. The good news is that it is a promotion. The less good news is that being a manager is a job. No one ever taught it to us.

From the synchronous assistant to the asynchronous agent

Let us recall the distinction, because everything follows from it. An assistant operates in a synchronous loop: you ask it a question, it answers, you keep control at each exchange. This is the classic conversation with a chatbot. An agent, on the other hand, operates in an asynchronous loop: you give it an objective, and it strings together the actions on its own, across several tools and several steps, until it reaches it. The difference is not a matter of power, but of nature: the agent acts in the real world, and it acts while you are doing something else.

This temporal gap is not a technical detail: it is the heart of the matter. When the answer is instantaneous, you remain an assisted doer. When it arrives in deferred time, you necessarily become something else: someone who has launched a piece of work, who awaits it, and who will have to answer for it. The researchers who study asynchronous software agents observe as much: the issue is no longer the speed of a request, but the ability to formulate a clear mandate, to break it down, and to verify what comes back [2]. Exactly the work of a supervisor.

The phenomenon is no longer marginal. Across the Microsoft 365 ecosystem, the number of active agents has multiplied by 15 in one year, and by 18 in large enterprises [1]. On maturity, the picture stays honest: according to McKinsey, 62% of companies are experimenting with AI agents in 2026, but only 23% scale them in at least one function [3]. The gap between these two figures is not down to the technology: it works. It is down to an organisational skill that few yet master: knowing how to make agents work in deferred time without losing the thread. In other words, knowing how to manage.

The shift: from doer to orchestrator

Let us put it simply. For two centuries, technical progress consisted in giving better tools to the doer. The AI agent does something different: it takes over execution, and hands back to the human the role that sat above. Microsoft calls this “the new agency equation”: as agents take on execution, the human gains latitude: “more room to direct the work, make the calls, and carry the results” [1]. The practitioners who already orchestrate several agents in parallel describe the same tipping point: you move “from the conductor, one musician guided in real time, to the orchestrator of an entire ensemble, coordinated asynchronously”; the mental model is no longer that of pair programming, it is that of managing a team [4].

This shift has a name that analysts have ended up popularising: everyone becomes an agent boss, a “boss of agents”. And it is not merely a turn of phrase. In the Microsoft survey, the most advanced workers (those who use agents for multi-step tasks and build multi-agent systems) still represent only 16% of the panel, but they give a glimpse of the job to come: they redesign their processes around intent and review, and 80% of them report producing work they could not have produced a year earlier [1]. Academic research on parallel agent loops reaches the same point by another route: performance depends on two factors, the genuinely parallelisable structure of the task, and “the delegation capacity of the central manager” [5]. The bottleneck is no longer the machine. It is the manager.

This is also, very precisely, what the MATIA Method™ describes when it places an SME on its five-level AI maturity scale: Spectateur, Artisan, Orchestre, Architecte, Pionnier. The move to the agentic is not a leap from Spectateur to Pionnier: it is the accession to the Orchestre level, the one where you stop tinkering with isolated uses and start coordinating a whole. The word was already in the model. Asynchronicity has merely given it its full reach.

Work now happens while you sleep

If the role changes, so does the tempo. The defining feature of the 2026 company is no longer flexibility of place, the legacy of remote work, but flexibility of time. Work unfolds continuously, often invisibly, sometimes outside human hours: one agent drafts a document at night, another monitors an indicator, a third prepares a report for Monday morning [6]. In software development, which often serves as an advanced laboratory, several agents run in parallel on separate branches, each on its own task, and the developer moves from writing to coordination: giving the direction, reviewing, correcting the trajectory [4].

This parallelisation has a virtue and a trap. The virtue: you no longer do one thing at a time. The trap: it is easy to confuse “many agents launched” with “much value produced”. The teams that measure productivity have learnt this at their own expense: the METR institute had to overhaul its entire experimental protocol in 2026, because time spent per task became uninterpretable once a single worker drives several competing agents [7]. The lesson is clear: value does not come from the number of agents, but from the quality of supervision. A manager who opens ten worksites they do not review is not ten times more productive. They are ten times more exposed.

The other side of the promotion: being a manager is a real job

It must be said with the same frankness as the rest, because this is where most projects stumble. Becoming a manager of agents is not merely reaping the time saved. It is inheriting the burdens of management. They are real.

The verification tax. Everything an agent produces must be reviewed by someone who answers for it. The more agents execute, the more the stakes move towards human evaluation: approving a bad result is harmless once, but at scale, the errors that slip through compound [1]. In the Microsoft survey, the two skills judged most important in the age of agents are not technical: they are quality control of AI outputs (50%) and critical thinking (46%); and 86% of advanced users say they treat any AI output “as a starting point, never a final answer” [1]. The manager is not the one who produces. It is the one who reviews, who decides, and who signs off.

The limit of supervision. You do not monitor asynchronous work the way you monitor synchronous work. The review comes after the fact, and it fails when the agent can cause harm before being reviewed, that is, when the time to harm is shorter than the time to control [8]. Recent work points to the “practical impossibility” of real-time, meaningful and continuous human supervision over long autonomous processes [8]. The answer is not to give up on delegating: it is to design the delegation so that no irreversible action escapes a control point. A good manager does not review everything; they know what to review before it becomes irreversible.

The deskilling risk. If the machine does the work, the human can unlearn. The literature speaks of deskilling (the loss of acquired expertise) and, more worrying, of never-skilling: beginners who never acquire the fundamentals because they lean on automation too early [9]. The most clear-sighted workers have understood this and guard against it: in the Microsoft survey, 43% of advanced professionals (against 30% of the others) say they deliberately do part of their work without AI to keep their skills sharp [1]. The good manager does not merely hand out the work: they keep their hand in the craft, otherwise they lose the ability to judge what they hand out.

The burden, finally. Not everyone dreams of becoming a manager. Coordinating, arbitrating, being accountable for others (even if they are agents) is work in itself, which does not mechanically lighten the mental load; it shifts it. This is the paradox Microsoft calls the “transformation paradox”: 65% of workers fear falling behind if they do not adapt quickly, but 45% find it safer to focus on their current objectives than to reinvent the way they work. Only 13% feel rewarded for that reinvention, even when it yields no immediate result [1]. The promotion to manager cannot be decreed. It is organised, or it exhausts.

What sets apart those who succeed: discipline, not the tool

The most useful data point of the year, for a business leader, fits in one sentence. When Microsoft measures what really explains the impact of AI, organisational factors (culture, managerial support, HR practices) weigh more than double the individual factors (67% against 32%) [1]. In other words: value does not come from the employee most gifted with AI, it comes from the environment that turns that talent into results. The same report shows that when managers use AI openly and set quality standards, teams' trust in the agentic climbs by 30 points [1] [10].

What the best do, concretely, is nothing mysterious. They do not seek to “do more things faster”; they redefine their value around what only a human brings: setting a clear intent (the expected result and the required standard) and designing the division of work between humans and agents [1]. They refuse to outsource their judgement. And they impose on themselves an execution discipline that we, at Junyr, formalise in the form of a triptych applied to each step of an agent: Plan → Execute → Verify. The agent first announces what it is about to do and the success criteria; it executes; then a distinct step checks the result before continuing. Verification becomes a native step of the process, not a control added after the fact, which is what makes the delegation auditable, and therefore sustainable. This is the principle on which Junyr Agents™ rests: a written mandate, documented supervision, a log of every act.

For an SME business leader: you are not buying agents, you are building a team

The practical consequence is simple to state and demanding to hold to. Deploying agents is not buying one more piece of software; it is taking charge of a team. Everything we know about management becomes relevant again. Three decisions follow.

First decision: write the job descriptions. An agent without a clear mandate is an employee without a job description: you can neither steer it nor audit it. Each agent receives an objective, an authorised scope of action, explicit limits and escalation cases. This looks bureaucratic; it is exactly the opposite: it is what makes it possible to delegate more, because you delegate with confidence.

Second decision: build the review infrastructure. As agents execute, the question is no longer “can the machine do it?” but “who answers for what it does?”. Microsoft sums up the task in three questions every organisation will have to settle: who reviews the performance of the agents? who has authority to modify the process they execute? how does a local gain spread to the whole organisation? [1]. An SME that can answer these three questions builds what Microsoft calls a “proprietary intelligence”, an in-house know-how that capitalises and that a competitor cannot copy [1].

Third decision: treat agents as entities to be governed. The software vendors themselves are switching to this logic: in May 2026 Microsoft rolled out a dedicated control platform for managing deployed agents, with identities, permissions and traceability [11]. For an SME, the point is not to have the same factory as Microsoft; it is to apply, at its own scale, the same rigour: an agent has an identity, a scope, a log. The rest is only a matter of method. It is precisely the move from the Artisan level (isolated uses) to the Orchestre level (a coordinated team), then Architecte (processes redesigned around agents) of the MATIA Method™.

The French context makes this milestone all the more accessible in that it is still early: 26% of French micro-businesses and SMEs declared they used at least one AI tool in 2025 (France Num) [12]. The window is not closing; it is wide open for those who methodically structure their trajectory rather than piling up tools. And employment is not disappearing in silence: LinkedIn counts 1.3 million AI-related opportunities created in two years, in jobs that did not exist five years ago [1] [13]. Work does not evaporate. It moves up a notch.

Three questions to ask this very week

To turn this observation into action, three questions (to put to your AI champion, your IT department or your provider) are enough to reveal your real maturity, with no tool and no budget:

Which processes have we genuinely entrusted to an agent that works in deferred time, and who, by name, reviews the result before it produces an effect? If the answer is “no one in particular”, you have a team without a manager.
Before which action must an agent absolutely stop for human validation? If the list does not exist, your supervision arrives after the harm, not before.
What are we doing so that our teams keep the skill to judge what the agents produce? If the answer is “nothing”, you are preparing a silent deskilling.

Conclusion: to lead, rather than to execute

The industry is turning a page without always noticing. The word “productivity” will survive, as “prompt” still survives. But the practice has already tipped over: in the age of asynchronous agents, you no longer do the work, you direct it. It is a promotion offered to everyone (from the intern to the business leader) and it is, at the same time, a demanding job that must be learnt: writing mandates, reviewing what matters, keeping a hand on judgement, organising trust.

The question is neither fear nor urgency; it is mastery. The agents provide the tireless arms; they do not provide the manager. That manager, from now on, is you. The only decision that matters this year is to take it on with method rather than to endure it. Everyone becomes a manager. It remains to learn how to be one.

Going further

AI Express Audit & Roadmap (MATIA Method™): 60 minutes over video call to place your SME on the maturity scale (Spectateur → Pionnier) and identify the two or three processes to delegate as a priority, at croissance-transitions.fr.
The 5-level AI maturity scale: understanding the Orchestre level and those that follow, at paulantoinetual.fr/blog.
Junyr Agents™: delegating operable, supervised and auditable AI agents, on the Junyr ERP (junyr.app).

Sources: verifiable, June 2026

[1] Microsoft, 2026 Work Trend Index Annual Report, “Agents, human agency, and the opportunity for every organization” (5 May 2026). Survey of 20,000 workers, 10 markets including France; M365 telemetry. https://www.microsoft.com/en-us/worklab/work-trend-index/agents-human-agency-and-the-opportunity-for-every-organization

[2] “Effective Strategies for Asynchronous Software Engineering Agents”, arXiv 2603.21489, 2026. https://arxiv.org/pdf/2603.21489

[3] McKinsey, The State of AI, the agentic era (2026), 62% of organisations are experimenting with AI agents, 23% scale them in at least one function. https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai

[4] Addy Osmani, “The Code Agent Orchestra, what makes multi-agent coding work” (2026). https://addyosmani.com/blog/code-agent-orchestra/

[5] “Self-Manager: Parallel Agent Loop for Long-form Deep Research”, arXiv 2601.17879, 2026. https://arxiv.org/pdf/2601.17879

[6] “2026: The Year Work Goes Asynchronous, and AI Agents Take Over”, The Fast Mode (2026). https://www.thefastmode.com/expert-opinion/47211-2026-the-year-work-goes-asynchronous-and-ai-agents-take-over

[7] METR, “We are Changing our Developer Productivity Experiment Design” (24 February 2026). https://metr.org/blog/2026-02-24-uplift-update/

[8] “Practical challenges of control monitoring in frontier AI deployments”, arXiv 2512.22154, 2026 (NIST framework). https://arxiv.org/pdf/2512.22154

[9] “From de-skilling to up-skilling: How artificial intelligence will augment the modern physician”, PMC (2026). https://pmc.ncbi.nlm.nih.gov/articles/PMC12955832/

[10] Microsoft People Science, “Research drop: Empowering managers for an AI-first future”, study of 1,800 workers. https://techcommunity.microsoft.com/blog/microsoftvivablog/research-drop-empowering-managers-for-an-ai-first-future/4468191

[11] Microsoft Agent 365, agent control platform, GA May 2026; coverage by Usine Digitale. https://www.microsoft.com/fr-fr/microsoft-agent-365 · https://www.usine-digitale.fr/big-tech/microsoft/microsoft-generalise-agent-365-pour-aider-les-entreprises-a-reprendre-le-controle-des-agents-ia-deployes-sans-supervision.5MGNTRBAVNAFXHLTSNZBIUQP7M.html

[12] France Num (2025), ~26% of French micro-businesses and SMEs declare they use AI. https://www.francenum.gouv.fr/

[13] LinkedIn, 2026 Labor Market Report (January 2026), ≥ 1.3 million AI-related opportunities created in two years. https://economicgraph.linkedin.com/content/dam/me/economicgraph/en-us/PDF/linkedIn-labor-market-report-building-a-future-of-work-that-works-jan-2026.pdf

[14] Deloitte Insights, “Agentic AI is scaling faster than guardrails” (2026), only one organisation in five has a mature model for agent governance. https://www.deloitte.com/us/en/insights/topics/emerging-technologies/ai-agents-scaling-faster.html

[15] Gartner, 40% of enterprise applications will embed specialised agents by the end of 2026 (vs < 5% in 2025); by 2029, ≥ 50% of knowledge workers will develop skills to work with, govern or create agents. https://joget.com/ai-agent-adoption-in-2026-what-the-analysts-data-shows/

Paul-Antoine Tual is an AI Transformation Leader. He runs Croissance et Transitions (SAS) and operates the Junyr™ suite, MATIA Method™ (a 5-level AI maturity methodology), Junyr Agents™ (AI agents for SMEs, junyr.app). He supports the leaders of French mid-caps and SMEs in their AI transformation, 60-minute diagnostic: croissance-transitions.fr.

Vibe coding is dead: enter UltraCoding

Tue, 23 Jun 2026 00:00:00 GMT

When the inventor buries his own word

On 2 February 2025, Andrej Karpathy (co-founder of OpenAI, former director of AI at Tesla) posted a message on X that went viral within hours. In it he described a new way of programming: describe what you want in natural language, let a model write the code, accept whatever comes back, run it again if it does not work. His phrasing hit the mark: “fully giving in to the vibes, embracing exponentials, and forgetting that the code even exists” [1]. The term vibe coding was born. Within months, it entered Merriam-Webster as a “trending” expression, then was crowned word of the year 2025 by the Collins Dictionary [3].

It is worth reading the rest of the original message, which collective enthusiasm often forgot. Karpathy specified that the practice was meant “for throwaway weekend projects”. The practitioner Simon Willison set out the most rigorous definition a few weeks later: doing vibe coding is not “using AI to code”: it is building software without reading the code the AI writes [2]. The move that defines vibe coding is not generation. It is the renunciation of review.

Yet, in early 2026, that same Karpathy has turned the page. He now declares vibe coding “passé” and proposes another term to describe the future of software: agentic engineering [4][5]. The distinction he draws is clear: vibe coding consists of describing and accepting; agentic engineering consists of designing the system, specifying the constraints, and reasoning about the architecture up front, then orchestrating agents that implement, under supervision [5].

The inventor has buried his own word. This is not a disavowal: it is a diagnosis. Vibe coding was not a mistake. It was a prototyping technique that the market mistook for a production method. And the bill for that misunderstanding can now be quantified.

The bill for vibe coding

Three serious studies, published in 2025, allow us to move beyond impressions and look at the figures.

Security first. In July 2025, the software vendor Veracode published its GenAI Code Security Report: 80 development tasks submitted to more than 100 different models. The result is stable and troubling: when a model has the choice between a safe and an unsafe way to write a piece of code, it picks the vulnerable version in 45% of cases [6]. The detail matters: the rate climbs to 72% in Java, and above all it does not improve from one model generation to the next, a finding stable across every family tested (OpenAI, Anthropic, Google). Veracode's March 2026 update extends the finding: after two years of releases billed as revolutionary, the rate stays at 45%, including on GPT-5.1/5.2, Gemini 3 and Claude 4.5/4.6 [6]. In other words, waiting for “the next, more intelligent model” will not solve the security problem. A more capable model writes more impressive code, not necessarily safer code.

Productivity next. In July 2025, the independent laboratory METR published a randomised controlled trial, the gold standard of scientific proof. Sixteen experienced open-source developers, 246 real tasks on projects they had known for five years on average. The verdict: with the AI tools of early 2025, they took 19% longer to complete their tasks. The most troubling part is not the slowdown. It is the perception gap: those same developers estimated that AI had made them 20% faster [7]. The machine does not merely slow you down: it gives the illusion of acceleration. This is exactly the trap of vibe coding: the sense of fluidity masks the real cost, which is paid later, in debugging.

Delivery last. Google Cloud's DORA 2025 report, the annual reference for measuring software performance, confirms the shift at the organisational level: adopting AI does increase throughput (you ship more), but it also increases the instability of production releases [9]. Its formulation says it all: “AI does not fix a team; it amplifies what is already there”. The following edition (Google Cloud/DORA, May 2026) confirms it by drawing a “J-curve”: the verification and instability tax comes before the gains. Without a control system (robust automated tests, continuous integration, small batches, fast feedback loops), more code shipped faster mechanically produces more breakage.

The common thread across these three findings is clear. The problem is not AI's ability to generate code. It has never been stronger. The problem is the absence of a verification system around that generation. Vibe coding removed review without putting anything in its place. This is precisely the void that the next practice fills.

From “agentic engineering” to UltraCoding

Karpathy's agentic engineering describes a stance: no longer writing the code yourself 99% of the time, but orchestrating agents that do it, while keeping control of design, constraints and supervision [4][5]. This is right. But a stance is not a method. “Keeping control” and “supervising” do not say how, nor how far, nor at what point you are allowed to ship.

It is this shift from stance to production line that we call UltraCoding. The word is deliberately strong, because the practice is. UltraCoding is an agentic software production line whose sole exit condition is a measured quality level, and which accepts consuming more resources to reach that level: more tokens, more compute, more machine time. Over-consumption is not a side effect to be corrected. It is the entry price of a quality that vibe coding structurally could not deliver.

UltraCoding comes down to four moves, plus one cross-cutting audit:

Multi-agent orchestration: several specialised AIs rather than a single generalist one.
Adversarial review: an AI whose job is to break another's code.
Three-layer testing: from the classic automated test to AI-driven visual inspection of the interface.
The threshold deployment loop: you start over until you reach the bar, then you ship.

And, cutting across each of these moves, an audit: security, internationalisation (i18n), accessibility.

Let us take each move in turn, with what justifies it.

Move 1: Multi-agent orchestration

The dominant intuition of 2023-2024 was that a single model, ever larger, would eventually do everything. The practice of 2025-2026 says the opposite: you get better results with several specialised agents working together than with a single overloaded one.

The most thoroughly documented demonstration comes from Anthropic. In the field report published in June 2025 on its multi-agent research system, the architecture is of the orchestrator-workers type: a lead agent plans, launches 3 to 5 specialised sub-agents in parallel, each with its own context window, then synthesises their work. Measured result: this arrangement outperforms the single agent (Claude Opus 4 alone) by 90.2% on their internal evaluation [8].

The same field report gives the price of this performance, and it is essential for a business leader: the system consumes roughly 15 times more tokens than an ordinary conversation, and the quantity of tokens used alone explains nearly 80% of the performance variance [8]. In plain terms: quality is paid for in resources, in an almost linear fashion. The economic rule Anthropic draws from this is the right compass for UltraCoding: the multi-agent architecture makes sense only when the value of the task exceeds the cost of the tokens. To produce software destined for production, for customers, for money in transit, this condition is almost always met.

In concrete terms, an UltraCoding team does not launch “an AI that codes”. It orchestrates an architect that breaks the work down, several implementers that write in parallel, a tester that stress-tests it, a reviewer that challenges it. Each has a role, a scope, and a context window of its own.

Move 2: Adversarial review

This is the heart of UltraCoding, and it is what most clearly separates it from vibe coding. An AI must never be its own judge. Asking it “is your code correct?” amounts to asking a candidate to mark their own paper: they will give themselves a good mark.

The countermeasure has been known to research since 2023 and has a name: the generator-critic loop. One agent produces; another agent, independent, has the sole mission of finding the flaw. OpenAI trained a model dedicated to this purpose, CriticGPT, specifically to critique and flush out the bugs in code written by other models [10]. The academic frameworks Self-Refine (generate → critique → correct) and Reflexion (remembering one's failures so as not to repeat them) formalised this mechanism of self-improvement through confrontation [11]. Code has a unique advantage here over text: it is executable. The feedback is not an opinion, it is a compilation or test result, an objective fact on which the loop can rely.

Adversarial review does not stop at functional correctness. It extends to security: a “red team” agent whose role is to think like a malicious actor: hunting for prompt injection, data leakage, the missing authorisation. Given the 45% rate of vulnerable code measured by Veracode [6], this confrontation is not a luxury: it is the only credible defence against a flaw that models, left to themselves, reproduce generation after generation.

The rule, in UltraCoding, admits no exception: the generator never validates its own output.

Move 3: Three-layer testing

Testing, in UltraCoding, does not mean “running the unit tests”. It means stacking three layers of verification, from the most mechanical to the most subtle.

Layer 1: the classic automated tests. Unit, integration, end-to-end tests. This is executable truth: does the code do what it claims to do? This layer is necessary but insufficient: it checks the logic, not the experience or the safety.

Layer 2: the audit. Two checks that vibe coding almost always ignores. First, security: static analysis, detection of secrets in plain text, verification of headers and of protections against classic attacks, the direct answer to Veracode's 45% [6]. Then internationalisation and accessibility: pseudo-translation to flush out hard-coded text, handling of right-to-left languages, date and currency formats, and compliance with accessibility criteria (WCAG) so that the interface remains usable by everyone. These audits are not “felt” in a demo; they reveal themselves in production, often at the worst possible moment.

Layer 3: AI-driven visual inspection of the interface, with UX/design critique. This is the most recent and most spectacular layer. Until 2024, testing an interface came down to comparing pixels: fragile, noisy, blind to meaning. The arrival of vision-language models (VLMs) changed the game. These models look at the rendered screen and understand it as a human would: no longer “this pixel moved by 3 points”, but “the Buy button is partly hidden by a banner, so the primary action is compromised” [14]. The tool Percy launched, in late 2025, a visual review agent that makes review roughly 3 times faster and filters out nearly 40% of false positives [14]. Beyond rendering, academic work such as UXAgent (CHI 2025) simulates panels of users and runs automated heuristic evaluations of usability before any real human test [12].

A note of honesty is required here, because this is the least mature layer. AI UX critique is not infallible: in March 2025, two Microsoft UX researchers measured accuracy rates of only 50 to 75% for heuristic evaluations conducted by AI, whereas specialised, trained systems such as Baymard's claim up to 95% [13]. The operational conclusion is clear: AI-driven visual inspection speeds up and broadens review; it does not replace human design judgement. It flags, it pre-sorts, it documents. The aesthetic decision and the experience call remain human.

Move 4: The threshold deployment loop

Here is what makes UltraCoding a method and not a collection of tools. The three preceding moves run in a loop (generate, confront, test, correct) and this loop does not stop when the developer is tired or when the deadline looms. It stops when an explicit, measurable quality threshold is reached: test coverage, zero critical vulnerabilities, accessibility compliance, visual review score. As long as the threshold is not crossed, the loop starts over. As soon as it is, deployment is triggered.

This inversion runs deep. In vibe coding, you ship when “it looks like it works”. In UltraCoding, you ship when the bar is reached, not when the human is spent. This is exactly the prescription of the DORA 2025 report: pairing AI's increased throughput with automated guardrails, small batches, fast feedback loops; failing which speed is paid for in instability [9]. The threshold makes the decision to ship objective, reproducible and defensible.

The honest counter-argument: it is expensive, but it can be internalised

It must be said plainly: UltraCoding consumes. Fifteen times more tokens for multi-agent orchestration as measured by Anthropic in 2025 [8], compute for each test layer, time for each turn of the loop, latency. The 2026 measurements put the agentic-workflow range at 5 to 30 times (TechCrunch, June 2026). It is not free, and pretending otherwise would be dishonest.

The consequence is that there are cases where UltraCoding is a waste. For a throwaway script, a macro you will run once or a weekend prototype meant to validate a hunch and then be deleted, vibe coding remains perfect, and Simon Willison is right to say that it “rocks” in that register [2]. Deploying a three-layer adversarial pipeline for that would be like putting a car through its MOT the day before it is sent to the scrapyard.

UltraCoding, by contrast, becomes non-negotiable as soon as the cost of a bug exceeds the cost of the tokens: software in production, an interface facing customers, a regulated system, a flow that touches money or personal data. There, saving 15 times the tokens in order to risk a security flaw (45% probability by default, as measured by Veracode in mid-2025 [6]) or an unstable production release [9] is not a saving. It is a transfer of cost to later, with interest.

And this cost is not even inevitable: it can be internalised. The 15-fold consumption is only a problem as long as you pay for it by the unit, as a running expense (OpEx), to a frontier API billed by the token. Nothing dictates this model. A dedicated inference server, an open-source model well calibrated to its domain, and an agentic loop developed in-house turn this over-consumption into an amortised investment (CapEx): a fixed hardware asset that runs without a meter, on infrastructure you own. The amortisation calculation often tips in favour of internalisation as soon as the volume is there, and it brings what billed cloud will never give: sovereignty. The company's code, data and secrets no longer leave its perimeter.

Better still: these loops have no reason to run at peak hours. They can execute in hidden time, when the servers are not tied up with real-time tasks: at night, at the weekend, in the troughs of load. The compute cost then becomes a hidden cost: you draw on capacity that is already paid for and otherwise unused. All that remains, at the margin, is energy, itself optimisable (off-peak hours, load management, low-carbon electricity). UltraCoding thus ceases to be a variable expense imposed upon you and becomes an owned industrial capability. The honest trade-off is that it takes sufficient volume to amortise the hardware and internal skill to calibrate the model and keep the loop running: this path is not for everyone, but it is open to anyone who wishes to escape renting.

And the human, in all this? The METR study is a salutary guardrail [7]: AI left to itself can slow things down and mislead about its own speed. UltraCoding does not remove the human. It moves them to where they create the most value: defining the threshold, writing the specification, adjudicating the cases the machine does not settle, and signing off the delivery. The developer no longer types every line; they become the architect of the standard and the judge of last resort.

What this changes for a business leader

One can read this article as a matter for engineers. That would be a mistake. UltraCoding is first and foremost a quality requirement that any business leader can, and should, formulate, without writing a single line of code.

You do not have to build the pipeline yourself. What you have to know is how to demand the bar and budget the resources when the stakes justify it. Three questions are enough to reveal the maturity of a supplier or an internal team: Who reviews the code your AI produces, and with which other tool? How do you test the interface and the security before delivery? What measurable threshold triggers a production release, and who signs it off? A team that “does vibe coding” has no answer to these three questions. A team that practises UltraCoding has them all.

This is exactly how we build Junyr Suite. Our conviction, in the MATIA Method™, is that in the AI era scarcity does not shift towards the generation of code (that becomes abundant) but towards verification: the ability to guarantee, prove and hold a level of quality. This is where the AI transformation of companies is now decided: not “producing faster”, but “producing to a standard that can be defended before a customer, an auditor or a regulator”.

And this is where UltraCoding goes beyond software. The logic that underpins it (generate, pit against an adversary, stress across three layers, ship only at the threshold) has nothing specific to code about it. It is a loop of trust transposable to any work produced by an AI. This is precisely what we do with the MATIA Method™: we implement UltraCoding in companies, then we apply the same loop to agents across every function: sales, finance, marketing, support, HR. An agent that drafts a proposal, qualifies a prospect or prepares a report follows the same cycle as an agent that writes code: production, adversarial review, control, quality threshold before any action. Generation becomes abundant everywhere in the company; verification, for its part, becomes the standard everywhere. That, at bottom, is the AI transformation: not plugging in models, but installing the loop that makes their work reliable.

Vibe coding promised to forget that code exists. UltraCoding makes the opposite bet: to remember that, behind every smooth interface, there is code that someone answers for. This responsibility cannot be delegated to a “vibe”. It is organised.

In one sentence

Vibe coding taught the world that AI can write anything. UltraCoding teaches it the only thing that matters now: how to know that what it has written is good, before shipping it.

To turn UltraCoding into a quality standard across your software projects, see the MATIA Method™ and the AI Express Audit & Roadmap.

Sources

[1] Andrej Karpathy, post on X, 2 February 2025 (origin of the term vibe coding), synthesis and quotation via Vibe coding, Wikipedia. https://en.wikipedia.org/wiki/Vibe_coding [2] Simon Willison, “Not all AI-assisted programming is vibe coding (but vibe coding rocks)”, 19 March 2025. https://simonwillison.net/2025/Mar/19/vibe-coding/ [3] Collins Dictionary, vibe coding, word of the year 2025; Merriam-Webster, “slang & trending” mention, March 2025 (via Wikipedia). https://en.wikipedia.org/wiki/Vibe_coding [4] “Vibe coding is passé. Karpathy has a new name for the future of software”, The New Stack, 2026. https://thenewstack.io/vibe-coding-is-passe/ [5] “Andrej Karpathy Has Renamed Vibe Coding. Here's What Engineering Leaders Need to Do About It”, SD Times, 2026. https://sdtimes.com/ai/andrej-karpathy-has-renamed-vibe-coding-heres-what-engineering-leaders-need-to-do-about-it/ [6] Veracode, 2025 GenAI Code Security Report, 30 July 2025 (45% vulnerable code; 80 tasks, 100+ models; Java 72%; stability GPT-4→GPT-5/Claude/Gemini); confirmed by the Spring 2026 GenAI Code Security update, 24 March 2026 (pass rate still ~55%, 150+ models evaluated, plateau held through GPT-5.1/5.2, Gemini 3 and Claude 4.5/4.6). https://www.veracode.com/resources/analyst-reports/2025-genai-code-security-report/; https://www.veracode.com/blog/spring-2026-genai-code-security/ [7] METR, “Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity”, 10 July 2025 (arXiv 2507.09089), 19% slowdown, perception gap. https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/ [8] Anthropic, “How we built our multi-agent research system”, June 2025 (orchestrator-workers, +90.2% vs single agent, ~15× tokens, ~80% of the variance). https://www.anthropic.com/engineering/multi-agent-research-system [9] Google Cloud / DORA, 2025 State of AI-assisted Software Development Report (throughput up, instability up; “AI amplifies what is already there”). https://cloud.google.com/blog/products/ai-machine-learning/announcing-the-2025-dora-report [10] OpenAI, CriticGPT, model trained to critique and find bugs in code generated by other models (2024), via LLMs-as-Judges: A Comprehensive Survey (arXiv 2412.05579). https://arxiv.org/pdf/2412.05579 [11] Self-Refine (Madaan et al., 2023) and Reflexion (Shinn et al., 2023), generate→critique→correct loop and verbal reinforcement through memorisation of failures (via arXiv 2412.05579). https://arxiv.org/pdf/2412.05579 [12] UXAgent: An LLM Agent-Based Usability Testing Framework for Web Design, CHI 2025 (arXiv 2502.12561; Amazon Science), simulated personas and automated heuristic evaluation. https://arxiv.org/abs/2502.12561 [13] Baymard Institute, “AI Heuristic UX Evaluations with a 95% Accuracy Rate” (and caveat: Microsoft UX researchers, March 2025, accuracy 50-75% depending on the case). https://baymard.com/blog/ai-heuristic-evaluations [14] Vision-language models for interface testing and agentic visual review (Percy AI Visual Review Agent, late 2025: ~3× faster, ~40% of false positives filtered out; semantic reading of the screen), TestMu (LambdaTest). https://www.testmuai.com/blog/visual-testing-ai-agent/ [15] A Survey of Vibe Coding with Large Language Models (arXiv 2510.12399, 2025), academic framing of the practice and its limits. https://arxiv.org/pdf/2510.12399

Junyr: you are not buying an AI. You are buying the base your AI works on.

Sun, 14 Jun 2026 00:00:00 GMT

A Croissance & Transitions manifesto

The idea in one sentence

Most of the “AI” software of 2026 sells you tokens: it pays for the artificial intelligence, resells it with a margin, then bills you by usage. Junyr does the opposite. We sell you your sovereign business database and the complete MCP on which your AI acts. The expensive AI is not something we pay for: it is something you bring.

This reversal is not an accounting detail. It is what makes a simple, flat, readable price possible, where all the others are condemned to count your requests and to unnerve you with variable bills.

Why everyone else's price is unstable (and ours is not)

State-of-the-art AI reasoning (the kind that runs a company, cross-references data, carries out complex actions) is expensive in compute. Any software vendor that builds this intelligence into its product absorbs the cost of the token [1][2][3]. It then has only three options: meter your usage, cap it, or pass on a bill that swells when you work more. In all three cases, you pay for your own productivity as though it were a penalty.

Junyr escapes this trap because the most expensive layer never goes through our bill. We are not token resellers. We are the sovereign, agent-ready management system onto which intelligence, yours, plugs in.

The three-layer architecture of AI cost

Layer	Who pays for the AI	What it covers
Complex / agentic reasoning	You: your Claude, ChatGPT or Gemini subscription, connected via MCP (Cowork)	Complex requests, running the company through the MCP, on your sovereign data, in keeping with your confidentiality
AI-native day-to-day	Junyr: credits included in the plan	Morning briefing, Nightly Reflections, virtual executive committee, Voice / Voice MCP, simple actions, routine agents
Total sovereignty	You: your own model (sovereign server or API key)	Gradual emancipation from the American giants, at your own pace

The reading is simple: the most expensive layer is never on our bill. The AI-native day-to-day is included in a predictable plan. And if you want to go all the way to total sovereignty, you plug in your own model.

Better still: we do not merely charge less, we make you consume less. Junyr does not settle for exposing raw data: it exposes pre-computed answers. Rather than forcing the AI to pull hundreds of rows and then aggregate everything itself (more tokens, more latency, more calculation errors), some twenty tools return the ready-to-use result in a single call: morning briefing, financial KPIs, sales pipeline, margin per deal, three-month cash-flow forecast, customer health score, anomaly detection. “What is my situation this morning?”, “my margin on deal X?”, “my cash position at quarter-end?”: one question, one call, one answer. Where an MCP that exposes only raw CRUD multiplies the requests and makes the model do the maths, Junyr does the computation server-side, that much less time and that many fewer tokens at every interaction, on your subscription.

A direct consequence: a single, flat, readable plan is not merely possible with Junyr: it is structurally superior to what AI products that pay the token on your behalf can offer.

Pillar 1: The expensive AI is what you bring

In 2026, many companies already pay for a Claude or ChatGPT subscription. That intelligence, you already have. What you lack is not one more model: it is a clean, structured, permissioned business base on which that intelligence can genuinely act.

That is exactly what Junyr gives you. You connect your AI via MCP, and it runs your company (quotes, deals, contacts, documents, finance) on data you control. The costly intellectual work runs on your subscription. The Junyr plan, for its part, stays flat.

And what if you do not yet have an external AI subscription? The offer stands up all the same: the virtual executive committee, the Voice feature and the Nightly Reflections run on the operational AI that is included. This is a floor, not a bonus. You start with a platform that is already intelligent, and you add state-of-the-art AI the day you decide to.

Pillar 2: Sovereignty and confidentiality are technical assets, not slogans

When we say “connect ChatGPT without exposing your data”, this is not a marketing promise. It is a mechanism.

Junyr's MCP already enforces, at the technical level, a per-company partitioning, a sanitisation of the data exposed, and a read refusal on accounts set to Total confidentiality. Your confidentiality tiers (Total, Secured) are not tick-boxes: they are rules the platform enforces on every request, including those of a third-party AI.

This is what turns a legitimate worry (“if I connect an American AI to my data, what leaves?”) into a clear answer: what leaves is scoped, filtered, permissioned, per company. And for those who want to take the logic all the way, the “total sovereignty” layer lets you bring your own model (sovereign server or API key) and free yourself, gradually, from the American giants.

Pillar 3: Complementary to Claude and ChatGPT, never a competitor

The mistake would be to think that Junyr fights against the large models. It is the opposite: we ride the wave, we do not fight it.

In 2026, agents (Claude Cowork, the ChatGPT agents) are searching everywhere for structured, permissioned, trustworthy business data to act on. Junyr is precisely that, with sovereignty on top. We are the data + MCP layer that these agents lack in order to become truly useful in an SME.

And it is a complete layer. An agent runs only what the MCP makes accessible: if half the company stays out of reach, the agent remains a gadget. Where most connectors expose only the CRM (or generic data that the AI has to reassemble piece by piece), Junyr's MCP natively covers the full range of business surfaces: email, CRM, deals, quotes and invoicing, purchasing and suppliers, catalogue, production, finance, expense reports, calendar, video conferencing, HR, CSR, electronic signature, documents, steering. The agent can therefore run the whole company, not just its address book.

This has two strong consequences. First, the more your company's Claude or ChatGPT workflows go through our MCP, the more Junyr becomes the infrastructure of your operation, and the more costly it would be to leave. That is what justifies an annual commitment and an unapologetic price. Second, you rise with the ecosystem instead of betting against it: every advance in Claude or ChatGPT makes your Junyr more powerful, without changing anything on your bill.

A new category, outside the usual pricing grids

Junyr is not “a CRM”. Not “a webmail”. Not one more “AI wrapper”. It is the sovereign, agent-ready management system: the data and MCP layer on which your artificial intelligence operates your business.

Because it is a new category, there is no term-for-term comparison with the tools you know. And that is deliberate: the value metric is neither the seat (the AI does the work, counting seats no longer makes sense) nor the token (that is not our expense). The value metric is the company running its business on Junyr. A platform plan per company, slightly adjusted to your usage. Flat. Simple. Readable.

In concrete terms, what is it?

Technically, Junyr is a structured data model that lets you run your company through MCP, a single central piece of software for your entire IT: email, documents, projects, quotes, electronic invoicing, catalogue, strategic steering, where you used to juggle ten siloed tools. You run your whole business through a web interface, in natural language, or by voice. No more navigating through ten screens: you ask out loud, the way you would speak to a colleague, the AI acts on your data, in keeping with your confidentiality rules.

And the mechanics of MCP are not only elegant, they are economical: by connecting your company to your consumer AI subscription (Claude or ChatGPT) rather than to the token-billed API, the operating cost per user is divided by a factor of the order of 20 [4][6] (and often far more [5]) compared with a metered API integration. The subscription is flat; the token is not. Our own case bears this out: at Croissance & Transitions, running our business on Junyr through our consumer AI subscriptions costs us roughly €4,000 over the period, whereas the same volume of reasoning billed by the token through the API would represent something of the order of €80,000 [1]: a cost divided by 20.

And for advanced functions, we go further: we configure bespoke agents for you, fully integrated into your workspace, in the service of your staff. Each person then has AI assistance that genuinely knows the company's business, because it works on its real data.

Finally, Junyr is designed for the coming decade, not to catch up on the previous one: certified, time-stamped documents and emails (provable content and date), post-quantum security by design, and an architecture ready for agent-to-agent (A2A) business. The day it is your customers' and suppliers' agents that negotiate, order and invoice, your company is already able to deal with them, under your rules.

What you get, in concrete terms

A Junyr plan per company is not “a database”. It is the whole: the management system, the email, the business modules, the MCP, the confidentiality tiers, the operational AI included (virtual executive committee, Briefing, Voice, Reflections), and the migration support by a consultant.

State-of-the-art intelligence, you bring when you want. Sovereignty, you switch on at your own pace. And the price, for its part, does not wobble.

How we begin together

The path is short and risk-free:

An AI Express Audit & Roadmap (60 minutes), online. We look at your business, your data, your usage, and we identify where sovereign AI saves you time from the very first month.
A personalised demo of Junyr, connected to your business reality.
A migration possible in 72 hours. You do not wait for months: you switch over to your sovereign base in three days, supported by a consultant.

Book your AI Express Audit & Roadmap (60 minutes): it is free, and you leave with a clear vision, whether you choose Junyr or not.

To go further

Token budgets and AI APIs: the FinOps guide for SMEs: why the token-based cost is structurally unstable, and how to govern it.
Junyr Agents™: delegating AI in your SME without losing control: how agents run your processes on your real data, under mandate and audit log.
The “All-Cloud” is dead: 5 risks that make strategic On-Premise essential: the sovereignty and confidentiality framework within which Junyr sits.

Junyr is a platform published within the Croissance & Transitions ecosystem.

Sources: verifiable, June 2026

[1] Anthropic, Claude API pricing, accessed June 2026. URL: https://platform.claude.com/docs/en/about-claude/pricing; official price list per million tokens (input/output) by model; reference base for costing usage billed by the token. The API equivalent “of the order of €80,000” is an internal calculation derived from this list (a factor of ≈ 20× the subscription), and not a figure appearing on the page.

[2] OpenAI, API Pricing, accessed June 2026. URL: https://developers.openai.com/api/docs/pricing; token-based pricing with distinct input and output prices (per million tokens); confirms that state-of-the-art intelligence is, at OpenAI too, billed by usage.

[3] Google, Gemini API, Pricing, accessed June 2026. URL: https://ai.google.dev/gemini-api/docs/pricing; pricing per million tokens (input/output, reasoning tokens included in the output price) for the Gemini language models.

[4] API易 (apiyi), Claude Max Subscription vs API Pay-As-You-Go, 2026. URL: https://help.apiyi.com/en/claude-max-vs-api-pay-per-use-pricing-comparison-claude-code-savings-guide-en.html; third-party analysis; costed case of usage of ~10 billion tokens over 8 months: ~$15,000 at the API rate against ~$800 of subscription, that is ~94% saving (≈ ÷17, “of the order of 20”). Commercial source (API reseller), to be read as an independent analysis and not as official Anthropic data.

[5] Indie Hackers (Khadin Akbar), “I used $30,983 of AI tokens last month in Claude Code on $200/mo plan”, May 2026. URL: https://www.indiehackers.com/post/i-used-30-983-of-ai-tokens-last-month-in-claude-code-on-200-mo-plan-3337a369a6; a developer on a $200/month plan documents ~17 billion tokens consumed in one month, whose value at the API rate would reach $30,983 (≈ 150× the price actually paid). Self-declared figure (linked to the promotion of the tokenflex.ing tool): to be read as an anecdotal illustration of the gap between token billing and flat plan, not as a bill issued.

[6] Product Compass (Paweł Huryn), Claude Code Pricing, April 2026. URL: https://www.productcompass.pm/p/claude-code-pricing; estimates that a coding agent run via a flat subscription comes out “about 15 to 30× cheaper” than the same work billed by the token via the API (concrete case: ~$1,588 of tokens in API equivalent covered by a $200 subscription).

Article written by Paul-Antoine TUAL, AI Transformation Leader, creator of the MATIA Method™.

Where is AI's value heading? Three signals of commoditisation that business leaders must read before the markets do

Sun, 07 Jun 2026 00:00:00 GMT

By Paul-Antoine TUAL · AI Transformation Leader, Croissance et Transitions · June 2026.

Stance. What follows was not true a week ago: the launch of Gemma 4 12B (the compact variant of the Gemma 4 family released on 2 April) on 3 June 2026 shifted the frontier of what runs on a personal machine. There is much debate about whether AI is a bubble. That is the wrong question. The right question, for a business leader and an investor alike, is: into which layer of the value chain is value migrating? Three signals from 2026 sketch out a coherent answer: open models catching up, compact local models being good enough, and the software layer being cloned overnight. It does not condemn AI; it condemns certain valuations. And it points precisely to where an SME should invest.

1. First signal: an open model overtakes a proprietary flagship

On 7 April 2026, Z.ai released GLM-5.1, an open-weight model under the MIT licence (a 754 billion-parameter MoE, 40 billion active per token, 202,752-token context, autonomous execution for up to 8 hours) [1]. On release, it took the top of the SWE-Bench Pro ranking with 58.4%, ahead of GPT-5.4 (57.7%), Claude Opus 4.6 (54.2-57.3% depending on the harness) and Gemini 3.1 Pro (54.2%) [1][2]. On Terminal-Bench 2.0 it also beats Gemini 3.1 Pro (63.5% vs 56.9%), and it is the first open-weight model in history to reach the top 3 of Code Arena [2]. On coding and agentic work, precisely the two use cases driving enterprise AI transformation, a model that can be downloaded for free thus pulled level, if only for a few weeks, with the flagship models billed by usage.

The nuance that avoids sensationalism: the frontier struck back within weeks. Claude Opus 4.8 (Anthropic, 28 May 2026) retook a clear lead on SWE-bench Pro at 69.2%, Claude Opus 4.6 keeps the lead on Terminal-Bench 2.0 (68.5%), and proprietary models still dominate pure abstract reasoning (GPQA Diamond: 94.3% vs 86.2%) [2]. Epoch AI puts the average lag of the best open-weight models at 3-4 months behind frontier models [3]. The frontier has not disappeared; it has shrunk to a few points, on a few benchmarks, for a few months' lead. A gap imperceptible for 95% of enterprise use cases.

But the economics, for their part, have already tipped. The price of a given level of performance is collapsing: Epoch AI measures a 40-fold drop in price per year to reach GPT-4 level on doctorate-level scientific questions [3]; the Gemini 3.1 Flash API costs $0.10/M input tokens where GPT-4 cost $30 in 2023, a ~99.7% drop in three years [4]; Gartner forecasts that inference on a 1 trillion-parameter model will cost 90% less in 2030 than in 2025 [5].

Interim conclusion: when the underlying asset (the model) catches up with proprietary quality and trends towards the marginal cost of compute, the rent from the model alone evaporates. What still sells at a premium is the few months' lead, an asset that depreciates at the pace of releases.

2. Second signal: the compact local model is enough for most agentic tasks

This signal is very recent: it dates from the launch of Gemma 4 12B on 3 June 2026 (the compact variant of the family released on 2 April), four days before these lines were written. Thanks to its unified, encoder-free architecture, Gemma 4 12B runs entirely locally on a machine with 16 GB of RAM, a standard professional laptop, with a 256,000-token context and native audio/vision [6][7]. Its agentic performance is serious: 69.0% on tau2-bench (agent simulation in a real enterprise environment), against 76.9% for its bigger 31B sibling. And the 12B beats the older Gemma 3 27B, twice as heavy, on document vision and reasoning [7]. Google explicitly documents “local agentic workflows” on a laptop [8].

The strategic reading matters as much as the technical feat. By releasing for free a local model that covers most agentic use cases, Google applies the old principle “commoditise your complement” formulated by its own chief economist, Hal Varian: it deliberately cannibalises the model layer, the one from which labs with no other revenue source draw their rent, because its own value lies elsewhere: distribution, cloud, hardware, advertising. When the best-resourced player in the sector decides that the model is a loss-leader, commoditisation is no longer a market drift; it is a deliberate strategy.

How many enterprise tasks does this cover? NVIDIA Research's position paper “Small Language Models are the Future of Agentic AI” estimates that 80 to 90% of agentic invocations fall into the “a small model is enough” category (tool calls, structured reasoning, orchestrated steps) for an inference cost 10 to 30 times lower [9].

The 2026 hardware scale (orders of magnitude observed):

Need	Machine	Budget	What runs
Everyday local agentic work	Laptop / 16 GB mini-PC	< €1,000	Gemma 4 12B and quantised equivalents
Advanced agentic + multimodal	Mac Studio 64 GB or AMD/Qualcomm mini-PC with 128 GB of unified memory	~€2,000-3,500	Gemma 4 31B and beyond
Frontier-class model run locally	Mac Studio 512 GB, option pulled from Apple's catalogue in March 2026 (DRAM shortage); speculative secondary market or a cluster of 256 GB machines	> €12,000 (estimate)	GLM-5.1 quantised to 8-bit

Two readings of this table. First, the new generation of unified-memory mini-PCs (AMD Ryzen AI Max+ 128 GB between ~$2,400 and $3,200, against $4,699 for the equivalent NVIDIA DGX Spark station) drives down the cost of the local token: at market prices, the ratio of tokens generated per dollar invested tilts clearly in favour of these commoditised machines [26]. Second, the disappearance of the Mac Studio 512 GB is an absurd illustration of this article's thesis: chasing frontier infrastructure locally has become a speculative game, whereas the €2,500 machine that covers 80-90% of use cases remains on the shelf.

And the curve keeps falling thanks to quantisation innovations: TurboQuant (Google Research, presented at ICLR 2026) combines random vector rotation, aggressive quantisation and 1-bit residual correction to divide the memory footprint of the weights by 3.2. In a 4-bit + 8-bit residual configuration, the perplexity loss is strictly zero compared with the 16-bit model [10][11]. Models that required a data-centre GPU cluster now run on consumer or semi-professional hardware [11]. Quantisation does not nibble away at the commoditisation of inference: it accelerates it structurally, shifting the data-centre/local frontier by one notch every six months.

The countercurrent not to be hidden: the cost of memory. It would be dishonest to present the shift to local as plain sailing. The crisis is not a factory accident, it is a strategic choice: to serve data centres' demand for AI chips, Samsung, SK Hynix and Micron have converted the majority of their lines to HBM (high-margin), sacrificing the classic DRAM in our PCs, servers and Macs. The result measured by TrendForce: DRAM contracts +90-95% in Q1 2026, then +58-63% in Q2 (NAND: +70-75%), with hyperscalers locking up supply through long-term contracts [23]. Gartner expects a cumulative increase of up to 130% over the year [24], and no easing is expected before the new plants ramp up in volume: late 2027 at the earliest, with a return to 2024 price standards more likely around 2028 [23][25]. It is this shortage that pushed Apple to pull the Mac Studio 512 GB from its catalogue rather than display absurd prices [26].

The practical consequence for a business leader fits in two lines. Standard needs (16-64 GB): buy despite the increase. An extra cost of around €100-150 remains marginal against the productivity gain from compact models, and this increase hits new hardware, not cloud API prices, which keep collapsing. Massive needs (128 GB and above): wait or rent. The segment is in speculative overheating; it is better to consume tokens via API over the next 12-18 months than to overpay for physical infrastructure bound to depreciate sharply when supply eases around 2028. The local/cloud trade-off is therefore made use case by use case. And quantisation partly offsets the increase by dividing the memory requirement at near-equal quality. One last point of caution: self-hosting assumes engineering skills that many SMEs do not yet have in-house. This is precisely where the value of support lies (section 5).

What still justifies heavy infrastructure: long-running frontier reasoning, real-time at large scale (voice, video, trading), very long contexts, training. This is real, but these are cutting-edge use cases, not the daily reality of 99% of SMEs and mid-caps. For most companies, the relevant AI infrastructure fits within a workstation budget, not a data-centre contract.

3. Third signal: the software layer was cloned overnight

On 31 March 2026, a .map file published by mistake in the Claude Code npm package exposed the complete internal architecture of Anthropic's coding agent (512,000 lines, 1,906 files) [12]. Within hours, the open-source community produced a clean reimplementation of it (Claw Code, Python/Rust, without copying a single line of proprietary code), which reached 100,000 GitHub stars in 24 hours, an all-time platform record [12][13].

The lesson goes beyond the anecdote: the software layer surrounding the model (the “harness”: agentic loop, tools, permissions) is architecturally transparent. Protecting it through secrecy or intellectual property is very difficult; replicating it costs a motivated community one night's work. What remains defensible in this layer: installed distribution, brand, trust, release cadence, deep integrations. In other words, commercial assets, not code.

4. Consequence: are the labs becoming mere compute landlords?

If the model becomes commoditised (signal 1), if inference moves down to the workstation (signal 2) and if the software is cloned (signal 3), what is left for the labs? The temptation is to answer: renting out compute power, a telecoms business, with returns compressed by the price war.

The 2026 figures already sketch out this telecoms economy. Anthropic passed $30 billion in annualised revenue in April 2026 and overtook OpenAI (~$25 billion). But the consumer subscription is plateauing, with a non-GAAP operating margin of -122% at OpenAI in Q1 2026 [14]: growth comes from the enterprise API, that is, from selling tokens by volume, literally renting out cognitive compute. And the reality is more uncomfortable still: most labs do not even own the compute they would be renting out. They themselves rent it from the hyperscalers and chipmakers, often through circular arrangements (the Nvidia-OpenAI deal reportedly represents ~13% of Nvidia's projected 2026 revenue [15]) that recall the vendor financing of late-1990s telecoms. The hyperscalers' 2026 capex (~$725 billion, including $180-190 billion for Alphabet alone, in the region of 2.5% of US GDP) [15], the applications-revenue “gap” (more than $500 billion needed to justify ~$1 trillion of infrastructure, according to Sequoia and Goldman) [15], and the NBER study of February 2026 (across ~6,000 executives: 90% measure no impact of AI on their company's productivity over three years, for an average use of 1.5 h/week) [16] say the same thing: the current valuations of the model and application layers assume a rent that the three signals above are dissolving.

The labs' counter-arguments exist and deserve to be taken seriously: vertical integration into products (the lead is sold as a product, not as an API), usage data, consumer brand, enterprise relationships. But each of these assets belongs to distribution and service, not to the model.

And there is an even more explicit admission: the labs and their allies are themselves launching consulting practices. The IBM-Google Cloud alliance of 4 June 2026 mobilises thousands of consultants around Gemini Enterprise [18]; the model vendors are developing their own client-facing integration and deployment arms. A model vendor that hires consultants, a service-margin business where the API promised a software margin, acknowledges three things: that the model alone no longer sells for enough, that the value is in the last mile it did not control, and that the migration thesis is correct. When the vendor moves downstream along the chain, it is because the rent from its original layer is eroding.

5. Where value takes refuge

The grid is now readable, layer by layer:

Layer	Value trend	Why
Semiconductors, energy, data-centre real estate	Sustained but cyclical	Real physical scarcity; risk of overcapacity
Models (labs)	Compressing	Open-weight + collapse in inference prices
Software layer / agents	Compressing fast	Cloneable in hours; no lasting protection
Living proprietary data	Rising	Continuous operational cost, not replicable by code
Approvals, compliance, distribution	Rising	Administrative and contractual barriers
Workflow integration + human support	Rising strongly	The service IS the product; Infosys puts the AI services opportunity at $300-400 billion by 2030 [17]; IBM and Google Cloud form an alliance on 4 June 2026 precisely around “human expertise at scale” [18]; in France, AI already accounts for more than 11% of Capgemini's bookings in Q1 2026 [19]

Human support is therefore not “what is left” by default: it is the layer towards which all rational players are converging, labs included. For an SME, the consequence is direct: sustainable AI spending is neither the model licence nor the tool of the moment. It is the method, governed data and internal skills. Exactly the thesis of the MATIA Method™: structuring the organisation to absorb models and tools that have become interchangeable, rather than marrying a single vendor.

Three questions to ask first thing on Monday. 1. What share of our AI use really requires a frontier model billed by usage, and what share would run on a €2,500 machine? 2. If our main AI tool were to disappear or increase tenfold in price tomorrow, what would we lose: code (replaceable) or data and skills (our own)? 3. Does our AI budget fund the rents of vendors on the way to commoditisation, or internal assets that appreciate?

6. What this says about valuations, without doom-mongering

Should we conclude that a “bursting” is imminent? No: recent multi-method work [20] and layer-by-layer analysis [21] converge on a more useful diagnosis. There is not one bubble; there are overvalued layers and sustainable layers, with different correction horizons. The concentration of the indices and a Shiller P/E of 42.3, very close to the record of 44.2 in 2000, make the correction of the fragile layers painful for everyone: on 3 June 2026, Broadcom lost 15% and $280 billion in market cap in a single session over a forecast barely below expectations [22]. This does not change the direction of the migration. The business leader does not have to predict the date; they have to position themselves on the right side of the migration. It is a schedule of decisions, not an alarm.

In practice

Croissance et Transitions supports the leaders of SMEs and mid-caps in positioning themselves on the right side of this migration: maturity diagnostic (MATIA Method™), local/cloud trade-offs use case by use case, governance of proprietary data and upskilling of teams. These are the assets that appreciate as models and tools become commoditised.

→ AI Express Audit & Roadmap: 60 minutes over video call. croissance-transitions.fr

Paul-Antoine TUAL · AI Transformation Leader · Croissance et Transitions (SAS) · MATIA Method™

FAQ

Is an open-source model really on a par with proprietary models in 2026? On coding and agentic work, yes for some benchmarks: GLM-5.1 (MIT licence) beats Gemini 3.1 Pro on SWE-Bench Pro (58.4% vs 54.2%) and on Terminal-Bench 2.0 (63.5% vs 56.9%). Proprietary models keep a lead on other composites, a lead measured in points and months, no longer in orders of magnitude.

Does an SME need heavy cloud infrastructure for agentic AI? In 80 to 90% of agentic invocations (NVIDIA Research estimate), a compact model is enough, and runs locally from 16 GB of RAM (Gemma 4 12B), or on a Mac Studio 64 GB (~€2,500) for the 31B version.

What does quantisation change for an SME? Techniques such as TurboQuant (Google Research, ICLR 2026) divide the memory footprint of models by ~3.2 with zero to marginal quality loss: models once reserved for clusters run on semi-professional hardware. The data-centre/local frontier keeps coming down.

Can the software layer of AI agents be protected? Weakly: the architecture of Claude Code, exposed by mistake in March 2026, was reimplemented in open-source overnight (Claw Code, 100,000 GitHub stars in 24 h). Durable defence is commercial (distribution, trust, integrations), not technical.

Where should an AI budget be invested in 2026? In the layers that appreciate: governed proprietary data, compliance, workflow integration and internal skills, not in the rents of models or tools on the way to commoditisation.

Sources

[1] Z.ai, GLM-5.1: Towards Long-Horizon Tasks (7 April 2026). https://z.ai/blog/glm-5.1 · Documentation: https://docs.z.ai/guides/llm/glm-5.1

[2] DeepInfra, GLM-5.1 Model Overview (SWE-Bench Pro / Terminal-Bench 2.0 / Code Arena). https://deepinfra.com/blog/glm-5-1-model-overview · Artificial Analysis, GLM-5.1 vs Gemini 3.1 Pro. https://artificialanalysis.ai

[3] Epoch AI, Open-weight models lag state-of-the-art by around 3 months on average. https://epoch.ai/data-insights/open-closed-eci-gap · LLM inference price trends. https://epoch.ai/data-insights/llm-inference-price-trends

[4] AI Magicx, The LLM Pricing Collapse of 2026. https://www.aimagicx.com/blog/llm-pricing-collapse-developer-guide-building-cheap-ai-2026

[5] Gartner, By 2030, Performing Inference on an LLM With 1 Trillion Parameters Will Cost Over 90% Less Than in 2025 (25 March 2026). https://www.gartner.com/en/newsroom/press-releases/2026-03-25-gartner-predicts-that-by-2030-performing-inference-on-an-llm-with-1-trillion-parameters-will-cost-genai-providers-over-90-percent-less-than-in-2025

[6] Google, Introducing Gemma 4 12B: a unified, encoder-free multimodal model (3 June 2026). https://blog.google/innovation-and-ai/technology/developers-tools/introducing-gemma-4-12b/

[7] Sierra Research, τ²-bench. https://taubench.com/ · VentureBeat, Gemma 4 12B runs entirely locally on a typical 16GB enterprise laptop. https://venturebeat.com

[8] Google Developers Blog, Bringing Gemma 4 12B to your Laptop: Unlocking Local, Agentic Workflows. https://developers.googleblog.com/bringing-gemma-4-12b-to-your-laptop-unlocking-local-agentic-workflows-with-google-ai-edge/

[9] NVIDIA Research, Small Language Models are the Future of Agentic AI (arXiv:2506.02153). https://arxiv.org/abs/2506.02153

[10] Google Research, TurboQuant: Redefining AI efficiency with extreme compression (ICLR 2026). https://research.google/blog/turboquant-redefining-ai-efficiency-with-extreme-compression/ · https://arxiv.org/abs/2504.19874

[11] AI Indigo, TurboQuant Explained: How This New Compression Method Changes Local LLM Inference. https://aiindigo.com/blog/turboquant-explained-how-this-new-compression-method-changes-local-llm-inference

[12] Zscaler ThreatLabz, Anthropic Claude Code Leak. https://www.zscaler.com/blogs/security-research/anthropic-claude-code-leak

[13] GitHub, claw-code (ultraworkers). https://github.com/ultraworkers/claw-code · Cybernews, Leaked Claude Code source spawns GitHub's fastest repo. https://cybernews.com

[14] SaaStr, Anthropic Just Passed OpenAI in Revenue. https://www.saastr.com · Medium, Anthropic just passed OpenAI in revenue. Here is why it matters. https://medium.com/@david.j.sea/anthropic-just-passed-openai-in-revenue-here-is-why-it-matters-e3dd9bb04069

[15] Allianz Research, AI capex cycle: war-proof for now (25 March 2026). https://www.allianz.com · Goldman Sachs, AI: In a Bubble? https://www.goldmansachs.com · IDC, Circular financing has muddied the AI story. https://www.idc.com/resource-center/blog/circular-financing-has-muddied-the-ai-story-watch-the-application-layer-instead/

[16] NBER, Firm Data on AI (Working Paper w34836, February 2026). https://www.nber.org/papers/w34836 · NBER w34851, Does Generative AI Narrow Education-Based Productivity Gaps? https://www.nber.org

[17] Infosys, AI First Value Framework: AI Services Opportunity of Over $300 Billion. https://www.infosys.com/newsroom/press-releases/2026/unveils-ai-first-value-framework.html

[18] IBM, IBM and Google Cloud Announce Strategic Partnership to Scale AI with Human Expertise (4 June 2026). https://newsroom.ibm.com/2026-06-04-ibm-and-google-cloud-announce-strategic-partnership-to-scale-ai-with-human-expertise-and-ai-powered-delivery

[19] Capgemini Investors, Q1 2026 revenues. https://investors.capgemini.com

[20] arXiv, Boom, Bubble, or Buildout? A Multi-Method Evaluation (2606.01575). https://arxiv.org/html/2606.01575

[21] VentureBeat, Stop calling it 'The AI bubble': It's actually multiple bubbles. https://venturebeat.com/infrastructure/stop-calling-it-the-ai-bubble-its-actually-multiple-bubbles-each-with-a

[22] TradingKey, S&P 500 valuation, Shiller P/E 42.32, Broadcom -15%. https://www.tradingkey.com/analysis/stocks/us-stocks/261950917-sp500-valuation-bubble-ai-concentration-shiller-pe-buffett-indicator-fed-hawkish-yield-market-nifty-fifty-strategy-tradingkey

[23] TrendForce, AI Server Demand to Drive Memory Contract Price Increases in 2Q26 (31 March 2026). https://www.trendforce.com/presscenter/news/20260331-12995.html · Tom's Hardware, DRAM prices predicted to jump 63% in Q2, NAND up to 75%. https://www.tomshardware.com/pc-components/dram/dram-and-nand-contract-prices-to-climb-again-in-q2

[24] TechTimes, RAM Prices 2026: Gartner Forecasts 130% Memory Cost Surge (5 June 2026). https://www.techtimes.com/articles/317872/20260605/ram-prices-2026-buy-now-wait-gartner-forecasts-130-memory-cost-surge.htm

[25] SaaS Sentinel, RAM Shortage Could Last Until 2028 as AI Demand Reshapes Memory Markets (April 2026). https://saassentinel.com/2026/04/19/ram-shortage-could-last-until-2028-as-ai-demand-reshapes-memory-markets/

[26] Tom's Hardware, Apple pulls 512GB Mac Studio upgrade option. https://www.tomshardware.com/tech-industry/apple-pulls-512-mac-studio-upgrade-option · MacRumors (5 March 2026). https://www.macrumors.com/2026/03/05/mac-studio-no-512gb-ram-upgrade/ · TechSpot, AMD Ryzen AI Halo mini PC, 128GB, vs DGX Spark. https://www.techspot.com/news/112287-amd-ryzen-ai-halo-mini-pc-coming-june.html · Liliputing, Ryzen AI Max+ mini PCs with 128GB. https://liliputing.com/more-ryzen-ai-max-395-mini-pcs-with-128gb-are-now-available-if-you-can-afford-one/

Video prospecting in 2026: the standard for a first meeting that converts

Tue, 02 Jun 2026 00:00:00 GMT

By Paul-Antoine TUAL, AI Transformation Leader, Croissance et Transitions. May 2026.

The rarest meeting has become the most decisive

Here is the paradox that every sales leader must confront in 2026. Your buyers have never needed you so little to inform themselves, and never needed you so much to decide.

The figures are unequivocal. According to Gartner, 67% of B2B buyers say they prefer a rep-free buying experience for at least part of their journey, and 70% prefer a fully digital, self-service journey [5]. McKinsey was already observing in its B2B Pulse surveys (2021-2024) that two thirds of buyers preferred remote or digital interactions to in-person ones at many stages, and that e-commerce had become the leading revenue-generating channel in B2B, ahead of in-person selling [1, 2]. The modern buyer does the work alone: comparing, downloading, querying generative AI tools. Gartner estimates that buyers devote only a fraction of their buying time to meeting suppliers [4], and predicted as early as 2020 that 80% of sales interactions would take place through digital channels by 2025 [3].

One might conclude that the sales rep is fading away. That would be a misreading. A second Gartner survey (May 2026) reveals the other side of the coin: 69% of B2B buyers turn to a sales rep to validate the analyses produced by AI [6]. Better still: by 2030, Gartner anticipates that 75% of buyers will prefer sales experiences that prioritise human interaction over AI [7]. In other words, the more information becomes digital and automated, the rarer the moment of human contact becomes, and the more decisive it becomes. The 2026 buyer arrives at the meeting already informed, sometimes better than your junior sales rep, and what they come looking for is no longer data: it is trust, judgement and validation.

In 2026, this moment of contact is most often a video call. Not a trip, not a lunch: a window of thirty to forty-five minutes, screen shared, camera on. That is where everything is decided. And that is precisely where most organisations remain amateurish, because they have transposed the reflexes of in-person meetings onto video without rewriting the rules. This article proposes a standard of execution, a golden standard, for this first remote meeting: how to prepare it, how to conduct it, how to show, how to put a figure on it live, and how to turn it into a next step through the meeting summary and follow-up. Without gimmicks, and without giving in to technological panic.

Before the meeting: the standard begins with the frame

A first video meeting is not won during the call. It is won beforehand, in the quality of the frame you set. Three elements distinguish a professional exchange from an improvised call.

The first is factual preparation. Since the buyer arrives informed, the sales rep must be more so. This does not mean reciting the company profile: it means arriving with a substantiated value hypothesis, two or three questions that demonstrate an understanding of the sector, and a point of view. The buyer who spends the bulk of their time informing themselves alone no longer tolerates the sales rep who “discovers” their business during the session.

The second is the agenda shared in advance. Sending, the day before, a three-point agenda and a clear objective for the session is not a formality: it is what turns a “presentation” passively endured into a jointly built conversation. The 2026 standard goes further with what analysts call the Digital Sales Room, a shared digital space between buyer and seller that centralises the agenda, the materials, the demo, the business case and, later, the mutual action plan. The digital sales room replaces the scattered email thread and the lost attachments: it gives the buyer a single, persistent place that they can reopen and share with their committee.

The third, the most neglected, is audiovisual credibility. A poorly lit frame, a microphone that clips, a downward camera angle from a laptop placed too low: these details are not cosmetic. In front of a decision-maker, they signal a lack of command. Conversely, a sharp image, clear sound and eye-level framing establish a professional presence within seconds. In person, your suit and your handshake spoke for you; on video, it is your image and your sound. The standard has become a prerequisite, not a luxury.

During the meeting: listen more than you speak

Then comes the moment of the meeting itself. And here, the most useful data point in all the sales literature comes down to two figures.

The software firm Gong analysed more than 326,000 B2B sales calls (analyses published in the late 2010s, still the profession's reference). The talk-to-listen ratio of the highest-performing sales reps sits around 43% talking to 57% listening [8]. The best talk less than average, and above all they keep this ratio stable whether they win or lose the deal. The lower performers, by contrast, see their talking time climb as soon as the deal gets complicated, as if talking more could convince. It is the opposite that works. Gong also observes that the sellers who win the deal ask about fifteen to sixteen questions over the call, neither too few, nor the interrogation barrage of those who ask twenty and lose [9]. Discovery is not a questionnaire: it is a guided conversation.

This discipline of listening takes on particular importance on video, because the screen adds a difficulty that in-person meetings did not have: engagement must be actively created. You do not capture a decision-maker's attention behind a screen the way you capture it in a meeting room. A few principles make the difference. Ask a question in the first two minutes, before any slide, to turn the expected monologue into a dialogue. Name what you are going to do and how long it will take, to reduce the mental load. Watch the weak signals (a gaze that drifts away, a camera that gets switched off) and respond with an open question rather than by speeding up the pitch.

The proven qualification frameworks (MEDDIC, SPIN Selling, Challenger Sale) remain perfectly valid; they simply need to be adapted to the format. MEDDIC, for example, requires identifying the decision criteria and the internal champion: on video, where you do not run into the other stakeholders in the corridor, this mapping must be explicitly prompted through questions. SPIN structures discovery around situation, problem, implication and need: this is precisely the backbone of a good first remote meeting, because it forces you to listen before proposing. The format changes; the method remains.

Showing: the live demonstration, not the slideshow

Once discovery is done, comes the moment to show. And that is where screen sharing becomes a weapon, or a trap.

The trap is well known: rattling through forty feature slides. The principle that works is the opposite, and it has a name: show, don't tell. What makes an impression on a buyer is not the list of features, it is the story behind each one: the problem it solves, the transformation it enables. A live demonstration, anchored in the buyer's concrete case, is worth ten generic slideshows. The data confirms it: according to figures published by the interactive-demo players (Navattic, Demostack; vendor data, to be read as orders of magnitude), 35% of B2B buyers interact with an interactive demonstration during their purchase, and the conversion rate of prospects who have handled a demo is markedly higher than that of others, of the order of +63% on qualification [10, 11]. Showing by having the buyer handle the product, rather than showing by talking, changes the nature of the engagement.

On video, this imposes a technical requirement that is simple but real: smooth screen sharing, a rehearsed demonstration, and the ability to go back if the buyer wants to dig into a point. The worst-case scenario remains the sales rep who hunts for their tab, waits for a load, or discovers a bug live. Command of the demonstration tool is part of the standard just as much as command of the pitch.

Putting a figure on it live: co-building the business case in the session

This is the step that separates the amateur from the professional in 2026: the ability to put a figure on the value, in the session, with the buyer.

The 2026 buyer does not buy a solution, they buy a return on investment that they will have to defend before their committee. Presenting them with a price without a value framework is leaving them alone in the face of the budget objection. The standard consists of co-building the calculation in front of them: opening a ROI calculator or a configurator, entering their own figures (volume, current cost, time spent) and letting the result appear live. Value stops being a seller's promise and becomes a calculation that the buyer has watched take shape, with their own data. It is also an excellent test: if the buyer challenges an assumption, you adjust it together, and the conversation about value takes place during the meeting rather than afterwards, by email, when the sales rep is no longer there to carry it.

This shared quantification prepares the next step, which strongly determines the outcome: the mutual action plan. It is a shared roadmap listing the next steps, the owners and the deadlines, on the seller's side and the buyer's side. Outreach's data shows that deals equipped with a mutual action plan post a win rate that is 26% higher, even though only a minority of sales reps use one systematically [13]. Concluding a first video meeting without having set, on screen, the date and purpose of the next exchange is letting the deal go cold.

After the meeting: the summary and the speed of follow-up

The meeting is over. The work, however, has only just begun, and this is often where deals are lost.

The founding data point is old but has never been disproven. A study published in the Harvard Business Review (Oldroyd, McElheran, Elkington), covering more than 2,200 companies and 100,000 prospects, establishes that responding to a prospect in under five minutes makes reaching them a hundred times more likely, and qualifying them twenty-one times more likely, compared with a response after thirty minutes [12]. The speed of follow-up is not an incidental quality: it is a multiplier. After a first meeting, the summary sent within the hour (recapping what was said, the quantified value and the next steps) carries more weight than the same message sent two days later.

This is precisely where AI applied to sales brings concrete value, and not a gimmick. Modern meeting assistants transcribe the exchange, produce a structured summary of it, extract the action items and sync them with the CRM. The sales rep who, yesterday, spent their evening writing up their summaries can now review, correct and send in a few minutes, and devote the time saved to the relationship. Gartner anticipated that 75% of B2B sales organisations would augment their practices with AI-guided selling solutions, and already finds that those which provide their sales reps with AI-assisted “next best actions” are 2.6 times more likely to achieve commercial growth [14, 15]. AI does not replace the follow-up: it makes it fast and systematic.

Tooling and sovereignty: method before tool

One watchword runs through everything above: technology is in the service of discipline, never the other way around. The 2026 ecosystem offers categories of tools that are now mature: video conferencing, interactive sharing and demonstration, AI recording and transcription, conversational intelligence, digital sales rooms. Stacking these tools without method produces nothing. It is the mistake I see most often in my advisory work: organisations over-equipped and under-methodical.

And since we are talking about recording and transcribing conversations, one subject can no longer be relegated to a footnote: compliance and data sovereignty. Having a transcription bot join a meeting without having warned the participants is not a convenience, it is a legal risk. The CNIL is clear: the recording and transcription of a professional meeting rest on explicit, informed and freely given consent [17]. French law penalises the recording of a private or confidential conversation without consent [18]. And since February 2025, the European AI Act (Article 4, softened by the Digital Omnibus) has required organisations to support their teams' AI literacy, an obligation of means, not of result: deploying a meeting assistant without training teams in how it works runs against that requirement [19]. This framework is not a hindrance: it is a standard of seriousness. Choosing solutions hosted in Europe, announcing consent at the start of the session and controlling where transcriptions are stored are today part of a sales rep's credibility in front of a decision-maker attentive to their data.

Forrester, moreover, puts a figure on the cost of recklessness: its 2026 B2B predictions estimate that companies will lose more than $10 billion in value because of ungoverned use of generative AI, and observe that a share of buyers using AI tools feel less confident in their decision because of inaccurate information [16]. The lesson, for a business leader, is at once calm and firm: AI powerfully augments remote prospecting, provided it is governed: data controlled, teams trained, results validated by humans.

What if video impoverished the relationship? This is the most frequent objection, and it is not without foundation: a screen filters out body language, tires attention, and makes disengagement easier. But the data does not argue for a step backwards: it argues for a change of standard. The buyer wants digital for informing themselves (70% prefer self-service) and the human for deciding (69% validate their analyses with a sales rep) [5, 6]. Video conducted well does not impoverish the relationship: it concentrates it on the moment when it matters. The risk is not video; it is amateurish video.

Five questions to ask yourself before your next video meeting 1. Have my agenda and my objective been shared before the session? 2. Am I going to speak for less than half the time, and have I prepared my fifteen good questions? 3. Is my demonstration anchored in the buyer's concrete case, or is it a generic slideshow? 4. Am I able to quantify the value live, with their own figures? 5. Have I planned for consent to recording, and a summary sent within the hour?

The standard is discipline, not equipment

The 2026 golden standard for video prospecting does not lie in a purchase of licences. It lies in a discipline of execution: a prepared frame, listening that exceeds speaking, a demonstration that shows rather than tells, value quantified in the session, a mutual action plan, and a fast follow-up tooled by AI but governed by humans. The technology of 2026 makes each of these gestures easier; it exempts you from none of them.

This is also the logic of the MATIA Method™: advancing an organisation not by accumulating tools, but by raising its level of maturity in the use it makes of them, from scattered craft to mastered orchestration. Note-taking, summarising and CRM synchronisation are precisely the type of task that governed AI agents can take on, provided the data remains controlled and the result validated by the sales rep. This is the purpose of Junyr Agents™: delegating repetitive execution to AI to give teams back the time for the relationship, the only thing the 2026 buyer still comes looking for in front of a human.

In practice

Croissance et Transitions supports the leaders of SMEs and mid-caps in making the first video meeting a standard of execution that converts: preparation, conduct, demonstration and a quantified business case in the session, then follow-up governed by AI, with the MATIA Method™ applied to sales performance.

→ AI Express Audit & Roadmap: 60 minutes by video. croissance-transitions.fr

Paul-Antoine TUAL · AI Transformation Leader · Croissance et Transitions (SAS) · MATIA Method™

FAQ

What is the ideal length for a first video prospecting meeting? Thirty to forty-five minutes is generally enough for a first exchange. Beyond that, attention drifts behind a screen. The essential thing is to devote the first half to discovery (listening, questioning) before any demonstration, and to reserve the last few minutes for the mutual action plan.

What is the right talk-to-listen ratio in a sales meeting? Gong's analysis of more than 326,000 calls puts the ratio of the best sales reps at around 43% talking to 57% listening. Talking less, asking around fifteen good questions, and keeping this ratio stable even when the deal gets complicated is what sets the high performers apart from the rest [8, 9].

Should you record your video meetings, and how do you stay GDPR-compliant? Recording and AI transcription are permitted, but rest on the explicit, informed and freely given consent of the participants, announced at the start of the session (CNIL recommendations) [17]. Favour solutions hosted in Europe, control where transcriptions are stored and train your teams: the AI Act (Article 4) has required organisations to support their AI literacy since February 2025 [19].

Are AI meeting-summary tools worth the investment? Yes, provided they are governed. They remove hours of writing and enable follow-up within the hour, and responding in under five minutes makes reaching a prospect a hundred times more likely [12]. The gain is not the summary itself, it is the time given back to the relationship and the speed of follow-up.

Is the sales rep disappearing in favour of self-service? No. While 67% of buyers prefer to go rep-free for part of their journey, 69% turn to a sales rep to validate the analyses produced by AI, and Gartner anticipates that by 2030, 75% will prefer experiences that prioritise the human [5, 6, 7]. The sales rep's role is shifting from information towards judgement and trust.

How do you increase the conversion rate after a first meeting? Three cumulative levers: co-building the quantified value in the session, formalising a mutual action plan (+26% win rate according to Outreach [13]), and sending an actionable summary within the hour. The speed and clarity of the follow-up weigh as much as the quality of the meeting itself.

Sources

[1] McKinsey & Company, The future of B2B sales is hybrid. https://www.mckinsey.com/capabilities/growth-marketing-and-sales/our-insights/the-future-of-b2b-sales-is-hybrid

[2] McKinsey & Company, Five fundamental truths: How B2B winners keep growing (B2B Pulse 2024). https://www.mckinsey.com/capabilities/growth-marketing-and-sales/our-insights/five-fundamental-truths-how-b2b-winners-keep-growing

[3] Gartner, Gartner Says 80% of B2B Sales Interactions Between Suppliers and Buyers Will Occur in Digital Channels by 2025 (research on the B2B buying journey). https://www.gartner.com/en/newsroom/press-releases/2020-09-15-gartner-says-80–of-b2b-sales-interactions-between-su

[4] Gartner, The B2B Buying Journey: Key Stages and How to Optimize Them. https://www.gartner.com/en/sales/insights/b2b-buying-journey

[5] Gartner, Gartner Sales Survey Finds 67% of B2B Buyers Prefer a Rep-Free Experience (9 March 2026). https://www.gartner.com/en/newsroom/press-releases/2026-03-09-gartner-sales-survey-finds-67-percent-of-b2b-buyers-prefer-a-rep-free-experience

[6] Gartner, Gartner Survey Finds 69% of B2B Buyers Turn to Sales Reps to Validate AI-Generated Insights (20 May 2026). https://www.gartner.com/en/newsroom/press-releases/2026-05-20-gartner-survey-finds-sixty-nine-percent-of-b-two-b-buyers-turn-to-sales-reps-to-validate-ai-generated-insights

[7] Gartner, Gartner Says by 2030 that 75% of B2B Buyers Will Prefer Sales Experiences that Prioritize Human Interaction Over AI (25 August 2025). https://www.gartner.com/en/newsroom/press-releases/2025-08-25-gartner-says-by-2030-that-75-percent-of-b2b-buyers-will-prefer-sales-experiences-that-prioritize-human-interaction-over-ai

[8] Gong, Mastering the talk-to-listen ratio in sales calls (analysis of 326,000 B2B calls). https://www.gong.io/blog/talk-to-listen-conversion-ratio

[9] Gong, Mastering Discovery Calls: Essential Questions and Tips. https://www.gong.io/blog/best-discovery-call-tips

[10] Gartner Peer Insights, Interactive Demonstration Applications Reviews ; Navattic, Interactive Demo Best Practices 2026. https://www.gartner.com/reviews/market/interactive-demonstration-applications | https://www.navattic.com/blog/interactive-demos

[11] Demostack, Sales Demo: 8 Tips to Improve Your Demo Conversion Rates (vendor source, to be read as an order of magnitude). https://www.demostack.com/post/sales-demo-roi

[12] Harvard Business Review, Oldroyd, McElheran & Elkington, The Short Life of Online Sales Leads (the 5-minute rule). https://hbr.org/2011/03/the-short-life-of-online-sales-leads

[13] Outreach, How to improve win rates by 26% with a best-in-class mutual action plan. https://www.outreach.io/resources/blog/how-to-use-mutual-action-plans

[14] Gartner, Gartner Predicts 75% of B2B Sales Organizations Will Augment Traditional Sales Playbooks with AI-Guided Selling Solutions by 2025. https://www.gartner.com/en/newsroom/press-releases/gartner-predicts-75–of-b2b-sales-organizations-will-augment-tra

[15] Gartner, Sales Organizations That Provide AI-Enabled Next Best Actions Are 2.6x More Likely to Achieve Commercial Growth (20 May 2026). https://www.businesswire.com/news/home/20260520536156/en/Gartner-Survey-Finds-Sales-Organizations-That-Provide-AI-Enabled-Next-Best-Actions-Are-2.6x-More-Likely-to-Achieve-Commercial-Growth

[16] Forrester, Forrester's 2026 B2B Marketing, Sales, And Product Predictions: B2B Companies Will Lose More Than $10 Billion Because Of Ungoverned Use Of Generative AI (28 October 2025). https://www.forrester.com/press-newsroom/forrester-b2b-marketing-sales-product-2026-predictions/

[17] CNIL, AI and GDPR: the CNIL publishes new recommendations to support responsible innovation. https://www.cnil.fr/en/ai-and-gdpr-cnil-publishes-new-recommendations-support-responsible-innovation

[18] DPO Partagé, Quand l'IA s'invite dans vos réunions : transcription automatique, comptes rendus intelligents et conformité RGPD (framework of art. 226-1 of the French Criminal Code). https://www.dpo-partage.fr/quand-lia-sinvite-dans-vos-reunions-transcription-automatique-comptes-rendus-intelligents-et-conformite-rgpd/

[19] European AI Regulation (AI Act), AI literacy obligation (art. 4, applicable since 2 February 2025). https://artificialintelligenceact.eu/article/4/

Junyr Agents™: delegating AI in your SME without losing control

Mon, 01 Jun 2026 00:00:00 GMT

According to McKinsey, 62% of companies are experimenting with AI agents in 2026, but only 23% scale them in at least one function [1]. The gap between these two figures is not explained by technology. It is explained by delegation.

An AI agent is not a tool. It is a colleague you can delegate to. And like any colleague, it needs three things to be useful without becoming a risk: a clear mandate, supervision, and a log of its actions. This is what Junyr Agents™ formalises.

AI agent, AI assistant: a distinction that changes everything

The confusion between “assistant” and “agent” is the source of most disappointments.

An assistant works in a synchronous loop: you ask it a question, it answers. ChatGPT in conversation is an assistant. The human stays in control at every step.

An agent works in an asynchronous loop: you give it an objective, and it chains actions together, across several tools and across several steps, until it reaches that objective. The difference is not a question of power. It is a question of nature: the agent acts in the real world. It sends emails, writes to a database, triggers invoicing.

This capacity to act is precisely what creates value, and what demands a control framework. Deploying an agent without a framework is handing over a mandate without a job description.

The three principles of Junyr Agents™

Junyr Agents™ rests on three simple principles, which transpose to AI the basic rules of managerial delegation.

First principle: the explicit mandate. Every agent receives a written, versioned brief, validated by a business champion. This brief sets out the objective, the authorised scope of action, the limits, and the escalation cases. Exactly as a colleague receives a job description before starting. An agent without a written mandate is an agent that can be neither steered nor audited.

Second principle: documented human supervision. Every impactful decision (an outbound send, a write to the database, the triggering of a payment) goes through human validation, unless prior authorisation is explicitly defined in the mandate. Supervision is not a brake: it is what makes it possible to delegate more, because you delegate with confidence.

Third principle: the auditable log. Every action of the agent is traced, time-stamped, attributable, and retained. This log serves day-to-day steering, and it also answers, by design, the record-keeping obligation for logs that the AI Act imposes on high-risk systems. Compliance is not added after the fact: it is built in from the design stage.

The architecture

Junyr Agents™ relies on a stack designed for sovereignty and control: multi-agent orchestration, hosting on controlled infrastructure in France for GDPR compliance, and a mechanism of nightly self-reflection cycles, the “Night Reflections”, that allows the agents to consolidate and verify their work outside production hours.

All of it is integrated into Junyr Mail™ through an “Email Routing” system: the agents can be delegated to, triggered and audited by email, the channel that every SME already masters. And they run across the eight modules of an integrated ERP: HR, accounting, CRM, projects, inventory, purchasing, invoicing, reporting.

Five ready-to-use agents

The quote agent qualifies an incoming request, produces a quote from the catalogue, and passes it to the sales representative for validation. On a documented B2B distribution engagement, this agent brought the time to produce a quote down from 4.2 days to 1.1 days, with a 24% rise in conversion.

The invoicing agent generates compliant invoices (Factur-X format) from accepted orders, checks their consistency, and transmits them securely.

The reporting agent compiles a one-page dashboard every Monday from the past week's ERP data.

The market-watch agent monitors competitors' publications and compiles the news into a weekly note.

The level-1 customer-service agent qualifies incoming requests, answers documented questions, and escalates the rest to human support.

On a second documented engagement (B2B e-commerce with the integration of five AI assistants into the ERP), together they reduced order processing time by 58% and freed up the equivalent of 1.5 full-time posts, with return on investment reached in 9 months.

What an AI agent must never do

The control framework is defined as much by its prohibitions as by its permissions. Four limits are non-negotiable.

An agent never takes a legally binding decision without human validation: contracts, HR decisions, sanctions. It never scores people: the AI Act explicitly prohibits this under the heading of prohibited practices. It never modifies financial data irreversibly without dual validation. And it performs no unlogged external action: every outbound communication is traced.

These limits do not restrain delegation. They make it possible, because they define a clear playing field in which the agent can act fast and the human can keep trusting it.

How to get started

Implementation follows four steps, and the first requires no technology.

First, identify the three most time-consuming processes in the company. Then, for each, pinpoint the sub-process that is genuinely delegable: often qualification, generation, or transmission, rarely the final decision. Next, launch a pilot on a single agent, over 30 days, with a before/after measurement. Finally, if the measurement is conclusive, industrialise, putting the audit log in place from this very step.

This progression is exactly that of the MATIA Method™: you do not skip a step, you build one level (Spectateur → Artisan → Orchestre → Architecte → Pionnier) before moving to the next.

To go further

A demonstration of Junyr Agents™ is available on request. The AI Express Audit & Roadmap (60 minutes by video conference, with no commitment) identifies the three priority agent use cases for your context and positions your SME on the MATIA Method™ Scale.

The white paper “AI Maturity of French SMEs 2025-2026” is available at croissance-transitions.fr. Contact: paul@croissance-transitions.fr.

Sources: verifiable, May 2026

[1] McKinsey (2026), The State of AI, the agentic era, 62% of organisations are experimenting with AI agents, 23% scale them in at least one function.

[2] Regulation (EU) 2024/1689 (AI Act), Article 14 (human oversight), Annex III (high-risk systems), Article 5 (prohibited practices, including the scoring of people).

[3] LangChain, documentation on multi-agent orchestration, langchain.com.

[4] Internal case studies from Croissance et Transitions, B2B distribution and B2B e-commerce engagements, 2024-2025.

Article written by Paul-Antoine TUAL, AI Transformation Leader, creator of the MATIA Method™.

The end of the telephone: why the chat + video pairing is becoming the standard for professional communication in 2026

Sun, 31 May 2026 00:00:00 GMT

By Paul-Antoine TUAL · AI Transformation Leader, Croissance et Transitions · May 2026.

Standpoint. The telephone is not dying. But the voice call that arrives unannounced on a line open to everyone has ceased to be the default channel of the relationship. In its place, a pairing is taking hold: chat for the asynchronous, video for the moment that matters, both now augmented by AI and automatic appointment booking. For an SME leader, this is not a generational fad: it is a change in the infrastructure of the customer relationship, with a sovereignty question at stake. This article sets out the facts, then the method.

What the figures say: the voice call is receding, without disappearing

Let us start by setting sensationalism aside. No, Generation Z does not “hang up in silence”: a YouGov survey from March 2026 shows that this behaviour concerns only 3% of them, and that only 13% consider an unannounced call unacceptable, a figure close to that of other generations [1]. The telephone remains, according to Gartner, a channel valued across all generations [2]. Any narrative announcing “the death of the telephone” is misdiagnosing the situation.

The real phenomenon is more precise, and better documented. First, the volume of voice calls is falling continuously. In France, fixed and mobile voice consumption totalled 200 billion minutes in 2024, down 3% over the year and declining steadily since 2014 [3]. In the United Kingdom, the regulator Ofcom measures a continuing decline in call volume in 2023: −7.7% for calls originating from mobiles and −20% for fixed lines [4].

Next, the preference for the written word is becoming generational and structural. Again according to YouGov, the share of those who prefer to reach their friends and family by telephone falls from 62% among baby-boomers to 17% among Generation Z; 65% of that generation favour the written word (email, text, messaging) [1]. Discomfort with calling exists, but it is targeted: 65% of young people are uncomfortable calling a stranger, against only 15% for calling someone close [1]. In other words, it is not the telephone that is rejected, it is the unannounced call from a stranger.

Finally, and this is the decisive point for a business: calls are not disappearing, they are migrating. CREDOC's 2024 Baromètre du numérique, produced for Arcep, establishes that 85% of the population uses instant messaging apps and that 78% now make their calls through these applications rather than over the traditional telephone network [5]. The same regulator observes that voice communications made over the internet have passed the 415 billion minute mark, while traditional fixed-line telephony has contracted to a fraction of that volume [20]. Voice is doing well: it has simply changed pipe, and that pipe is called WhatsApp, Messenger or a video call.

Why the open line has lost its value: advertising and fraud saturation

If the traditional voice channel is receding, it is also because it has been damaged by its own abuses. The telephone number has become a field saturated with unwanted solicitations, to the point where answering an unknown call is now a gamble.

The figures on exasperation are unequivocal. According to UFC-Que Choisir (October 2024), 97% of French people say they are annoyed by commercial canvassing, 72% are pestered every week on their mobile, for an average of six unsolicited calls per week [6]. Fraud has followed: Arcep recorded 10,973 alerts for abusive calls and messages in 2024, against 2,029 in 2023, a fivefold increase. Number spoofing alone rose from a few hundred reports to more than 8,500 [7]. The 33700 anti-spam service received more than one million reports at the operator Orange alone in 2025 [8], and in January 2026 Arcep opened an investigation into all operators on this subject [9]. On the SMS side, phishing (smishing) remains the leading threat recorded by Cybermalveillance.gouv.fr [10], and as early as 2023 Proofpoint noted that 75% of organisations had suffered attacks of this kind [11].

The behavioural consequence is measured: 46% of unidentified calls now go unanswered, and 92% of consumers believe that a call from an unknown number is probably fraudulent [12]. A line where nine calls out of ten are suspect is no longer a channel of trust.

The legislator has drawn the conclusions. Contrary to what is sometimes read, there is no “law of 13 August 2025”: the reform appears in article 13 of law no. 2025-594 of 30 June 2025, which reverses the paradigm of telephone canvassing [13]. From 11 August 2026, canvassing switches from opt-out to opt-in: it will be prohibited to call a person who has not given their prior consent, defined as an expression of will that is “free, specific, informed, unambiguous and revocable”, with the burden of proof falling on the professional [14]. The Bloctel scheme disappears on the same date [15], and administrative penalties reach €75,000 for a natural person and €375,000 for a legal person [16], beyond which abuse of weakness falls under criminal law (up to €500,000 or 10% of revenue, and five years' imprisonment). The message sent to the market is clear: interrupting someone by telephone without their having asked for it is no longer a commercial strategy, it is a legal risk.

The new pairing: chat for the flow, video for the moment that matters

As the open line loses its value, two channels rise together and divide the roles between them. Chat absorbs the asynchronous, the follow-up, the quick question; video conferencing carries the decisive moment: discovery, demonstration, negotiation, arbitration. These are not two competing modes, it is a complementary pairing.

On the chat side, the scale is now that of a global infrastructure: 94.1% of internet users use a messaging application every month [17], with WhatsApp having passed 3 billion monthly active users in 2025 and handling on the order of 100 billion messages per day [18]. In France, daily use of instant messaging rose from 18.8% of the population in 2018 to 49.2% in 2023 [19], and Arcep measures that 86% of French people aged 12 and over now use it to communicate [20]. The counterpart to this rise is the collapse of the traditional SMS: −27% in one year in the fourth quarter of 2025 (64.7 billion SMS against 88.5 billion a year earlier), and 59 messages per month per customer against 250 at the 2016 peak [21]. Arcep explicitly attributes this decline to the rise of RCS and enriched messaging. The integration of the RCS protocol into the iPhone (iOS 18, late 2024) tipped that standard over, with Google claiming more than one billion RCS messages per day in the United States [22].

This shift is not merely personal: it is redefining the customer relationship. Global business messaging traffic is estimated at 2,000 billion messages in 2025, projected to reach 3,000 billion in 2030 [23]. Meta indicates that one billion users exchange messages with a business every week through its applications, for 600 million conversations per day [24]. And Gartner anticipates that, as early as 2027, self-service and live chat will overtake the telephone and email as the most important customer service technologies [25].

On the video side, the movement is now established rather than spectacular: hybrid meetings have become the norm of work [26], and Microsoft Teams exceeded 320 million monthly active users as early as the start of 2024 [27]. Video has established itself as the format for the first remote appointment, an area that Croissance et Transitions has documented in detail in its 2026 standard for prospecting by video. The point to remember here is complementarity: chat qualifies and follows up, video engages and decides.

The third player: conversational AI and automatic appointment booking

The chat + video pairing would not deliver on its promise without the component that makes it operable at the scale of an SME: AI assistance, and in particular the automation of the first mile of the relationship (qualifying, routing, and booking appointments without friction).

Usage is spreading fast. Salesforce measured (State of Service, early 2025) that the adoption of AI agents in customer service had been multiplied by 1.7 in one year, rising from 39% to 66% of organisations, of which 70% observe measurable value in under 60 days [28]. In France, the 2025 Baromètre France Num indicates that 26% of micro-businesses and SMEs report using AI, with chatbots and conversational assistants already representing 14% of uses [29]. In concrete terms, AI takes on what the telephone did badly: answering at any hour, offering a slot, confirming an appointment, chasing a no-show. According to estimates from vendors in the category (Jamie, Neuwark), AI-assisted scheduling tools save executives on the order of 3 to 5 hours per week [30], and, in healthcare, automated reminders reduce missed appointments by 30 to 40% [31].

Gartner pushes the projection further: by 2028, 70% of customer service journeys would begin and be resolved within conversational assistants integrated into devices [32], and agentic AI could autonomously resolve 80% of common problems by 2029 [33]. These figures are forecasts, to be handled as such, all the more so as the same Gartner warns that more than 40% of agentic AI projects will be cancelled by the end of 2027, for lack of clear business value or risk control (a prediction reiterated in Gartner's Hype Cycle for Agentic AI, April 2026) [34]. The lesson is not “automate everything”, but “automate the repetitive under governance”.

For this productivity seam has a regulatory counterpart that the MATIA Method™ places at the centre. As early as July 2025, the CNIL published its first recommendations on the application of the GDPR to AI [35], and the AI Act imposes human oversight, traceability and transparency. An agent that books an appointment, reads a calendar or drafts a follow-up handles personal data: it must remain an assistant validated by a human, not an opaque automaton. Automatic appointment booking is an excellent gateway to AI in an SME precisely because it is at once high-leverage and manageable in risk, provided the framework is set.

The blind spot: escaping the Google / Microsoft lock-in

There remains the question that most businesses do not ask when adopting the chat + video pairing: with whom, and under what jurisdiction? Because replacing a telephone call with a Teams video call or a Google appointment booking means trading a commoditised channel for a dependence on an integrated ecosystem whose price, terms and applicable law one does not control.

On price, the 2025-2026 trend is clear. Google integrated its Gemini AI into Workspace in January 2025 while raising most subscriptions by around 17%, including for customers who do not use the AI [36]. Microsoft raised its 365 Personal offer to €9.99 per month with Copilot included, against €6.99, a 43% increase [37], with new commercial rates taking effect on 1 July 2026 [38]. Embedded AI is becoming a reason for an increase that the customer does not choose.

On concentration, the regulatory signal came from Brussels. In September 2025 the European Commission accepted binding commitments from Microsoft to close an antitrust investigation opened in 2023: the publisher must now offer Office and 365 without Teams at a lower price, guarantee interoperability and the portability of Teams data, for a period of 7 to 10 years [39]. That decoupling a video tool from an office suite required a two-year European procedure says a great deal about the reality of the lock-in.

On the law, finally, the subject goes beyond commercial convenience. The 2018 CLOUD Act obliges providers under United States jurisdiction to hand over the data they control wherever it is stored, including within the Union [40], while the European Data Act, applicable since September 2025, seeks to close that door [41]. The conflict remains structurally unresolved: complying with one may breach the other [42]. In March 2026, the Conseil d'État judged the risk “sufficiently contained” to approve the hosting of health data on non-European infrastructure, but under strict conditions, the organisation concerned having itself initiated an exit [43]. The point is not to designate a culprit: Croissance et Transitions does not reason in terms of the national origin of a risk, but in terms of applicable jurisdiction and infrastructure dependence. The point is that an SME which entrusts its customer communication to a non-European integrated ecosystem accepts a risk that it has, most often, neither measured nor weighed.

The alternative exists, and it is maturing. The French state gives the strongest signal: its sovereign messaging service Tchap already equips nearly 600,000 civil servants, and in January 2026 minister David Amiel announced the general roll-out of the sovereign application Visio, targeting 200,000 civil servants by the end of 2026 then 2.5 million by 2027, with technical blocking of Teams, Zoom and Google Meet on the inter-ministerial network [44]. On the vendor side, the French ecosystem is credible: Tixeo, video conferencing qualified by ANSSI since 2017 with end-to-end encryption by default [45]; Olvid, the first CSPN-certified messaging service; the open source solutions Jitsi and Element/Matrix; and a collective of collaboration-software vendors, Whaller, Jamespot, Wimi, that position themselves explicitly as an alternative to Microsoft 365 [46]. All of it underpinned by ANSSI's SecNumCloud framework, one of whose criteria is precisely imperviousness to the extraterritoriality of non-European laws [47]. Escaping the lock-in is no longer an activist act: it is an available and qualified architecture option.

What the MATIA Method™ says: from an imposed channel to a governed channel

The MATIA Method™ does not advance an organisation by accumulating tools, but by raising its level of maturity in the use it makes of them. The end of the telephone-by-default illustrates exactly this shift: moving from imposed communication (waiting for the call, filtering the spam, depending on the cheapest suite of the year) to governed communication, where each channel has a role, a data framework and a level of human validation.

In concrete terms, the trajectory comes down to four moves. First, map your channels: who contacts the business, by which channel, and which are today imposed rather than chosen. Second, give chat and video back their place: chat for the asynchronous and follow-up, video for the appointment that commits, abandoning the illusion that the open telephone line remains the channel of trust. Third, automate the first mile under governance: appointment booking, qualification and follow-up entrusted to AI agents, but in human-validated draft mode, never in blind automatic sending. Fourth, choose the infrastructure with awareness of jurisdiction: prefer, when the data is sensitive, a sovereign and auditable foundation.

This is exactly what the Junyr Visio-IA™ module of the Junyr ERP suite embodies: a meeting assistant that augments the salesperson across the whole cycle of the remote appointment (preparing the framing, transcribing, calculating an ROI during the session, laying down a shared action plan, generating the minutes and synchronising the CRM), according to three non-negotiable principles: sovereign processing by default (local transcription and models, resorting to the cloud only after pseudonymisation), explicit and time-stamped consent (CNIL compliance and article 226-1 of the Criminal Code, AI literacy notice under the AI Act), and systematic human-in-the-loop (no minutes, no follow-up, no CRM write leaves without validation). The Junyr Mail™ layer (eIDAS messaging) makes every agent action traceable, auditable and revocable. Technology serves the relationship; it does not replace it.

The counter-argument, taken seriously

Three objections deserve an honest answer.

“The telephone is not dead, you are exaggerating.” That is fair, and it is why the title says “the end of the telephone” in the sense of the default channel, not the channel altogether. Voice remains essential: it simply migrates towards video and in-app calls. Gartner also recalls that the telephone remains valued across all generations [2]. The correct reading is not “remove voice” but “stop making it the sole, imposed point of entry”.

“AI automation will dehumanise my customer relationship.” The risk is real if one automates without a framework. Hence Gartner's warning about the 40% of agentic projects cancelled by 2027 (a prediction reiterated in the Hype Cycle for Agentic AI, April 2026) [34]. The MATIA Method™'s answer is precisely the opposite of dehumanisation: delegating the repetitive (booking a slot, confirming, chasing) to give time back to the relational (the video appointment, the arbitration, the trust). AI unburdens; the human decides.

“Sovereignty is more expensive and less convenient.” The convenience of integrated suites is real, and the Conseil d'État itself judged certain non-European uses “acceptable” under conditions [43]. But the equation is shifting: the price rises linked to embedded AI [36][37], the European antitrust procedure [39] and the growing maturity of ANSSI-qualified alternatives [45][47] are narrowing the gap. Sovereignty is not an activist absolute; it is a trade-off to be made with awareness, sensitive data on one side, cost and convenience on the other.

Three questions to ask this very week. 1. Which inbound channels do we endure (an open line saturated with spam) rather than choose? 2. Does our appointment booking still rest on an unannounced outbound call, which becomes a legal risk on 11 August 2026? 3. If we transcribe our meetings with an AI, do we know where the data is stored, and have we provided for explicit consent?

In practice

Croissance et Transitions supports mid-cap and SME leaders in transforming their customer communication from an imposed channel into a governed one: maturity diagnosis (MATIA Method™), redesign of the chat + video trajectory, governed automation of appointment booking, and selection of a sovereign infrastructure when the data requires it.

→ AI Express Audit & Roadmap: 60 minutes over video. croissance-transitions.fr

Paul-Antoine TUAL · AI Transformation Leader · Croissance et Transitions (SAS) · MATIA Method™ · Junyr Visio-IA™ · Junyr Mail™

FAQ

Will the telephone really disappear for businesses? No. Voice remains essential and valued across all generations (Gartner). What is receding is the unannounced voice call on an open line: the volume is falling continuously (−3% in France in 2024) and calls are migrating towards messaging apps and video. The right reflex is not to remove voice, but to stop making it the sole, imposed point of entry [1][2][3].

What changes for telephone canvassing on 11 August 2026? Law no. 2025-594 of 30 June 2025 (article 13) switches canvassing from opt-out to opt-in: it becomes prohibited to call a person who has not given, beforehand, free, specific, informed and unambiguous consent. Bloctel disappears on that date, and penalties reach €75,000 (natural person) and €375,000 (legal person) [13][14][15][16].

Why speak of a “pairing” of chat + video rather than a single channel? Because the two channels divide the roles: chat absorbs the asynchronous, the quick question and the follow-up (94% of internet users use a messaging service every month), while video carries the moment that commits (discovery, demonstration, negotiation). Gartner anticipates that as early as 2027 self-service and chat will overtake the telephone and email in customer service [17][25].

Is appointment-booking AI compatible with the GDPR? Yes, provided it is governed. An agent that reads a calendar, offers a slot or drafts a follow-up handles personal data: the CNIL (July 2025 recommendations) and the AI Act impose human oversight, traceability and transparency. The right model is “human-validated draft”, never blind automatic sending [35].

Must one leave Teams or Google Meet for a sovereign solution? It is not an obligation, it is a trade-off. The price rises linked to embedded AI (+17% Google Workspace, +43% Microsoft 365 Personal), the European antitrust procedure over Teams and the CLOUD Act / GDPR conflict justify evaluating the option. ANSSI-qualified alternatives exist (Tixeo, Olvid, Jitsi, Element/Matrix), underpinned by the SecNumCloud framework. The rule: a sovereign foundation when the data is sensitive [36][37][39][45][47].

Sources

[1] YouGov, Mythbusting claims about Gen Z and their phone habits (2 March 2026). https://yougov.com/en-gb/articles/54114-mythbusting-claims-about-gen-z-and-their-phone-habits

[2] Gartner, Traditional Customer Service Channels Are Losing Ground to Mobile and AI Innovations (10 February 2025). https://www.gartner.com/en/newsroom/press-releases/2025-02-10-traditional-customer-service-channels-are-losing-ground-to-mobile-and-ai-innovations

[3] Arcep, Observatoire des marchés des communications électroniques en France, Année 2024 (2025). https://www.arcep.fr/cartes-et-donnees/nos-publications-chiffrees/observatoire-des-marches-des-communications-electroniques-en-france/marche-communications-electroniques-france-2024-resultats-definitifs.html

[4] Ofcom, Communications Market Report 2024 (18 July 2024). https://www.ofcom.org.uk/siteassets/resources/documents/research-and-data/multi-sector/cmr/cmr24/communications-market-report-2024.pdf

[5] CREDOC / Arcep, Baromètre du numérique 2024 (2025). https://www.economie.gouv.fr/files/files/media-document/Barometre_numerique_2024.pdf

[6] UFC-Que Choisir, Démarchage téléphonique : mettons enfin un terme au harcèlement marketing (October 2024). https://www.quechoisir.org/action-ufc-que-choisir-demarchage-telephonique-mettons-enfin-un-terme-au-harcelement-marketing-n132374/

[7] Arcep / Next, Spam : l'Arcep confirme une explosion des appels et messages abusifs en 2024 (April 2025). https://next.ink/178902/spam-larcep-confirme-une-explosion-des-appels-et-messages-abusifs-en-2024/

[8] Orange, Le 33700 contre le smishing (2025). https://bienvivreledigital.orange.fr/securite/attention-arnaques/victime-de-smishing-le-33700-peut-vous-aider.html

[9] Arcep, Protection des consommateurs, enquête usurpation de numéro et spam (January 2026). https://www.arcep.fr/actualites/actualites-et-communiques/detail/n/protection-des-consommateurs-290126.html

[10] Cybermalveillance.gouv.fr, Rapport d'activité 2024 (March 2025). https://www.cybermalveillance.gouv.fr/medias/2025/03/250327_RA_2024_SCREEN.pdf

[11] Proofpoint, 2024 State of the Phish (2024). https://www.proofpoint.com/us/resources/threat-reports/state-of-phish

[12] Hiya, State of the Call 2024 (2024). https://blog.hiya.com/2024-state-of-the-call-consumers-prefer-voice-but-spam-and-fraud-are-threats

[13] Légifrance, Article 13, LOI n° 2025-594 du 30 juin 2025 (30 June 2025). https://www.legifrance.gouv.fr/jorf/article_jo/JORFARTI000051824325

[14] Légifrance, Code de la consommation, art. L223-1 (in force on 11 August 2026). https://www.legifrance.gouv.fr/codes/article_lc/LEGIARTI000051830285/2026-08-11

[15] service-public.gouv.fr, Démarchage téléphonique : les nouvelles règles (2025). https://www.service-public.gouv.fr/particuliers/actualites/A18384

[16] economie.gouv.fr, Professionnels : respecter la réglementation sur le démarchage (2025). https://www.economie.gouv.fr/entreprises/developper-son-entreprise/innover-et-numeriser-son-entreprise/professionnels-comment-respecter-la-reglementation-sur-le-demarchage

[17] We Are Social / Meltwater (via Statista), Chat and messenger service usage (Q2 2025). https://www.statista.com/statistics/1489440/chat-and-messenger-service-usage/

[18] Meta (via Rest of World), How WhatsApp for Business changed the world (2024-2025). https://restofworld.org/2024/how-whatsapp-for-business-changed-the-world/

[19] Médiamétrie (via Siècle Digital), Usage du numérique en France, tendances 2024 (2025). https://siecledigital.fr/2025/02/17/usage-du-numerique-en-france-les-tendances-et-chiffres-2024-de-mediametrie/

[20] Arcep, Observatoire des communications électroniques, T4 2025 (9 April 2026). https://www.arcep.fr/cartes-et-donnees/nos-publications-chiffrees/observatoire-des-marches-des-communications-electroniques-en-france/t4-2025.html

[21] Arcep, Observatoire T4 2025 (SMS decline −27%) (9 April 2026). (same source as [20])

[22] Google (via 9to5Google), Google RCS passes 1 billion daily messages in the US (13 May 2025). https://9to5google.com/2025/05/13/google-rcs-messages-billion-daily-us/

[23] Juniper Research, Conversational use cases fuel global messaging boom (1 September 2025). https://www.juniperresearch.com/press/conversational-use-cases-fuel-global-messaging-boom/

[24] Meta (via CX Today / Rest of World), Business messaging, 1 billion users/week (2024). https://restofworld.org/2024/how-whatsapp-for-business-changed-the-world/

[25] Gartner, Self-Service and Live Chat Will Surpass Traditional Channels by 2027 (27 August 2025). https://www.gartner.com/en/newsroom/press-releases/2025-08-27-gartner-survey-finds-self-service-and-live-chat-will-surpass-traditional-channels-as-top-customer-service-technologies-by-2027

[26] Owl Labs, State of Hybrid Work 2024 (September 2024). https://owllabs.com/state-of-hybrid-work/2024

[27] Microsoft (via Office365ITPros), Teams passes 320 million users (2023-2024). https://office365itpros.com/2023/10/26/teams-number-of-users-320-million/

[28] Salesforce, State of Service Report (6th edition) (2025). https://www.salesforce.com/service/state-of-service-report/

[29] France Num / Bpifrance Big Média, Baromètre France Num 2025, chatbots for micro-businesses and SMEs (2025). https://bigmedia.bpifrance.fr/nos-dossiers/chatbots-pour-tpe-ameliorer-le-service-client-avec-lia-conversationnelle

[30] Jamie, AI scheduling assistant, time savings (2025). https://www.meetjamie.ai/blog/ai-scheduling-assistant

[31] Neuwark, AI patient engagement, reducing no-shows (2025-2026). https://neuwark.com/blog/ai-patient-engagement-reduce-no-shows-conversational-ai-2026

[32] Gartner, 70% of customer service journeys resolved within conversational assistants by 2028 (10 February 2025). (see [2])

[33] Gartner, Agentic AI Will Autonomously Resolve 80% of Common Customer Service Issues by 2029 (5 March 2025). https://www.gartner.com/en/newsroom/press-releases/2025-03-05-gartner-predicts-agentic-ai-will-autonomously-resolve-80-percent-of-common-customer-service-issues-without-human-intervention-by-20290

[34] Gartner, Hype Cycle for Agentic AI (April 2026), reiterates the initial prediction of June 2025. https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027

[35] CNIL, GDPR & AI recommendations (22 July 2025) & 2026 work programme. https://www.cnil.fr/fr/accompagnement-des-professionnels-le-programme-de-travail-de-la-cnil-pour-2026

[36] Incentro, Google Workspace price increase 2025 (+17%) (January 2025). https://www.incentro.com/en-EAF/news/google-workspace-price-increase-2025

[37] Microsoft 365 Blog, Advancing Microsoft 365, pricing update (4 December 2025). https://www.microsoft.com/en-us/microsoft-365/blog/2025/12/04/advancing-microsoft-365-new-capabilities-and-pricing-update/

[38] Microsoft Licensing, 2026 M365 packaging & pricing updates FAQ (2026). https://www.microsoft.com/en-us/licensing/news/2026-m365-packaging-pricing-updates-faq

[39] European Commission, Commitments accepted, Microsoft Teams (IP/25/2048) (12 September 2025). https://ec.europa.eu/commission/presscorner/detail/en/ip_25_2048

[40] LeMagIT, CLOUD Act : entre le marteau et l'enclume (2025). https://www.lemagit.fr/tribune/CLOUD-Act-entre-le-marteau-et-lenclume

[41] Kiteworks, EU Data Act, RGPD et conflit avec le cloud (September 2025). https://www.kiteworks.com/fr/conformite-rgpd/eu-data-act-rgpd-conflit-cloud/

[42] Kiteworks, Structural CLOUD Act / GDPR conflict (2025). (see [41])

[43] Usine Digitale, Le Conseil d'État valide Microsoft pour l'hébergement des données de santé (March 2026). https://www.usine-digitale.fr/cybersecurite/data-protection/donnees-de-sante/le-conseil-detat-valide-microsoft-pour-lhebergement-des-donnees-de-sante-et-juge-le-risque-lie-au-cloud-act-acceptable.T6ZKEMHAKRFAJKURNJOFS3WFEU.html

[44] Journal du Geek, L'État va se passer de Teams et Zoom avec l'app souveraine Visio (January 2026). https://www.journaldugeek.com/2026/01/31/letat-va-se-passer-de-teams-et-zoom-avec-lapp-souveraine-visio/

[45] Le Monde Informatique, Tixeo devient une suite collaborative souveraine (2025). https://www.lemondeinformatique.fr/actualites/lire-tixeo-devient-une-suite-collaborative-souveraine-99777.html

[46] Blog Whaller, Cloud souverain : 8 acteurs français alternative à Microsoft 365 (2024-2025). https://blog.whaller.com/presse/cloud-souverain-8-acteurs-francais-serigent-en-alternative-a-microsoft-365/

[47] ANSSI, SecNumCloud, enjeux technologiques / cloud (2025). https://cyber.gouv.fr/enjeux-technologiques/cloud/

Cryptography 2026: transport, storage, identities and the post-quantum transition, the decision framework for business leaders and CISOs

Fri, 29 May 2026 00:00:00 GMT

Cryptography is no longer a matter for engineers. It has become a leadership decision. Three underlying shifts are converging in 2026: the post-quantum transition, the industrialisation of attacks on identity, and a European regulatory framework that is becoming clearer. Each one calls for trade-offs that only a business leader can settle: which data to protect first, what level of sovereignty to aim for, at what pace to modernise.

The good news is that the standards exist, the timeline is known, and the path is marked out by ANSSI, NIST and industry best practice. There is no cause for panic; there is a method to apply. This article sets out the decision framework in three stages (data in transit, data at rest, identities), then places it within the maturity grid of the MATIA Method™ so that you can decide where to focus your effort this year.

The one-sentence takeaway. The winning reflex in 2026 is not to buy yet another tool, but to establish crypto-agility: the ability to change algorithm without overhauling the information system with every new development.

1. The post-quantum transition: a timeline, not an alarm

1.1. What is changing, and why it is manageable

The asymmetric cryptography that secures the Internet (RSA, elliptic curves) relies on mathematical problems that a sufficiently powerful quantum computer, described as “cryptographically relevant” (CRQC), would be able to solve quickly thanks to Shor's algorithm. This CRQC does not yet exist. The issue for a business leader is therefore not tomorrow's machine, but a practice of today to be anticipated calmly.

This practice has a name: “Store-Now, Decrypt-Later” (SNDL), or “harvest now, decrypt later”. Actors with substantial resources can intercept and store encrypted communications today (TLS traffic, VPN tunnels), with the intention of decrypting them once quantum computing power becomes available. The consequence is simple and actionable: any data whose confidentiality must hold for more than five to ten years deserves to be protected right now by quantum-resistant mechanisms.

This is why priority goes to confidentiality (key exchange) rather than to signature: impersonating an identity through signature would require a quantum computer operational in real time, whereas data intercepted today simply waits.

1.2. Stable standards to build on

In August 2024, NIST published its first finalised post-quantum cryptography standards. They no longer rely on factorisation but on different mathematical families, designed to resist both classical and quantum attacks.

Standard	Origin	New name	Function	Primary use
FIPS 203	CRYSTALS-Kyber	ML-KEM	Key exchange (KEM)	Confidentiality in transit, protection against SNDL
FIPS 204	CRYSTALS-Dilithium	ML-DSA	Digital signature	Server authentication, certificates, PKI
FIPS 205	SPHINCS+	SLH-DSA	Digital signature	Fallback to ML-DSA, software/firmware integrity
SP 800-208	LMS / XMSS	LMS / XMSS	Stateful signature	Firmware updates

These standards allow software vendors, operating systems and agencies to begin large-scale integration. The ground is mapped out.

1.3. The transition doctrine: hybridisation, a prudent choice

Post-quantum algorithms are robust, but recent: they do not yet have the decades of analysis that RSA or elliptic curves have undergone. Rather than replacing everything at once, industry and the European agencies recommend hybridisation: combining, for the duration of the transition, a proven classical algorithm (ECDH) with a post-quantum algorithm (ML-KEM). To compromise the session, an adversary would have to break both at once, one requiring a quantum computer, the other being immune to it. This is a belt-and-braces approach, and therefore reassuring.

National doctrines diverge on the tempo, which has a concrete implication for international companies:

Agency	Country	Stance on hybridisation
ANSSI	France	Required; primary and immediate transition strategy
BSI	Germany	Strongly recommended; aligned with ANSSI / ENISA
NCSC	United Kingdom	Transitional measure; prefers a direct transition in the long run
NSA (CNSA 2.0)	United States	Direct transition to pure PQC for national security systems
CCCS	Canada	Neutral

The practical consequence: designing crypto-agile architectures, able to negotiate hybrid modes in Europe (ANSSI/BSI compliance) while being able to switch to pure PQC modes elsewhere. This is precisely the flexibility that crypto-agility aims for.

2. Data in transit: modernising without breaking the network

Three protocols structure secure transport: TLS (web, API), IPsec (site interconnections, remote working) and SSH (administration). Their move to post-quantum is above all a matter of network engineering: post-quantum keys are larger, and some older equipment copes poorly with bigger packets. The good news: the solutions are known and documented.

2.1. TLS 1.3 and QUIC

TLS 1.3 accounts for more than 93% of secure connections (Cloudflare, 2024) and serves as the foundation for QUIC (HTTP/3). The priority is to hybridise the key establishment phase, with configurations such as X25519MLKEM768. A point to watch: a hybrid key exceeds one kilobyte (compared with ~32 bytes for a classical exchange), which can push the first message (“ClientHello”) beyond the standard size and cause fragmentation that is poorly handled by older firewalls or load balancers, a phenomenon of network “ossification”.

The recommended action is simple: audit the network inspection chain to check its tolerance for these larger packets. The GREASE mechanism (built into TLS 1.3) already helps to prevent filtering rules from becoming frozen. For hybrid X.509 certificates, ANSSI recommends waiting for the IETF standards to be finalised: the priority remains confidentiality, not authentication.

2.2. IPsec and IKEv2

The IPsec data plane (ESP/AH, using AES-256-GCM) resists quantum attacks well. It is the establishment phase, driven by IKEv2 over UDP, that requires hybridisation. As UDP handles very large packets poorly, ANSSI and the IETF rely on a stack of standards that are now available:

RFC 7383: fragmentation at the IKE level (avoids blind IP fragmentation);
RFC 9242: intermediate exchange (IKE_INTERMEDIATE) to carry large payloads reliably;
RFC 9370: negotiation of multiple key exchanges (hybridisation proper), integrated into StrongSwan 6.0+;
RFC 8784: post-quantum pre-shared keys (PPK), an immediate stopgap for closed systems.

The concrete action: update the firmware of VPN concentrators and firewalls to versions supporting these RFCs. Without this upgrade, hybridisation would block the establishment of tunnels, hence the importance of planning these updates before switching over.

2.3. SSH

SSH is the gateway to administration. OpenSSH was a pioneer: from version 9.0, hybrid key exchange by default (X25519 + NTRU Prime); with 10.0, ML-KEM hybridisation by default, aligned with FIPS 203. The recommended action: upgrade the administration fleet to OpenSSH 10.0+. Key authentication (users and hosts) remains classical for want of a finalised IETF standard; in the meantime, frequent key rotation is automated (Vault, Ansible) and administration networks are segmented.

2.4. Group messaging: the MLS standard

For end-to-end encrypted (E2EE) group communications, the historical approach (Signal's “Double Ratchet”) scales poorly: adding or removing a member in a group of n people costs an effort proportional to n. In July 2023 the IETF published the Messaging Layer Security (MLS, RFC 9420) standard, which reduces this cost to a logarithmic order thanks to a tree structure (TreeKEM). MLS is already in production (Cisco Webex, RCS (Universal Profile 3.0)) and was designed to natively incorporate future post-quantum mechanisms, a useful marker for choosing a collaboration platform that will last.

3. Data at rest: the real question is that of keys

For storage, the encryption algorithm is not the issue: AES-256 remains the norm and resists quantum attacks. Grover's algorithm halves the effective strength of a symmetric key: an AES-128 key would fall to a vulnerable level, whereas AES-256 retains a comfortable security margin. The real challenge shifts to key management and sovereignty.

3.1. Envelope encryption

Modern cloud relies on a hierarchy of keys: a DEK (data encryption key, ephemeral) encrypts the files; a KEK (master key) encrypts the DEK. The DEK is never stored in the clear. KEKs must be generated, managed and stored in HSMs (hardware security modules) validated to FIPS 140-3 Level 3, via a KMS service. Every decryption operation takes place in the protected memory of the HSM.

3.2. Sovereignty: the question to settle at leadership level

The risk is not only criminal (ransomware targeting backups), it is also legal. The CNIL, drawing on ANSSI, warns: if the cloud provider holds both the data and the master keys, it can technically be compelled, by legislation with extraterritorial reach (the US CLOUD Act, Section 702 of FISA, intelligence laws of third-party jurisdictions), to provide data in the clear. Sovereignty is therefore reasoned in terms of jurisdiction and infrastructure, not the presumed national origin of threats.

Three levels of key control, to be weighed according to the sensitivity of the data:

By default: the provider manages everything. Simple, but no protection against an extraterritorial order.
BYOK (Bring Your Own Key): you generate your keys and import them. Better control of rotations, but the provider retains technical access in memory during processing.
HYOK / External KMS (the CNIL's “Approach 4”): you retain physical control of the KEKs in your own HSMs (on-premises or a sovereign provider). This is the only option that guarantees strict cryptographic sovereignty; it is more demanding to integrate. It is also the spirit of ANSSI's SecNumCloud qualification.

3.3. Protecting the integrity of backups

Encryption is no longer enough: attackers target backups in order to erase or alter them. Robust hash functions (SHA-384, SHA-512, the SHA-3 family) and integrity mechanisms (MAC) are used to guarantee the immutability of offline backups, the condition for recovering calmly from an attack.

4. Identities: the password alone has had its day

This is probably the project with the best effort-to-impact ratio for an SME in 2026.

4.1. Storing passwords correctly

In a database, a password is never stored in the clear, and fast hashes (MD5, SHA-1), crackable by GPU farms, are banned. The OWASP/ANSSI rule: use “memory-hard” functions, hungry for RAM, which make the attack economically unattractive.

Algorithm	Status	Configuration benchmark
Argon2id	Recommended (de facto standard)	m = 19 MiB minimum, t = 2, p tuned to the server
scrypt	Good alternative	N = 2¹⁷ (≈128 MiB), r = 8
bcrypt	Legacy	Work factor > 10; beware of truncation beyond 72 bytes
PBKDF2	Tolerated (compliance)	Not “memory-hard”: to be avoided for any new project

A salt (unique, against rainbow tables) is systematically added and, in high-security settings, a pepper (a global secret key stored in an HSM). These mechanisms are quantum-resistant as long as the underlying hash function is robust.

4.2. Why “classic” MFA no longer protects

MFA via SMS, one-time code (TOTP) or push notification was considered sufficient. It is no longer sufficient against “Adversary-in-the-Middle” (AiTM) proxy attacks, which have become industrialised.

The principle, put simply: the attacker places a transparent proxy between you and the real site. You see the genuine login page (relayed in real time), you enter your credentials and then your MFA code. The attacker captures, in passing, the session token issued after validation. From that point on, your password no longer matters: the attacker holds a legitimate session. These kits are sold “turnkey” (Phishing-as-a-Service), sometimes combined with vishing (a phone call from a fake support desk, possibly assisted by a synthetic voice). Recent incidents have led, by this method, to the theft of SSO credentials and the exfiltration of CRM databases, something no amount of vigilance training could have prevented.

The observation is stated without dramatising: no method relying on the human repetition of a secret withstands interception by proxy. The countermeasure exists, and it is mathematical.

4.3. FIDO2 / passkeys: the countermeasure by design

Passkeys (the FIDO2 / WebAuthn standard) solve the problem through a principle called origin binding. At registration, your device generates a key pair specific to the service; the private key never leaves the secure enclave (TPM, Secure Enclave, a hardware key such as a YubiKey). At sign-in, the browser itself verifies the exact domain and only signs the challenge if it matches the registered domain. On a fake site, the signature is refused automatically: the attack fails without the user having to judge the validity of the URL.

Adoption reached critical mass in 2026: more than 5 billion passkeys in use, 68% of organisations deploying them. Pioneering companies (Cloudflare, Snap) report zero compromise through phishing after a full switch-over and the deactivation of vulnerable fallback methods.

5. The European regulatory framework: a horizon becoming clearer

Two European matters deserve the attention of business leaders, not to cause worry but to anticipate.

eIDAS 2.0 (Article 45 / QWAC). The regulation creates a European digital identity wallet, a useful step forward. One technical provision is contested: the obligation on browsers to embed certain certificate authorities endorsed by member states (QWAC) without being able to remove them freely. Browser vendors, cryptographers and digital-rights organisations (EFF, Internet Society) warn that this could slow incident response (revocation of a compromised authority) and set a precedent liable to be invoked by other jurisdictions with weaker democratic safeguards, fuelling the fragmentation of the Internet. Leadership action: for critical applications, put certificate pinning in place and monitor Certificate Transparency.

“Chat Control” and end-to-end encryption. The proposal to pre-emptively scan messaging services in order to combat child sexual abuse material ran up against a technical impossibility: scanning end-to-end encrypted messages would require “client-side scanning” (analysis on the device before encryption), which amounts to weakening encryption for everyone and undermines professional confidentiality (lawyers, doctors, journalists). In late 2025 and early 2026, at Denmark's instigation, the Council of the EU amended the text to exclude any obligation of forced detection on E2EE spaces, a welcome clarification. The matter is entering the trilogue phase; vigilance remains warranted, but the direction is reassuring.

6. Where to focus effort in 2026: reading it through the MATIA Method™

Not everything is done at once. The MATIA Method™ grades AI maturity across five levels (Spectateur, Artisan, Orchestre, Architecte, Pionnier). For cryptography, effort concentrates on three of them (Artisan, Orchestre, Architecte), which provide a clear and workable order of march. Here is how cryptography fits into them:

Artisan level: securing the foundations (0-6 months). The high-impact, low-complexity moves: roll out passkeys / FIDO2 for privileged accounts (leadership, IT, finance) and disable vulnerable fallback MFA methods; check that password storage uses Argon2id; confirm that all encryption at rest uses AES-256. This is where most of the resilience is won, and quickly.

Orchestre level: structuring crypto-agility (6-18 months). Map encrypted flows and certificates (a cryptographic inventory), enable TLS 1.3 hybridisation on exposed services, plan network updates (IPsec/IKEv2, OpenSSH 10.0+), and weigh the level of key sovereignty (BYOK vs HYOK) according to the sensitivity of the data. The objective: to be able to change algorithm without rebuilding everything.

Architecte level: steering by data and duration (18 months and beyond). Classify data by the required confidentiality period and prioritise post-quantum protection for the longest (the SNDL logic); integrate a sovereign External KMS / HYOK architecture; choose collaboration platforms built on durable standards (MLS); integrate regulatory monitoring (eIDAS, E2EE) into governance.

This progression avoids two pitfalls: wait-and-see (“we will see when quantum arrives”) and costly haste (“let us replace everything”). For organisations that wish to delegate part of this orchestration (inventory, certificate monitoring, key rotation), Junyr Agents™ can automate repetitive tasks under supervision, in a logic of method rather than an accumulation of tools.

Conclusion

Cryptography has changed status: it now determines an organisation's digital trust as much as its compliance. But the picture in 2026 is not that of an insurmountable threat. It is that of a mapped-out path: stable standards (NIST, ANSSI), a prudent doctrine (hybridisation), proven countermeasures (FIDO2), and a European framework that is becoming clearer. The business leader's task is not to do everything at once, but to structure crypto-agility and to place the effort in the right place, in the right order. It is achievable, and it is even a competitive advantage for those who go about it methodically.

To position your organisation on the maturity scale and build your roadmap: AI & Security Express Audit & Roadmap.

Sources: direct verified links (audit 29 May 2026)

Links verified on 29 May 2026.

Post-quantum, standards and transition 1. NIST, NIST Releases First 3 Finalized Post-Quantum Encryption Standards 2. NIST CSRC, Post-Quantum Cryptography 3. Akamai, A Guide to International Post-Quantum Cryptography Standards 4. Akamai, Post-Quantum Cryptography Implementation Considerations in TLS 5. Palo Alto Networks, A Complete Guide to Post-Quantum Cryptography Standards 6. ANSSI, Follow-up position paper on Post-Quantum Cryptography (addendum 2023)

Transport (TLS, QUIC, IPsec, SSH) 7. ANSSI, Transition post-quantique du protocole TLS 1.3 8. ANSSI, Transition post-quantique de SSHv2 (PDF) 9. ANSSI, Transition post-quantique du protocole IPsec 10. DataGuidance, France: ANSSI releases guide on Post-Quantum Transition of IPsec 11. blog.ogwilliam.com, Concrete Technical Steps for Post-Quantum TLS, SSH, and IPsec 12. AWS Security Blog, Enable post-quantum key exchange in QUIC with the s2n-quic library 13. Cloudflare Blog, The state of the post-quantum Internet

Group messaging / MLS 14. IETF Datatracker, RFC 9420, The Messaging Layer Security (MLS) Protocol 15. Feisty Duck, RFC 9420: Messaging Layer Security 16. Gopher Security, Understanding Messaging Layer Security 17. YouTube (37C3), RFC 9420 or how to scale end-to-end encryption with Messaging Layer Security

Storage, KMS, cloud sovereignty 18. Google Cloud, Envelope encryption / Chiffrement encapsulé (Cloud KMS) 19. AWS, Protection des données dans AWS Key Management Service 20. Entrust, Quelles sont les meilleures stratégies de gestion des clés d'entreprise ? 21. CERT-FR (ANSSI), Secteur du cloud, État de la menace informatique (CERTFR-2025-CTI-001) 22. CNIL, Les pratiques de chiffrement dans l'informatique en nuage (cloud) public 23. ANSSI, Recommandations de déploiement d'un service IaaS OpenStack SecNumCloud (ANSSI-BP-104)

Password hashing 24. OWASP, Password Storage Cheat Sheet 25. ANSSI / MonServiceSécurisé, Protéger les mots de passe stockés sur le service 26. ANSSI, Guide des mécanismes cryptographiques : règles et recommandations (ANSSI-PG-083)

AiTM, MFA, FIDO2 / passkeys 27. The Hacker News, How AitM Phishing Attacks Bypass MFA and EDR, and How to Fight Back 28. Luxgap, Okta/SSO hit by vishing: how FIDO2 blocks MFA bypass 29. WorkOS, Passkeys stop phishing. Your MFA fallbacks undo it. 30. FIDO Alliance, The State of Passkeys 2026: Global Consumer and Workforce Report

eIDAS 2.0 31. EFF, eIDAS 2.0 Sets a Dangerous Precedent for Web Security 32. R Street Institute, Cybersecurity Score, European Union Electronic Identification, Authentication, and Trust Services (eIDAS 2.0)

Chat Control / E2EE 33. European Newsroom, Privacy vs. child protection: EU's “chat control” plans split member states 34. EFF, EU Parliament Blocks Mass-Scanning of Our Chats, What's Next? 35. EFF, After Years of Controversy, the EU's Chat Control Nears Its Final Hurdle: What to Know 36. EDRi, Chat Control is in the final stretch, but it could be a marathon, not a sprint 37. Tech Policy Press, How Europe's “Chat Control” Regulation Could Compromise American Communications

The end of prompt engineering: why your teams must stop talking to AI and start commanding it

Tue, 19 May 2026 00:00:00 GMT

By Paul-Antoine TUAL (AI Transformation Leader, Croissance et Transitions). Updated 19 May 2026.

The founding misunderstanding

On 30 November 2022, OpenAI launched ChatGPT. The interface looked like a messaging app. You wrote in it as you would write to a colleague. The machine replied with surprising politeness, nuance, sometimes humour. And one word settled into the vocabulary of businesses: prompt. People learned to “talk to AI properly”. Teams were trained in “the perfect query”. A job title was coined: prompt engineer.

Three and a half years later, in May 2026, this word has had its day. Not because artificial intelligence has disappointed: it has never performed better. But because the analogy that carried it, talking to AI as if to a human, was an illusion of interface, not a technical truth. Andrej Karpathy, former head of AI at Tesla and a guiding figure of the field, wrote so publicly as early as June 2025 on X: he now prefers the term context engineering to prompt engineering, because “prompt suggests a short task description, whereas in every industrial LLM application what matters is the delicate art of filling the context window with exactly the right information for the next step” [1]. Tobi Lütke, head of Shopify, backed the same shift [2]. In the summer of 2025 Gartner published a note whose title sums up the pivot: Lead the Shift to Context Engineering as Prompt Engineering Fades [3].

This text rests on a simple thesis. A large language model is not a human. It has no intuition, no goodwill, no memory of the previous conversation unless it is re-injected. It does one thing, and one thing only: at each step, it computes the most probable next token given all the tokens that precede it. “Talking” to it as to a human is a useful cultural convention for the general public; in a business, it is a strategic mistake. The right way to instruct it, in 2026, is the way one adopts with any desktop computer: clear, explicit, structured instructions. And the form best suited to carry these instructions on the market's dominant models, Claude (Anthropic), GPT-5 (OpenAI), Gemini (Google), is structured markup: XML, JSON or explicit delimiters depending on the model and the task.

This is not an opinion. It is what the vendors themselves recommend in their official documentation. It is what independent benchmarks measure. It is what every team that puts AI into production practises, without formalising it. And it is, for a French SME that wants to draw real value from its AI tools in 2026, the most poorly understood lever on the market.

What an LLM really is, for a leader who does not have time to read a research paper

To understand why the “talking to a human” analogy is misleading, you only need to look under the bonnet. No mathematics. Three mechanisms are enough.

The tokenizer. Before a model “reads” your sentence, a deterministic program cuts it into small pieces called tokens. A token is neither a word, nor a syllable, nor a character. It is a statistical unit learned from billions of pages of text. Each token is given a unique identifier number in a vocabulary of the order of 100,000 to 200,000 entries [4]. When you write “Draft the minutes of a meeting”, the machine does not see your sentence: it sees a sequence of numbers.

The transformer. This is the neural architecture that has prevailed since 2017. Its defining feature: it takes the sequence of tokens as input, and produces a single thing as output: a probability distribution over the next token. The model does not “understand” your question. It computes: given this sequence of numbers, what is the most probable next number? Then it chooses. Then it starts again. Word by word, token by token, it generates its answer [5].

Attention. At each step, each token “looks at” the other tokens in the sequence and computes how much it should rely on each one. When a model processes the word “lawyer” in “the client consulted their lawyer”, attention weights the tokens “client” and “consulted” more than “the”. This is also why the format of the input matters as much as its content [6, 7].

An important nuance deserves to be stated here, because it directly conditions practice: a model's attention does not treat every position in your prompt as equally valuable. A now well-documented phenomenon, sometimes called context rot or lost in the middle, shows that transformers massively favour the beginning of the query (primacy effect) and the end of the submitted text (recency effect). The probability that a crucial piece of information buried in the middle of a long narrative query is correctly taken into account can drop markedly as the context lengthens: work on long contexts measures major losses of usable reliability [26]. Practical conclusion: the longer and less marked-up the prompt, the more attention goes astray. Explicit markup is not a typographic affectation: it is a map you give to the attention mechanism so that it does not get lost.

There, in three mechanisms, is what an LLM is: a statistical splitter, a probabilistic predictor, an attention orchestrator. None is anthropomorphic. None “understands” in the human sense of the word. The academic debate on this point remains open [8, 9]. But for an SME leader who must decide how to train their teams, the practical conclusion is clear: the machine responds better to what looks like program instructions than to what looks like a café conversation.

Why the chat interface has distorted the debate

The analogy error was born of the interface, not the technology. Before ChatGPT, language models were APIs consumed by developers, in scripts, with formatted inputs. In November 2022, OpenAI made two product choices that changed collective perception: the chat window, and the “helpful assistant” tone. The first creates the expectation of a conversation. The second creates the illusion of intent.

The cognitive-science literature describes this phenomenon under the term anthropomorphism. Conversational interfaces reinforce this bias through the simulation of turn-taking, artificial response delays, and first-person vocabulary (“I think that…”) [10, 11]. A recent review speaks of a “double-edged sword”: anthropomorphism eases adoption, but it masks the crucial differences between humans and LLMs, which leads to over-confidence in the answers and a poor calibration of uses in business [11].

For an SME, this bias has a concrete cost. When a leader believes they must “talk to ChatGPT properly”, they steer their training towards the rhetoric of the query: polite phrasing, an “act as if you were an expert” example, promises of reward. Some of these recipes circulated massively on LinkedIn between 2023 and 2025. Recent studies show that most bring no measurable gain in accuracy, and that some degrade performance [12].

The 2025-2026 turning point: from “prompt” to “instruction”

What changes in 2025-2026 is that the vendors themselves are formalising the other path. Three signals converge, and they must be set in a broader frame. The DORA 2025 report (Google Cloud), which draws on nearly 5,000 responses from tech professionals and more than 100 hours of qualitative interviews, highlights a significant dissonance: while ~90% of developers say they use some form of AI assistance and 80% consider that it increases their individual productivity, organisational delivery indicators often remain flat: individual gain without systematic collective gain [27]. The gap between adoption and real productivity at the organisation level is the clearest indicator of the method problem.

Anthropic. The official Claude documentation explicitly recommends the use of XML tags to structure prompts. The wording is unequivocal: “Claude has been specifically trained to pay close attention to your structure” [13]. The guide cites use cases: <instructions>, <context>, <documents> wrapping each indexed <document>, <examples>, <thinking> and <answer> to distinguish reasoning from the answer.

OpenAI. The GPT-5 guide is explicit: “GPT-5 interprets prompts literally and exhaustively”, and recommends “structured XML specifications such as <[instruction]_spec>” to improve instruction-following [14, 15]. The model is tuned for precision: it will do exactly what is written, without liberal interpretation. This makes ambiguous instructions more costly in hallucinations than before.

Google. Gemini's official Prompting Guide makes the same recommendation: frame instructions, examples and reference content with explicit tags or separators, because the model uses these boundaries to focus its attention on the right portion of the context [16].

Beyond the vendors, academic research produced several works in 2024-2025 that consolidate this finding. StructEval (arXiv 2505.20139, 2025) proposes a comprehensive benchmark of LLMs' ability to produce structured outputs: state-of-the-art models reach 75/100 on average, and quality varies strongly with the requested format, which implies that specifying the format in the prompt is a performance lever in its own right [17]. Meaning Typed Prompting (arXiv 2410.18146, 2024) shows that a typed, structured specification of outputs improves reliability and reduces inference cost [18]. XML Prompting as Grammar-Constrained Interaction (arXiv 2509.08182, 2025) proposes a theoretical framework: XML markup acts as a grammar constraint that reduces the space of possible outputs, and therefore variance: a formal demonstration that a structured prompt is not an aesthetic whim, it is a reduction of entropy [19].

On the figures side, several sources converge, and they must be qualified to avoid over-extrapolation. Practitioners' analyses report that well-placed XML markup can reduce hallucinations by up to 40% on certain tasks [13]. A study published in 2024 in npj Digital Medicine (PMC11039454) on the interpretation of hepatology guidelines (hepatitis C) documents a jump in accuracy from 43.0% (GPT-4 Turbo alone) to 99.0% with a RAG framework combined with structured prompt engineering [20]: a strong result, but specific to this precise clinical use case; generalisation to any business task must be made with caution. On the security dimension, Anthropic published in late 2025 that Claude Opus 4.5 reduces the rate of successful prompt injections to ~1.4% in its browser-agent benchmark under new safeguards, versus ~10.8% for Claude Sonnet 4.5 under old safeguards [21]. Isolation through markup between instructions and data is one of the defences.

The honest counter-argument: when markup is not the answer

It must be said clearly: XML markup is not a magic formula applicable everywhere. Intellectual rigour requires presenting the counter-argument as it exists in the literature.

A benchmark published in May 2026 by Manish Ramavat compared, on Claude Sonnet 4.5, 150-token extraction prompts in two versions: flat prose and XML-marked-up prose. Result: the XML version costs 31% more input tokens for a negligible accuracy gap of −1.2 percentage points [22]. At 10,000 calls per day with this type of prompt, the XML overhead represents around $515/year wasted on Sonnet 4.5.

The same author's conclusion deserves to be read in full: “If your prompts are long, complex, multi-section, or handle untrusted inputs, use XML. If they are short, clear and templated, do without.” Wrapping a 10,000-token document in XML tags costs 4 additional tokens, but lets attention cleanly isolate the document from the instructions. The benefit/cost ratio therefore tips in favour of markup as soon as the context grows more complex.

Simon Willison, one of the most closely followed practitioners on these subjects, makes the same observation at a broader level. His recent study (9,649 experiments across 11 models and 4 formats: YAML, Markdown, JSON, TOON) shows that no format dominates universally, and that the model's familiarity with the format matters as much as the structure itself: the ultra-compact TOON format, little present in training corpora, paradoxically loses tokens because the model “hesitates” to follow it [23].

The operational rule that emerges from this, and which must be kept in mind, is therefore nuanced: structured markup becomes essential as soon as the task is complex, the context extensive, the sources multiple, or the inputs potentially untrusted. For simple uses (summarising an email, rephrasing a paragraph), clear prose is enough. But in business, few uses remain simple once the exploration phase is over.

What if the 2026 models had become good enough to understand natural language without XML?

This is the most legitimate objection a leader can raise, and it deserves a direct answer. Yes, Claude Opus 4.7, GPT-5.5 and Gemini 3.1 Pro understand natural language infinitely better than their 2023 predecessors. A free-prose prompt, on a simple case, will very often give a good answer on the first try.

But three forces mean that the discipline of structured instruction remains relevant, and becomes even more so with these models.

First, the more literal interpretation. OpenAI's official GPT-5 and GPT-5.5 guides say so explicitly: these models interpret prompts “literally and exhaustively”. A vague instruction will no longer be “softened” by the model; it will be executed to the letter.

Second, the stake is no longer one successful query, it is 10,000 reproducible queries. In solo exploration, a free-prose prompt works nine times out of ten. In production, over 10,000 calls per day, the 10% gap represents 1,000 non-compliant outputs per day, unacceptable for a business process.

Third, markup also structures human thinking. A team that cannot formulate the four zones “role / context / instructions / output format” cannot clearly formulate its business request either. The rigour of the format reveals the rigour of the thinking.

That said, the verdict 24 to 36 months from now will shift. If the next generation of models internalises a native understanding of vague intentions, the boundary will move. The prudent rule: for the next 18-24 months, structured markup is the standard; beyond that, to be re-assessed.

Why this changes everything for an SME in 2026

If the good practice in 2026 is no longer to “talk to AI properly” but to command it with structured instructions, several leadership decisions follow.

First, on team training. According to a survey published by a French training organisation (2026), fewer than 12% of employees have received structured training in prompting, and those who benefit from it produce results that are around 40% more accurate [24]. But the quality of the training matters more than its existence. Training that teaches how to write long polite sentences, promise rewards, make the model “play a role”, will be obsolete in six months. Training that teaches how to specify a task, structure an instruction, mark up a context, define an output format is durable.

Second, on the prompt templates used in production. Professional practice in 2026 consists of building libraries of versioned XML templates, shared across teams, tested on evaluation sets, audited at each model update. This looks far more like source-code management than like writing.

Third, on the governance of autonomous agents. The stake rises a notch as soon as AI is no longer consulted but delegated to. An autonomous agent (one that calls tools, writes files, sends emails) executes a stream of instructions composed in a chain. If the instructions are conversational, the slightest vagueness opens the door to aberrant behaviour. If they are structured and marked up, the agent stays on its rails [25].

Fourth, on the relationship with the vendor. An SME that masters the structured specification of its tasks is less captive to the model. A correctly written XML template runs, with a few adjustments, on Claude, on GPT-5, on Gemini, even on local open-source models. Portability is an underestimated strategic advantage.

Three questions to ask this very week

For an SME leader reading this article, here are the three questions that turn the observation into immediate action, to put to your IT department, your AI champion or your provider:

Do you have an inventory of the critical prompts currently used in production, and who is named as responsible for them? If the answer is “no” or “everyone”, there is a prompt-governance debt to open.
When was the last non-regression test on your AI's outputs, that is, the verification, on a standardised set of examples, that the answers remain compliant after a model or prompt update? If the answer is “never”, you are exposed to silent drift.
If you had to migrate tomorrow from Claude to GPT-5 or the other way round, how many prompts in your internal catalogue would you have to rewrite entirely? If the answer is “all of them” or “we do not know”, your portability is weak.

These three questions require neither tool nor budget. They reveal the organisation's real level of AI maturity, independently of the number of tools deployed.

What the MATIA Method™ recommends

The MATIA Method™ (the proprietary methodology of Croissance et Transitions, built around five levels of AI maturity) addresses this question within its level 3 (Orchestre): the industrialisation and operational governance of use cases. Three practices are laid down as a foundation.

1. A standardised minimum template. For any AI use in production, the prompt is written by explicitly distinguishing four zones: the role and the task, the reference context, the method instructions, the expected output format:

<role>You are a senior analyst tasked with…</role> <context> <document index="1">…</document> <document index="2">…</document> </context> <instructions>

First check the consistency between the documents
Identify the points of divergence
Propose a synthesis in … </instructions> <output_format> Answer in English, structured in three sections, with [n] citations referenced to the documents. </output_format>

This structure fits in a few dozen additional tokens. On a long prompt, it is amply amortised. On a short prompt, it is trimmed down: this is the 80/20 rule of markup.

2. A library of versioned templates. Critical prompts are stored in a versioned repository, with their test sets, their change history, their named owner. This is the discipline of application code. This library is organised around a simple structure, sometimes codified under the acronym CARE: Context, Action expected, Result, End-goal. Four creation processes coexist in business: steering by subject-matter experts (SME-Driven) for legal and financial uses; open participation (Crowdsourcing) for creative cases; AI-assisted generation then human filtering (AI-Generated) for large-scale optimisation; structuring by role (Role-Based) to standardise an entire department [28].

3. A delegation framework for agents. Junyr Agents™, the flagship product of the Junyr suite operated on junyr.app, embodies this discipline: delegation of AI agents that can be operated, triggered and audited by email via the Email Routing layer of Junyr Mail™. Each agent is defined by an XML instruction template, its tools are limited to an explicit scope, its outputs are constrained to a schema.

Conclusion: command, do not converse

The industry is turning a page. The word prompt will survive a few more years in everyday vocabulary. But in business, in 2026, professional practice is aligning on a more rigorous discipline: you do not talk to AI, you command it. With explicit instructions, marked-up contexts, specified output formats, versioned templates, and documented governance. Structured markup (XML, JSON or explicit delimiters) is the de facto standard on which the three big vendors converge.

For a French SME that wants to draw real value from its AI tools this year, the most structuring lever is not a better model. It is a discipline of instruction. The 2026 window remains open: 18 to 24 months to switch from an AI that is used to an AI that is architected. What is at stake is neither fear nor urgency; it is mastery. And mastery, as always, begins by changing the right word: here, replacing talk with command.

Frequently asked questions

Is prompt engineering really dead? No. The word survives in everyday vocabulary and in some job descriptions. But professional practice has shifted: people now speak of “context engineering”, encompassing structured markup, context management, template design and prompt governance. “Prompt engineering” in the narrow sense gives way to a discipline of orchestration.

Do you really have to use XML everywhere? No. The nuanced rule is: use structured markup (XML, JSON or explicit delimiters) as soon as the task is complex, the context extensive, the sources multiple, or the inputs potentially untrusted. For simple uses, clear prose is enough. The Ramavat benchmark of May 2026 documents that on short prompts (≈150 tokens), XML markup can be a useless overhead.

What is the difference between prompt engineering and context engineering? Prompt engineering focuses on the formulation of a given query at a given moment. Context engineering encompasses the whole filling of the context window: task description, few-shot examples, retrieval results (RAG), multimodal data, available tools, state, history.

My team is non-technical. Do they need to be trained to write XML? Not directly. You train them instead to specify a task, structure an instruction, mark up a context, define an output format. The XML templates are then carried by an AI champion, a developer or a consultant, and the teams fill them in, they do not write them each time.

Will XML markup age with the 2027 models? Probably, in part. If models internalise a greater understanding of vague intentions, the boundary between clear prose and strict markup will move. For the next 18-24 months, structured markup is the standard; beyond that, to be re-assessed. The lasting benefit is not XML in itself, it is the discipline of specification.

Does XML markup replace RAG, fine-tuning or system prompts? No. They are complementary. RAG injects the company's private data into the context; XML markup cleanly separates them from the instructions. Fine-tuning adjusts the model; markup structures the instruction. Modern system prompts are themselves structured markup in disguise.

Going further

MATIA Method™ Audit (AI Express Audit & Roadmap): a 60-minute video call to assess your current maturity level. croissance-transitions.fr
Junyr Agents™: delegation of AI agents for SMEs, operable and auditable by email. junyr.app
Junyr Mail™: eIDAS professional email. junyr-mail.com

Sources

Andrej Karpathy, X, 25 June 2025, “+1 for 'context engineering' over 'prompt engineering'”. [x.com/karpathy](https://x.com/karpathy/status/1937902205765607626)
Addy Osmani, “Context Engineering: Bringing Engineering Discipline to Prompts”, Substack, 2025. [addyo.substack.com](https://addyo.substack.com/p/context-engineering-bringing-engineering)
Gartner, Lead the Shift to Context Engineering as Prompt Engineering Fades (Report ID 6781234), 28 July 2025. gartner.com
“LLM Fundamentals, Tokens, Attention & Transformers (2026)”, MyEngineeringPath. myengineeringpath.dev
“How LLMs Work”, tutorialQ. tutorialq.com
“What is an attention mechanism?”, IBM. ibm.com
Sebastian Raschka, “A Visual Guide to Attention Variants”. [magazine.sebastianraschka.com](https://magazine.sebastianraschka.com/p/visual-attention-variants)
arXiv 2503.08980, 2025. [arxiv.org/abs/2503.08980](https://arxiv.org/pdf/2503.08980)
Grzankowski, A., arXiv 2408.04666, 2024. [arxiv.org/abs/2408.04666](https://arxiv.org/pdf/2408.04666)
So, J. et al., “Beyond Anthropomorphism: a Spectrum of Interface Metaphors for LLMs”, arXiv 2603.04613, 4 March 2026. [arxiv.org/abs/2603.04613](https://arxiv.org/abs/2603.04613)
“The Double-Edged Sword of Anthropomorphism in LLMs”, PMC. [pmc.ncbi.nlm.nih.gov](https://pmc.ncbi.nlm.nih.gov/articles/PMC7617520/)
“The $380 Million Prompt Engineering Lie”, Towards AI, 2025. [pub.towardsai.net](https://pub.towardsai.net/the-380-million-prompt-engineering-lie-why-act-like-an-expert-doesnt-boost-accuracy-0af5eb79b4ff)
Anthropic, “Use XML Tags to Structure Your Prompts”, official Claude documentation. [platform.claude.com](https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/use-xml-tags)
OpenAI, GPT-5 Prompting Guide. [developers.openai.com](https://developers.openai.com/cookbook/examples/gpt-5/gpt-5_prompting_guide)
OpenAI, Prompt Guidance. [developers.openai.com](https://developers.openai.com/api/docs/guides/prompt-guidance)
Google, Prompting Guide for Gemini API. [ai.google.dev](https://ai.google.dev/gemini-api/docs/prompting-strategies)
StructEval, arXiv 2505.20139, 2025. [arxiv.org/abs/2505.20139](https://arxiv.org/pdf/2505.20139)
Meaning Typed Prompting, arXiv 2410.18146, 2024. [arxiv.org/abs/2410.18146](https://arxiv.org/pdf/2410.18146)
Alpay F. & Alpay T., “XML Prompting as Grammar-Constrained Interaction”, arXiv 2509.08182, 9 September 2025. [arxiv.org/abs/2509.08182](https://arxiv.org/abs/2509.08182)
So J. et al., “Optimization of hepatological clinical guidelines interpretation by large language models”, npj Digital Medicine, 2024 (PMC11039454), jump 43.0% → 99.0% with RAG + structured prompt engineering on hepatitis C. [pmc.ncbi.nlm.nih.gov](https://www.ncbi.nlm.nih.gov/pmc/articles/PMC11039454/) · nature.com
Anthropic, Mitigating the risk of prompt injections in browser use, 2025, Claude Opus 4.5 brings the rate of successful injections down to ~1.4% (vs ~10.8% for Sonnet 4.5 without new safeguards). anthropic.com · pymnts.com
Manish Ramavat, “Benchmarking XML Delimiters in LLM Prompts”, May 2026. dev.to/manishramavat
Simon Willison, “Structured Context Engineering for File-Native Agentic Systems”, Feb 2026, 9,649 experiments, 11 models, 4 formats. simonwillison.net
Nerolia Formation, Prompt Engineering en français 2026. nerolia-formation.fr
Simon Willison, “New prompt injection papers”, 2025. [simonw.substack.com](https://simonw.substack.com/p/new-prompt-injection-papers-agents)
On context rot / lost in the middle, cf. arXiv 2504.02732 “Why do LLMs attend to the first token?”. [arxiv.org/abs/2504.02732](https://arxiv.org/pdf/2504.02732)
DORA 2025 Report, Google Cloud, Nearly 5,000 responses, 100 hours of interviews. [cloud.google.com](https://cloud.google.com/blog/products/ai-machine-learning/announcing-the-2025-dora-report) · faros.ai · dora.dev
CARE model and SME-Driven / Crowdsourcing / AI-Generated / Role-Based typology, Gemini Deep Research synthesis, 19 May 2026 (internal report).

Paul-Antoine TUAL is an AI Transformation Leader. He heads Croissance et Transitions (SAS) and operates the Junyr™ suite, MATIA Method™ (methodology), Junyr Agents™ (AI agents for SMEs, junyr.app), Junyr Mail™ (eIDAS email). He supports the leaders of French mid-caps and SMEs in their AI transformation, 60-minute diagnostic: croissance-transitions.fr.

Token and AI API budgets: the FinOps guide for SMEs in 2026

Mon, 18 May 2026 00:00:00 GMT

Introduction: the new economic paradigm of artificial intelligence in the enterprise

The integration of artificial intelligence into business processes has crossed a tipping point. As of May 2026, the technology landscape of small and medium-sized enterprises (SMEs) is marked by a deep structural transition: the shift from a software economy based on fixed per-user licences (SaaS) to a utility-consumption economy, dictated by an omnipresent billing unit: the token. This pricing shift has introduced new volatility into financial and technology planning.

Current statistics reveal an instructive duality. On the one hand, adoption rates have risen sharply: 88% of organisations use artificial intelligence in at least one business function (Stanford AI Index 2026), a figure the McKinsey barometer had already revised from 78% to 88% in its 2025 iterations [1]. On the other hand, this ubiquity comes with a financial reality that must be faced head-on: 43% of large AI initiatives are judged doomed to fail, for want of execution (HCLTech, May 2026) [3]. An MIT study (Project NANDA, August 2025) estimated that 95% of generative AI pilots had no measurable impact on the P&L [2]; the 2026 measurements paint a less binary picture: 11% of “AI leaders” (KPMG Global AI Pulse Q1 2026) and fewer than 10% of organisations having fully scaled AI in a function (Stanford AI Index 2026). The main cause lies not in the limitations of language models (LLMs), but in the absence of an architectural and organisational framework for managing inference costs at scale. This is, fundamentally, good news: a problem of method is fixed by method.

Companies are now observing an internal phenomenon sometimes referred to as “tokenmaxxing”: the consumption of computing power by development and operations teams is sometimes wrongly interpreted as an indicator of technological velocity. The financial consequences are concrete. Some SMEs find that spending on AI tokens has become one of their fastest-growing budget lines, sometimes overtaking the cost of the automation tasks it replaces. It is not uncommon to see a cloud infrastructure bill rise sharply. One documented case [6] mentions an autonomous agent that reached its injection cap of 150,000 characters and accumulated several hundred dollars in monthly overspend on unsupervised flows. This is what is called a “shadow budget”: AI spending that escapes financial control.

In the absence of control frameworks, consumption grows asymmetrically relative to the value generated. With the proliferation of complex requests and the emergence of autonomous multi-agent systems, inference spending in engineering departments becomes a budget line in its own right. Several field reports put it at close to 10% of staff costs on user teams, although no reference institute (IDC, Gartner) has to date validated this ratio as a consolidated average. Optimising AI costs is therefore no longer a mere matter of financial hygiene relegated to FinOps teams at the end of the quarter; it constitutes an architectural discipline in its own right.

This document sets the gold standard for optimising, distributing and governing token and API budgets for teams operating within SMEs in 2026. It details the structuring of quotas, routing gateway architectures, semantic caching strategies, the secure management of autonomous agents and frameworks for evaluating return on investment (ROI). The objective is simple: to provide a technical and financial foundation that turns a structurally inflationary technology into a lever for predictable, measurable profitability.

The regulatory framework and governance: foundations of profitability

Optimising technology budgets in 2026 is intrinsically tied to a company's ability to impose clear governance. The absolute freedom to experiment of previous years has given way to a regulated environment, where compliance guides the architecture of information systems. SMEs can no longer let each department deploy artificial intelligence models on an ad hoc basis, without centralised oversight.

The ISO/IEC 42001 standard: structuring accountability

The international standard ISO/IEC 42001:2023, dedicated to artificial intelligence management systems (AIMS), has established itself as the reference framework for structuring a responsible and financially viable use of AI [7]. Obtaining this certification (or, at a minimum, aligning rigorously with its requirements) is not a communications exercise: it is a prerequisite for budgetary control.

One of the standard's major contributions is the obligation to maintain a complete and up-to-date inventory of all artificial intelligence systems, deployed models and third-party providers used by the organisation. Without this visibility, it is impossible to attribute token consumption costs to the various profit centres. The standard requires that risk and impact assessment be carried out at the level of each specific application, and not generically at company level. This leads SMEs to link each API request flow to a designated owner, creating a direct line between the technology spend (the cost of tokens) and managerial responsibility (accountability).

Adopting ISO 42001 also helps to fill a notable decision-making gap. On one side, the Piper Sandler CIO Survey reports that 87% of CIOs expect an increase in their AI budget [4]. On the other, the Drexel LeBow / RGP work shows that only 14% of leaders say their organisation is prepared in terms of skills, and that 14% of CFOs measure a clear impact on the P&L [5]. These two studies do not overlap exactly, but their convergence points to the same reality: AI budgets are rising faster than governance maturity. Deploying an AIMS framework in line with ISO 42001 leads management committees to take ownership of consumption metrics, and turns technology spend into an auditable strategic asset.

European regulation (AI Act) and initiatives for frugal AI

On the regulatory front, the AI Act timetable has just changed. The “Digital Omnibus” political agreement reached at the European trilogue on 7 May 2026 has postponed the entry into force of binding obligations for high-risk AI systems: 2 December 2027 for standalone systems (Annex III: recruitment, credit scoring, biometrics) and 2 August 2028 for systems embedded in already-regulated products (Annex I: medical devices, industrial machinery) [8]. The transparency obligation remains set for 2 August 2026, with one exception: the machine-readable marking of generative content (Article 50(2)) benefits from a reprieve until 2 December 2026 (Digital Omnibus).

For French SMEs, this reprieve is not an invitation to stall: it is a useful window in which to structure governance (system inventory, per-application risk assessment, documented human oversight) before these requirements become binding. The penalties remain heavy on the horizon; and the real cost of being unprepared is paid first in emergency reorganisation, not in fines.

Alongside legal compliance, the concept of “frugal AI” has taken concrete form in France through practical reference frameworks. AFNOR Spec 2314 (12 July 2024), “General reference framework for frugal AI: measuring and reducing the environmental impact of AI”, sets out methodological guidelines [9]. Technological frugality aligns with budget optimisation: by minimising energy consumption (through the use of models with fewer parameters or through the reduction of superfluous API calls), SMEs mechanically lower their token bill.

Sector initiatives, such as the work carried out by Numeum on “Ethical AI” [10], reinforce this momentum. A manifesto built around three pillars (DO, COMMUNICATE, PROGRESS) and an application guide containing 117 recommendations in its 2024 edition, these tools help companies design architectures in which the accuracy of data prevails over quantity, which limits the overloading of models' context windows. AI governance, whether driven by ecology, ethics or the law, invariably leads to a rationalisation of data flows and, consequently, to the protection of the company's financial capital.

The AI gateway (LLM gateway): control infrastructure

Optimising budgets and managing quotas require an architecture capable of intercepting, analysing and directing every request sent by the SME's applications to model providers (OpenAI, Anthropic, Google, etc.). The traditional model (developers embedding API keys directly in the applications' source code) is no longer fit for purpose: it is difficult to audit and exposes the company to uncontrolled costs. The 2026 standard rests on the use of an AI gateway (LLM gateway) acting as a centralised control plane.

An AI gateway differs from a classic API gateway (REST or GraphQL) in its ability to understand the asynchronous, probabilistic and token-priced nature of generative workloads. Without this intermediary layer, companies face unexplained outages during provider incidents, uncontrolled proliferation of high-end models, and the inability to attribute costs to the different teams.

Every request passing through an enterprise gateway must be wrapped in four logical envelopes:

Identity: associating the request with a user, a team, a project or a cost centre, to enable internal chargeback.
Policy: applying rate limits, budgets, allowlists of authorised models and dynamic routing logic.
Security: inspecting in real time to filter out personally identifiable information (PII) and block prompt-injection attempts.
Observability: recording in detail the latency, the exact number of tokens consumed (input and output) and the cost of the transaction.

Comparative analysis of AI gateways in 2026

The market offers a range of solutions addressing different constraints of latency, deployment complexity and granularity of financial controls. The table below summarises the characteristics of the dominant platforms for SMEs [15].

Gateway solution	Architecture & deployment	Latency (overhead)	Cost and quota control	Use case and SME recommendation
Bifrost (Maxim AI)	Open source (Go) / fully managed	~11 µs at 5,000 RPS	Hierarchical budgets across 4 levels (organisation, team, key, user). Strict rejection (hard block) of out-of-budget requests. Millisecond-level cost analytics.	Gold standard for SMEs requiring very low latency on customer-facing applications, with enterprise-grade governance.
LiteLLM	Open source (Python)	Average under light load; P99 = 90.72 s at 500 RPS, memory crash at 1,000 RPS	Normalisation of requests across more than 100 providers. Spend tracking and strict enforcement of limits per virtual key and per project.	SMEs with platform teams able to manage the infrastructure, favouring open-source flexibility and portability; not to be exposed to high-volume real-time traffic.
Portkey	SaaS / private deployment	+65% latency compared with Kong AI Gateway	Advanced observability capturing more than 40 data points per request. Strict cost segmentation by workspace, team and user.	SME applications requiring complex firewalls, deep CI/CD integration and application-level rather than infrastructure-level management.
Braintrust Gateway	SaaS (free beta)	Average	Cost attribution via customisable tags (environment, feature). Detailed tree-structured traces (span-level).	Teams strongly focused on evaluating model quality (evals) and debugging reasoning chains.
Kong AI Gateway	Enterprise API gateway (Lua/Go)	Industry benchmark	Robust quota management and rate limiting via the existing plugin ecosystem. Enterprise security (mTLS, key rotation).	SMEs already using Kong for their traditional APIs and wishing to consolidate all traffic under a single governance.
Cloudflare AI Gateway	Edge infrastructure	Depends on the network	Real-time dashboards for token usage. Limited hierarchical budgeting capabilities, strong DDoS protection.	SMEs seeking immediate deployment and already using Cloudflare's content delivery network (CDN).

Beyond the functional comparison, the benchmarks published by gateway vendors in 2026 (to be cross-checked) [15] make one key point clear for SMEs: under real traffic, the differences in behaviour between gateways quickly become decisive. The choice of infrastructure therefore determines the company's financial resilience. Adopting a tool such as Bifrost or LiteLLM ensures that the financial safeguards run at the edge, and stop any excess request before the provider can even bill it.

Quota management by team: allocation and pragmatic enforcement

Treating tokens as an infinite resource is an architectural mistake. Token budgeting (Token Budgeting Architecture) consists of treating these units as a scarce and exhaustible resource, in the same way as random-access memory (RAM) in an operating system or processor time in a scheduler.

Structuring departmental quotas

The starting point is to set an overall budget not from the models' theoretical limits (which can accept up to 2 million tokens), but from economic projections. The architectural golden rule: an application should plan to use only 85% of its theoretical maximum envelope, with the remaining 15% serving as a safety margin to absorb estimation errors or the inevitable expansion of system messages.

The breakdown of this overall budget must be carried out precisely across the SME's teams, drawing on realistic consumption forecast models for 2026. Analysis of the workloads yields the following profiles.

Department / SME use case	Estimated task volume	Monthly consumption (tokens)	Financial impact and optimisation priority
Customer service (chatbots / support)	5,000 to 50,000 conversations / month	15 to 250 million	Very high. Near-systematic reliance on entry-level (budget-tier) models to avoid a cost explosion. The pricing gap reaches several orders of magnitude compared with flagship models.
Finance & accounting (invoices)	500 to 5,000 documents / month	1.25 to 75 million	Moderate. Structured extraction tasks. Using regular expressions or traditional OCR in pre-processing is recommended to limit the volume submitted to the LLM.
Software engineering (developers)	Intensive daily use (copilots, agents)	Hard to cap	Critical. The forecast budget per developer ranges, according to our field observations, between $1,000 and $3,000 per year in 2026 (MATIA estimate). Coding agents can consume 50,000 to 200,000 tokens per complex task.
Marketing (content generation)	Continuous stream of copy and trend analysis	Variable (high ratio of output tokens)	High. Content generation involves a high proportion of output tokens (output), billed 3 to 8 times more than input tokens [14]. Strict verbosity limits are imperative.

Enforcement mechanisms: from soft limits to hard cut-offs

Quota governance does not rest on merely watching post-billing financial dashboards. It requires pre-emptive controls implemented directly in the AI gateway, orchestrated along a rigorous, graduated scale.

Warnings and soft limits. Configured to trigger when the team reaches 70% or 80% of its daily or monthly allocation. This threshold does not disrupt end users' workflow; it triggers automated webhooks (Slack notifications, emails) that alert project managers and FinOps teams to a potentially abnormal acceleration in spending.

Conservative mode and slowdown (rate limiting). As the critical zone approaches (85% to 95% of budget), the gateway activates a throttling strategy. Requests are deliberately slowed to discourage non-essential usage. Above all, routing is altered: requests explicitly asking for access to expensive premium models are intercepted and automatically downgraded to standard models, unless the request is identified as coming from a critical process (whitelist).

Emergency mode and hard limits (hard limits & feature gating). When 100% of the quota is consumed, the gateway refuses to incur any new charges. The application undergoes a hard cut-off (hard reject) for standard requests, returning an HTTP 429 Too Many Requests code. To maintain the continuity of service perceived by users, the feature gating technique is used: advanced features are disabled in the interface, and residual basic traffic is routed exclusively to “nano” models whose inference cost is close to zero.

This hierarchical system protects the SME's gross margins from uncontrolled consumption, while preserving a controlled operational flexibility.

Dynamic model routing: maximising yield per token

One of the most common inefficiencies in enterprise AI deployment is the routine use of the most powerful (and most expensive) models to solve trivial problems. In 2026, the cost disparity between entry-level models and top-tier models is considerable. Using a flagship model to format a text or classify a customer intent is an economic aberration: the market now offers highly capable models for fractions of a cent.

A comparison of the pricing in force in May 2026 illustrates the extent of this gap [11][12][13].

Provider and model	Cost / 1M tokens (input)	Cost / 1M tokens (output)	Recommended use case for SMEs
OpenAI GPT-5 Nano	$0.05	$0.40	The champion for small budgets. Ideal for classification, simple data extraction and formatting.
DeepSeek V4-Flash	$0.14	$0.28	An ultra-economical open-weights alternative, for batch processing (batch) or high-volume pipelines.
Anthropic Claude Haiku 4.5	$1.00	$5.00	Routing of high-volume customer support flows requiring speed and consistency.
OpenAI GPT-5	$1.25	$10.00	General-purpose use cases, a balance between contextual nuance and moderate cost.
Anthropic Claude Opus 4.7	$5.00	$25.00	Flagship model. New tokenizer that can consume ~35% more tokens for the same text (higher real cost); a figure from a single secondary source, to be cross-checked. To be reserved for complex analysis and deep reasoning [12].

The gap between the cheapest model (GPT-5 Nano) and the most expensive (Claude Opus 4.7) represents a cost multiplier that exceeds 60 on output and 100 on input. Given that around 70% of a company's typical requests are basic extraction or simple question-and-answer, the absence of dynamic routing amounts to spending most of the IT budget on unused computing power.

The decision-making architecture (router logic)

Dynamic routing (Dynamic Routing) consists of inserting an algorithmic evaluation layer that intercepts the user's request, analyses it in a few milliseconds and directs it to the model offering the best cost/performance ratio for that specific task. The execution flow of a modern intelligent router follows a logical sequence:

Classification of intent and complexity. A very fast “nano” model, or a set of heuristic rules, evaluates the request: a simple rewording? reading a long context? a complex mathematical problem?
Tier selection (tiering). The request is assigned to a capability tier. The modern SME deploys its models as a portfolio: the vast majority of traffic is directed to the core layer (the low-cost models).
Quality check and fallback. If the small model's response has too low a confidence score, the gateway organises a transparent escalation to a higher model. This safety net ensures that the quality perceived by the user does not deteriorate, while achieving substantial savings on the bulk of requests handled on the first attempt.

Implementing this strategy translates into a portfolio approach: a large volume of requests routed to the cheapest models, a medium fraction to standard models, and a narrow reserve to elite models. Field reports and vendor comparisons cite API bill reductions ranging from 40% to 85% with such an architecture, without perceived quality degradation, provided the escalation confidence thresholds are calibrated correctly.

The reasoning-token caveat (thinking tokens)

2026 has seen the widespread adoption of so-called “reasoning” models (Reasoning Models), which simulate an internal chain of thought before formulating their answer. They are remarkably effective at solving software problems or complex mathematical logic.

This architecture, however, introduces an important point of caution for budget management. The “thinking tokens” generated during the internal cognitive process, although often hidden from the end user, are billed at the output-token rate (output tokens) [14]: depending on the provider, a price 3 to 8 times higher than that of input tokens.

As a result, a seemingly trivial request that triggers a prolonged reasoning loop can consume between 500 and 5,000 invisible tokens. To model correctly the budget of an SME using these advanced models, finance departments must apply a safety multiplier of 3 to 5 times the cost usually estimated for standard responses. This is why dynamic routing must formally isolate access to these reasoning models, prohibiting it for routine requests and front-line conversational agents.

Context engineering and compression: maximising the signal-to-noise ratio

Cost optimisation also involves reducing the volume of data ingested by the models. In a Transformer architecture, processing cost and latency scale quadratically with the size of the context window: doubling the amount of text provided multiplies the required computing power by approximately four. Filling this window with irrelevant documents or verbose instructions is not only costly: it also degrades the accuracy of the responses (the lost-in-the-middle phenomenon).

Traditional prompt engineering has given way to context engineering. The key skill in 2026 is no longer to craft a fine sentence, but to design the informational ecosystem in which the model operates, filtering out the noise. SMEs would do well to establish strict formatting rules.

Verbosity constraints and structured format. The most immediate technique for curbing output costs is to systematically require concise or formatted responses. Replacing long textual descriptions with instructions such as “Provide the answer as a Markdown table” or “Limit the answer to 50 words” directly reduces the most expensive part of the API bill. Likewise, using clear XML tags (<context>, <instructions>) allows the model to isolate variables quickly without the need for long explanatory sentences.

Algorithmic prompt compression (LLMLingua). Retrieval-augmented generation (RAG) systems inject large volumes of document fragments into the context window. To avoid token inflation, programmatic tools such as LLMLingua [16] are deployed. These algorithms, which rely on small language models (SLMs), compute the perplexity of each word and remove non-essential terms (stop words, syntactic flourishes) while preserving the semantic integrity of the information. Microsoft Research benchmarks report compression rates of up to 20x with limited performance loss, and 4x savings at a compression rate of 5x, reducing in typical cases an 800-token prompt to around 160, with minimal quality degradation.

Dynamic management via reinforcement learning (ContextBudget). At the frontier of optimisation in 2026, new frameworks such as “ContextBudget” and its BACM-RL method [17] treat memory management as a sequential decision problem subject to explicit budget constraints. Instead of relying on arbitrary chunking heuristics, the system dynamically learns to compress the conversation history as it progresses, thereby avoiding capacity overflows (overflow) while maximising the retention of critical information.

The discipline imposed by context compression is fundamental. By treating the context window as a virtual bank account where every word deposited costs a few cents, software architects learn to prioritise essential data and eliminate waste at the source.

Semantic caching: the most effective saving lever

While compression reduces the unit cost of a request, caching eliminates the need to query the model altogether. In enterprise environments, a massive share of traffic is inherently redundant: users continually ask the same technical support questions, request the same summaries of HR policies, or generate reports based on identical data.

Traditional caching (Exact Match) relies on the exact comparison of character strings or their hash (SHA-256). Its limitation is well known: a tiny variation in punctuation or wording (“What is the delivery time?” vs “When will I receive my parcel?”) invalidates the cache and triggers a fresh, full call to the API. On the natural language of real users, the hit rate remains modest.

Semantic caching resolves this inefficiency by understanding the intent behind the request. It is the optimisation with the most immediate return on investment for an SME.

The three-layer architecture

The robust implementation of a semantic cache (often hosted at the AI gateway level or via in-memory databases such as Redis) is orchestrated along a defensive three-layer architecture:

Exact match. Fast and free. The incoming prompt is normalised (whitespace removal, lower-casing), hashed, then compared. In the event of a perfect match, the response is served in less than a millisecond.
Semantic similarity (Semantic Cache). If the first layer fails, the system calls on a lightweight, inexpensive embedding model to convert the sentence into a multidimensional mathematical vector. This vector is compared with the requests previously stored in a vector database. By computing the distance between vectors (cosine similarity), the system assesses the closeness of meaning; if the score exceeds a rigorous confidence threshold (for example 0.95), the stored response is reused.
Recourse to the LLM (LLM Fallback). Only when the first two barriers have been crossed is a paid call triggered to the large model's API. The new response is then vectorised and stored to enrich the future cache.

Financial impact and lifecycle management

The metrics observed in production justify the integration effort. The principle is validated by the documentation of the main gateways: by intercepting redundant requests, API load and perceived latency are markedly reduced. The orders of magnitude often cited (API cost reductions of the order of 45% to 86% and latency improvements of around 88%) do not yet have a consolidated academic study as a reference; they serve as an indicative range to be validated on one's own scope. On the measured-cost side: vector computation adds a marginal overhead of around 20 milliseconds, negligible compared with the 850+ milliseconds of an avoided LLM call.

Characteristic	Traditional cache (Exact Match)	Semantic cache (Vector Similarity)
Matching method	Strict string comparison (hashing)	Vector distance (cosine similarity) reflecting meaning
Handling of rewordings	Systematic failure (cache miss)	Success if the similarity threshold is reached
Required infrastructure	Simple key-value store (e.g. Memcached)	Vector database + embedding model
Hit rate (hit rate)	Low on natural language (sensitive to wording variations)	High, but varies greatly with the recurrence of traffic (to be measured for one's own case)
Latency reduction	Instantaneous (< 1 ms)	Strong (minimal computation overhead ~20 ms, largely offset by the gain)

Semantic caching carries one caveat: the staleness of information (staleness). Serving a cached response about a financial procedure that was changed the day before poses a real reliability problem. The gold standard therefore requires meticulous management of the time to live (TTL, Time To Live) of cache entries. Highly volatile data (prices, stock levels) must have a short TTL (a few minutes); structural information (FAQs, product documentation) can persist for several days. Event-based invalidation mechanisms (event-based invalidation) must purge the cache as soon as the source database is updated.

Finally, API providers now offer server-side prompt caching solutions (Provider-Side Prompt Caching). This feature is particularly valuable for long system messages or static RAG contexts of more than 1,000 tokens: Anthropic advertises discounts of up to 90% for repeated access to the same prefix; on the DeepSeek side, a cache hit is billed at around $0.014 for an initial cost of $0.14, that is, the same discount of around 90% [18]. Combining the local semantic cache with the provider-side prompt cache forms the most robust financial shield against cost inflation.

Mastering agentic systems: circuit breakers (kill switches)

2026 is the year of “agentic” AI. Models no longer merely generate text in response to an isolated prompt: they are embedded in autonomous workflows where they plan, use tools (web browsing, code execution) and delegate tasks among themselves, in multi-agent systems built with frameworks such as LangGraph, CrewAI or AutoGen [19]. While this evolution sharply increases productivity, it also introduces new financial and security risks that must be kept in check.

The risk of infinite loops (infinite retry loops)

Agentic autonomy alters the cost dynamics: billing is no longer linear, it becomes quadratic. With each iteration of an agent trying to correct an error, the complete history of its previous actions must be reinjected into the context window to maintain the coherence of its reasoning. An agent stuck on a task, and persisting in solving it, therefore consumes more and more tokens with each attempt.

Silent failures exist and are documented [6]. An agent programmed to analyse a code base or validate invoices, which encounters a transient API error, can enter an infinite retry loop (infinite retry loop). If it runs at night, unsupervised, it can generate thousands of useless API calls and accumulate several hundred dollars in monthly overspend on the environment concerned. The remedy is not fear: it is architecture.

Containment architecture: three levels of circuit breaker

Preventing these incidents does not rest on improving prompts, but on a containment architecture operating below the application layer. Implementing “circuit breakers” (kill switches) and firewalls is a necessity. A resilient architecture is built around three blocking layers.

The budget and threshold circuit breaker (Quota Guard Pattern). Built into the heart of the AI gateway, this circuit breaker monitors the telemetry stream in real time. It imposes an absolute, non-negotiable ceiling on the number of iterations allowed per session (for example, forced stop after 3 unsuccessful attempts) or on the amount spent (for example, a cut-off at $5 for the current task). Beyond these thresholds, the gateway blocks communication with the LLM API, freezes the agent's state and requires the intervention of a human supervisor (Human-in-the-loop, HITL).

Cryptographic identity isolation (Identity Gate Revocation). In mature production environments, each autonomous agent is given a unique cryptographic identity (for example, SPIFFE certificates). When aberrant behaviour is detected (data leakage, excessive loops, unauthorised access attempts), the security system does not merely refuse requests: it revokes the agent's certificate. This cryptographic cut-off is absolute: the agent loses its mutual authentication capability (mTLS), its requests to the models are rejected, its access to internal databases lapses, and the other agents refuse to communicate with it.

Time-based tool confinement (Sandbox & Data Plane Gates). The principle of least privilege must govern access to external tools (reading emails, writing to a database). The architecture prohibits perpetual access: if an agent must audit a customer file, the system issues it a strictly time-limited authorisation token (timeboxed consent), for example for 60 minutes, and confined to a specific resource. Once the deadline expires, the data plane gate closes. Thus, even in the event of a hallucination or a malicious prompt injection, the potential damage is contained in space (restricted access) and in time (rapid expiry).

Thanks to these architectural barriers, an SME ensures that error, inevitable in any probabilistic system, remains contained, with no major financial or security consequence. The infrastructure protects the application from its own failures.

Evaluating return on investment (ROI) and FinOps practices

Governance, gateways, dynamic routing and semantic caching are the tools of profitability. But to sustain the funding of these initiatives, SME finance departments (CFOs) expect quantified evidence of their impact. The debate no longer concerns the theoretical capabilities of the technology, but the return on the capital deployed.

Although 88% of organisations use AI in at least one function (Stanford AI Index 2026) [1], financial validation remains demanding: the CIO Playbook 2026 (IDC-Lenovo) measures that only 46% of POCs reach production (even though 94% of organisations anticipate a positive ROI) and fewer than 20% of initiatives manage to scale to enterprise level. This gap is explained by a poor grasp of the total cost of ownership (TCO) and by the difficulty of monetising productivity gains.

Calculating the total cost of ownership (TCO)

Financial modelling of generative AI systems differs from that of traditional software. Classic SaaS licences had fixed, predictable costs; AI generates variable costs tied to the computing intensity of each interaction. Finance departments must analyse AI through the lens of cost of goods sold (CoGS, Cost of Goods Sold) or as a variable operating expense (OpEx).

The classic return-on-investment formula applies, provided the variables are defined rigorously:

ROI (%) = [ Net benefit (Gains − TCO) / Total cost of ownership (TCO) ] × 100 The most common mistake SMEs make is to equate the TCO with only the price billed by the provider's API (input and output tokens). The true cost (Fully Loaded Cost) is structurally broader and must include:

Data preparation and processing: data engineering, cleaning, structuring and vectorisation (embeddings). The Snowflake/ANZ surveys [20] identify this item as the leading operational blocker (lack of data diversity: 56%; lack of preparation: 59%) and estimate that it regularly accounts for 10% to 20% of the total budget, or even the majority of unexpected costs (vector databases, pipelines).
Infrastructure and orchestration: hosting the AI gateway, log storage, observability tools, server costs.
Technical integration and quality assurance: connector development, time spent by subject-matter experts (SME, Subject Matter Experts) annotating and evaluating the quality of responses (evals), and continuous prompt tuning.
Change management: training employees to ensure adoption of the tools, which often absorbs 10% to 30% of the project's overall budget, with training costs per employee ranging from $3,000 to $20,000 according to Snowflake feedback [20].
Governance and compliance: monitoring AI Act-related risks and maintaining IT security.

Quantifying the benefits: from intangible to financial

On the numerator side, measuring the benefits must go beyond superficial metrics of “employee satisfaction”. To justify the investment, time savings must be converted into financial value.

The most rigorous method is to monetise the hours saved: the time gained on a task is multiplied by the employee's loaded hourly cost (base salary plus 25% to 40% for social charges and benefits). For example, if automating a customer-returns classification process allows a team of 10 people to save 1,300 hours per year at a loaded cost of $87/hour, the gross productivity gain amounts to $113,100. Adding the reduction of manual errors and the decrease in rework, the financial value generated can easily multiply from the very first year of operation.

SMEs must also incorporate the notion of cost avoidance. If deploying a customer-support agent routed to economical models absorbs a 20% increase in the volume of incoming requests without additional hiring, the ROI includes the total cost of the salaries that did not need to be paid to support the growth.

The following table lists the key performance indicators (KPIs) useful for assessing budgetary impact by operational area.

Impact (department)	Financial gains (net benefits)	Operational indicators (KPIs)	Target time to value (TTV)
Finance (FP&A) & operations	Reduction in operating costs, increased operating leverage.	Hours saved / week (e.g. 2 to 4 h/employee), reduction in forecasting cycle time (−30%).	Fast (< 6 months) thanks to structured flows.
Customer service & support	Avoided recruitment costs, reduced customer attrition (churn).	Autonomous resolution rate (containment rate), reduction in average response time, increase in CSAT / NPS.	3 to 6 months. High-volume economical models excel here.
Security & compliance	Lower compliance costs, fewer high-risk errors.	Number of false positives, fraud detection speed, rate of unresolved incidents.	3 to 6 months.
Sales & marketing	Revenue growth, higher customer lifetime value (LTV).	Conversion rate, average order value (AOV), return on ad spend (MER) targeted at 5.0x.	Short to medium term. Requires monitoring of generation costs.

The adoption strategy to secure ROI

Securing ROI in an SME requires methodological prudence. The “shiny object syndrome”, which drives the adoption of AI for every problem, must be set aside. The gold standard recommends initially targeting only a single use case (Single Use Case), characterised by strong potential impact, low risk and internal data that is already structured and of good quality.

It is also prudent to apply a safety discount to projections. While industry reports and solution vendors cite productivity gains of 30% to 50%, a conservative finance department will reduce these estimates by 30% to 50% in its business case, to account for adoption frictions and the performance gaps specific to the real contexts of SMEs. By framing expectations and limiting pilot projects to a horizon of 2 to 4 weeks, companies ensure that their investments translate into measurable cash rather than costly laboratory experiments.

Conclusion: engineering profitability in the age of AI

2026 marks a break in the corporate technology ecosystem. Access to the most capable artificial intelligence models has become commonplace, which erases the competitive advantage tied to merely owning the technology. The real differentiator between SMEs now lies in the ability to master the underlying economic architecture of these systems. Intelligence has become an abundant commodity; it is profitable intelligence that is scarce.

The gold standard for optimising budgets and tokens is not improvised: it is architected. It begins with a solid governance framework, aligned with demanding reference standards such as ISO/IEC 42001, which turns experimentation into accountable, auditable processes. It takes shape through the deployment of AI gateways (LLM gateways), the true backbone of financial control: these proxies ensure that every fraction of a cent spent is identified, budgeted and subject to clear consumption limits.

Cost control then rests on precise execution strategies. Dynamic request routing demonstrates that a vast majority of tasks can be accomplished by low-cost models without sacrificing quality. Context engineering and semantic caching eliminate waste at the source, converting linguistic redundancies into economies of scale. Faced with the growing autonomy of agentic systems, operational resilience is ensured by the integration of circuit breakers (kill switches) and cryptographic firewalls, which protect the organisation against technical and financial runaways.

The reprieve offered by the “Digital Omnibus” on the AI Act, until 2027-2028, is not a respite: it is a window of opportunity for the SMEs that want to move out of experimentation and into architecture. Those that adopt these principles, rigorously measure their total cost of ownership and demand a tangible, documented return on investment equip themselves with a resilient infrastructure. They turn artificial intelligence, by nature unpredictable and costly, into a lever for lasting operational efficiency, and thereby cement their competitiveness in tomorrow's digital economy. This is precisely the logic of the MATIA Method™: structuring the method before piling up tools, and making the control of AI costs a discipline of architecture, not an emergency reaction.

Going further

The AI Express Audit & Roadmap (60 minutes by video call, no commitment) includes a review of your AI cost architecture: gateway, routing, caching and per-team quotas, with a quantified recommendation.

The white paper “AI Maturity of French SMEs 2025-2026” is available at croissance-transitions.fr.

Sources: verifiable, May 2026

[1] McKinsey & Company, The state of AI: How organizations are rewiring to capture value (and State of AI 2025/2026), late 2024 / 2025. URL: https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai-how-organizations-are-rewiring-to-capture-value; “78% of respondents say their organizations use AI in at least one business function” (revised to 88% in 2025; a figure taken up by the Stanford AI Index 2026).

[2] MIT (Project NANDA), The GenAI Divide: State of AI in Business 2025, August 2025. URL: https://mlq.ai/media/quarterly_decks/v0.1_State_of_AI_in_Business_2025_Report.pdf; “95% of enterprise AI pilots deliver zero measurable return on the P&L” (study of 300 deployments). The 2026 measurements paint a less binary picture: 11% of “AI leaders” (KPMG Global AI Pulse Q1 2026) and fewer than 10% of organisations having fully scaled AI in a function (Stanford AI Index 2026).

[3] HCLTech, survey of 467 decision-makers, May 2026. URL: https://www.hcltech.com; 43% of AI initiatives judged doomed to fail (execution gap).

[4] Piper Sandler, CIO Survey 2025/2026. URL: https://www.pipersandler.com/sites/default/files/document/cio_survey_sample.pdf; “87% [of CIOs are] expecting budget increases” for AI.

[5] Drexel LeBow / RGP, State of Data Integrity & Foundational Divide, 2025-2026., “14% of leaders responded that their organization is not prepared with the skills”; “only 14% of CFOs report clear, measurable impact”. Note: the “87% / 14%” combination is a conflation of two separate studies (see [4]).

[6] Niko Feith (Medium), The token tax: who pays when AI agents run in loops, 2026. URL: https://medium.com/@niko.feith/the-token-tax-who-pays-when-ai-agents-run-in-loops-59adef9eee1b; “total injection cap is 150,000 characters… hundreds of dollars per month in API costs… burns tokens on failed retry loops” (OpenClaw agent).

[7] ISO / ISMS.online, ISO/IEC 42001:2023, Artificial Intelligence Management System, December 2023. URL: https://www.isms.online/iso-42001/; AIMS requirements: “Conduct comprehensive AI risk assessments, AI impact assessments, Implement Ethical AI Practices”.

[8] European Commission / Modulos, Digital Omnibus Deal / AI Act FAQ, May 2026. URL: https://www.modulos.ai/blog/eu-ai-act-omnibus-deal/; high-risk obligations postponed to 2 December 2027 (standalone Annex III) and 2 August 2028 (Annex I products). The political agreement was reached on 7 May 2026.

[9] AFNOR, AFNOR Spec 2314, Référentiel général pour l'IA frugale : mesurer et réduire l'impact environnemental de l'IA, 12 July 2024. URL: https://www.afnor.org/en/news/artificial-intelligence/reference-framework-reduce-environmental-impact-ai/

[10] Numeum, Ethical AI Manifesto + Guides, 2021-2024. URL: https://ai-ethical.com/home/ ; https://ai-ethical.com/en/manifesto/; three pillars DO / COMMUNICATE / PROGRESS; 2024 edition of the guide = 117 recommendations.

[11] OpenAI / DevTk, OpenAI API Pricing Guide 2026, May 2026. URL: https://devtk.ai/en/blog/openai-api-pricing-guide-2026; GPT-5 Nano: $0.05 / $0.40 per million tokens; GPT-5: $1.25 / $10.00.

[12] Anthropic / Metacto, Anthropic API Pricing: A Full Breakdown, May 2026. URL: https://www.metacto.com/blogs/anthropic-api-pricing-a-full-breakdown-of-costs-and-integration; Claude Opus 4.7: $5.00 / $25.00 with a new tokenizer that can consume ~35% more tokens for the same text (a figure from a single secondary source, to be cross-checked); Claude Haiku 4.5: $1.00 / $5.00.

[13] DeepSeek / TLDL, DeepSeek API Pricing 2026, May 2026. URL: https://www.tldl.io/resources/deepseek-api-pricing; DeepSeek V4-Flash: $0.14 / $0.28.

[14] Anthropic Docs / Metacto, Extended thinking tokens billing, May 2026. URL: https://www.metacto.com/blogs/anthropic-api-pricing-a-full-breakdown-of-costs-and-integration; “Extended thinking tokens are billed as output tokens… charged at the standard output rate”; Input/Output ratio of 3 to 8x depending on the model (Opus 5x, R1 ~4x).

[15] Varshith V. Hegde (Dev.to), Top 5 LLM Gateways in 2026: A Deep Dive Comparison for Production Teams, 2026. URL: https://dev.to/varshithvhegde/top-5-llm-gateways-in-2026-a-deep-dive-comparison-for-production-teams-34d2; Bifrost: < 11 µs overhead at 5,000 RPS; LiteLLM: P99 = 90.72 s at 500 RPS, memory crash at 1,000 RPS; Portkey: +65% latency vs Kong.

[16] Microsoft Research, LLMLingua: Compressing Prompts for Accelerated Inference of Large Language Models, 2023-2025. URL: https://llmlingua.com/llmlingua.html ; arXiv: https://arxiv.org/html/2310.05736v2; “up to 20x compression with little performance loss” and “4x savings at a prompt compression rate of 5x”.

[17] Independent researchers (arXiv), ContextBudget: Budget-Aware Context Management, BACM-RL, April 2026. URL: https://arxiv.org/abs/2604.01664; “BACM-RL, an end-to-end curriculum-based reinforcement learning approach that learns compression strategies under varying context budgets”.

[18] Anthropic / DeepSeek (Finout synthesis), Provider-side prompt caching, May 2026. URL: https://www.finout.io/blog/claude-opus-4.7-pricing-the-real-cost-story-behind-the-unchanged-price-tag; Anthropic: “up to 90% savings with prompt caching”; DeepSeek cache hit ~$0.014 for an initial cost of $0.14 (≈ −90%).

[19] Radixia AI, Designing proactive AI agents. URL: https://blog.radixia.ai/designing-proactive-ai-agents/; agent frameworks (AutoGen, etc.) and design patterns.

[20] Snowflake / Scoop, Snowflake research ANZ: More organisations investing heavily in Gen AI than the global average, 2024-2025. URL: https://www.scoop.co.nz/stories/BU2504/S00311/snowflake-research-reveals-more-anz-organisations-investing-heavily-in-gen-ai-than-the-global-average.htm; lack of data diversity: 56%; lack of preparation: 59%; training costs per employee: $3,000 to $20,000; unexpected cost drift: 30% to 50% of budget.

Article written by Paul-Antoine TUAL, AI Transformation Leader, creator of the MATIA Method™.

The “All-Cloud” is dead: 5 risks that put strategic on-premise back on the table in 2026

Wed, 13 May 2026 00:00:00 GMT

More than 85% of the European cloud market is captured by American players (EU Perspectives, February 2026). In May 2026, five risks are converging simultaneously: the geopolitical fragmentation of the global network, a haemorrhage of intellectual property through AI coding tools, the first offensive AI capable of creating zero-days (GTIG report, 11 May 2026), SaaS inflation disconnected from the value delivered, and the ANSSI post-quantum deadline as early as 2027. The economic calculation has tipped. Here are the facts, and the quantified ROI of the alternative.

1. Geopolitical risks and digital “blackouts”: the Splinternet is coming

The concept of the “Splinternet”, an internet fragmented into incompatible geopolitical blocs, is no longer a researcher's hypothesis. It is the terrain on which French SMEs operate in 2026.

The dependency is structural. More than 85% of the European cloud market is captured by the American hyperscalers, led by AWS, Azure and Google Cloud (EU Perspectives, February 2026). All are subject to the 2018 CLOUD Act. On a US judicial order (18 U.S.C. §2713), the authorities can demand access to data hosted by an American company, even if the servers are physically in France. The great majority of French companies are in this situation.

The cut-offs are documented, not theoretical. In early 2025, the Trump administration brandished the threat of cutting Ukraine's access to Starlink; Microsoft blocked the International Criminal Court prosecutor's account on the basis of US sanctions (May 2025); Maxar Technologies suspended Ukraine's access to its satellite imagery (March 2025). The “kill switch” mechanism is no longer an abstraction. The Ifri puts it bluntly: Europe is “at Washington's mercy” in the digital domain.

The instability is getting worse. The repeated Outlook and Microsoft 365 outages in 2025-2026 (some lasting several hours and simultaneously affecting thousands of European companies) illustrate what “outsourcing your critical infrastructure” concretely means: when Microsoft goes down, you go down with it. With no leverage, no alternative, no SLA to compensate for the hour of lost productivity. And the Asahi affair (a return to pen and paper after a ransomware attack, October 2025) is a reminder, more broadly, of what an unavailable critical infrastructure costs.

BGP routing anomalies, incidents on transatlantic submarine cables, and climate events (the 2025 CME solar storm) expose an “illusion of sovereignty”: your data at rest is in France, but your data in transit passes through nodes beyond your control. The WEF Outlook 2026 ranks dependence on critical digital infrastructure among the world's top ten systemic risks.

The transatlantic trade tensions of 2026 and Microsoft's unilateral price rise (+5.2 to +16.6% depending on the plan from 1 July 2026, see section 4) materialise the economic risk. Geopolitical dependency is also pricing dependency.

2. Hidden dependencies: the silent leak of your intellectual property

In late April 2026, GitHub updated the terms of service for Copilot's non-Enterprise versions: exchanges with the AI may be used to improve Microsoft/OpenAI models. For an SME without an Enterprise licence (several hundred euros per developer per year), your source code, your software architecture and your internal comments may feed the training data of the Big Tech giants.

Comparison table: what each tool sends (May 2026)

Tool	Data sent	Privacy mode	Jurisdiction	CLOUD Act
GitHub Copilot (non-Enterprise)	Code + model training since April 2026	Opt-out not enabled by default	USA	Yes
GitHub Copilot (Enterprise)	Context code only	Contractually guaranteed	USA	Yes
Cursor	Context code + full session	Privacy Mode OFF by default	USA	Yes
Claude Code (Anthropic)	Prompts stored for 30 days by default	Configurable	USA	Yes
Local LLM (Ollama)	No outbound data	Total by definition	Your servers	Full sovereignty

This table is not an anecdote. This is intellectual property (pricing algorithms, business logic, the architecture of your systems) crossing the Atlantic with every keystroke.

Emerging risks: slopsquatting and prompt injection

Slopsquatting: LLMs hallucinate non-existent package names in 5% to 22% of code suggestions, according to studies published in 2025. Attackers register these names with malicious code. Your developer installs a package “recommended by the AI”: a Trojan horse.

IDE Prompt Injection: malicious code in your dependencies can inject instructions into your AI coding assistant, which then executes unauthorised actions (credential exfiltration, silent code modification). This is an attack vector documented in 2026 that specifically exploits this channel.

3. Offensive AI bots: the automated zero-day has arrived

11 May 2026: the Google Threat Intelligence Group (GTIG) publishes a landmark report. For the first documented time, an AI designed a functional zero-day exploit from start to finish, capable of bypassing a two-factor authentication (2FA) system, without human intervention. The exploit leveraged a semantic logic flaw: not a memory bug, but a behavioural inconsistency in the protocol's logic. The Python markers in the code confirm the AI origin.

The Google Threat Intelligence Group and Mandiant document, in their February 2026 reports, the active use of LLMs by state offensive groups:

UNC2814 (China): AI-driven vulnerability analysis and exploit generation
APT45 (North Korea): LLM-driven automation of spear-phishing campaigns
CANFAIL/LONGSTREAM (Russia): AI for identifying attack vectors in source code

The XBOW system (June 2025) had already demonstrated that an AI bot can submit hundreds of zero-day vulnerability reports to bug bounty programmes, without human intervention.

The operational conclusion: your code exposed on the public cloud is scrutinised around the clock by automated systems capable of creating their own exploits. Obscuring your architecture behind private servers mechanically reduces this attack surface.

4. The unbeatable ROI of on-premise: three layers to bring back in-house

4.1 Microsoft 365 / Exchange: time to take stock

Microsoft is applying a price increase from July 2026:

Plan	Current price	July 2026 price	Increase
Microsoft 365 Business Basic	€6.00/month/user	€7.00/month/user	+16.6%
Microsoft 365 Business Standard	€12.50/month/user	€13.80/month/user	+10.4%
Microsoft 365 Business Premium	€22.00/month/user	€24.00/month/user	+9.1%
Microsoft 365 E3	€36.00/month/user	€38.00/month/user	+5.6%
Microsoft 365 E5	€57.50/month/user	€60.50/month/user	+5.2%

These increases include AI features (Copilot for Microsoft 365) that are often unsolicited. For a 50-person SME on Business Standard, that is +€780/year for features nobody asked for.

Microsoft Exchange 2016 and 2019 reached their End-of-Life on 14 October 2025. No more security patches. SMEs that migrate to Microsoft 365 face these increases. There is a third way.

4.2 Local LLM: up to ~90% lower inference costs

Compact open-source models, from 7 to 30 billion parameters (Gemma 4 12B, Mistral Small, Qwen, distilled DeepSeek variants), cover a large share of common business sub-tasks (summarisation, classification, extraction). Available on Ollama (169,000 ⭐ on GitHub), with no token cost.

The maths: frontier APIs = $3 to $15 per million tokens at list price (May 2026), while the average real spend measured by Ramp in June 2026 falls to $0.72 per million. A sign the market is already shifting to lighter models. Local LLM = €0 per token, cost amortised on the hardware. On a hybrid deployment (local LLM for volume, cloud API for complex cases), routing to compact models can cut inference costs by up to ~90% (InfoWorld, May 2026). Break-even: a few months on our typical deployments (MATIA estimate).

The result: zero data exposure, zero contractual dependency, and easier GDPR compliance. Local hosting does not remove the other obligations, but it does remove the transfer question.

4.3 On-premise AI agents and a sovereign email server

Cloud iPaaS costs reach $48,000 to $180,000/year for mid-sized companies (ranges observed on vendors' public pricing grids; orders of magnitude). n8n, LangChain and Ollama enable entirely local agent architectures, auditable, with no external API dependency.

This is the Junyr Agents™ model (junyr.app): delegation of AI agents within business processes (HR, CRM, accounting, projects, invoicing), operated on-premise, triggerable by email via Junyr Mail™. Auditable, reversible, with no cloud dependency.

For email: a sovereign solution hosted in France, eIDAS-compliant, costs less than €10/month per domain, outside the CLOUD Act, outside Microsoft outages, outside unilateral price increases. Junyr Mail™ (junyr-mail.com): €9.90/month, OVH France, European legal standing.

5. Post-quantum cryptography: the urgency of crypto-agility

The SNDL attack: “Store Now, Decrypt Later”

The reasoning is simple: an attacker intercepts your encrypted data today and stores it. When Q-Day arrives (2035 horizon according to the coordinated European roadmap of June 2025), a quantum computer will break RSA-2048 and ECC-256 in a matter of hours. Your strategic data from 2026 will be readable in 2035.

This is the SNDL (Store Now, Decrypt Later) attack. It is considered already under way by Western cybersecurity agencies, which is precisely what justifies the migration timetables below.

The recommended post-quantum algorithms (NIST standards)

Algorithm	Use	NIST standard
CRYSTALS-Kyber (ML-KEM)	Key exchange (KEM)	FIPS 203
CRYSTALS-Dilithium (ML-DSA)	Digital signature	FIPS 204
Falcon (FN-DSA)	Compact digital signature	FIPS 206 (draft in progress)
SPHINCS+ (SLH-DSA)	Stateless signature (hash-based)	FIPS 205

The timetables (not to be confused):

2027: no more ANSSI-qualified products without hybrid post-quantum cryptography (ANSSI)
2030: procurement of security products incorporating PQC (ANSSI); migration of high-risk use cases (EU roadmap)
2035: intermediate use cases (coordinated European roadmap, June 2025)

In late 2025, Thales and Samsung received the first ANSSI approvals incorporating PQC algorithms. The compliance window is open, but it is closing.

On shared cloud infrastructure, you depend on your provider's migration timeline. On on-premise infrastructure, you control the quality of your random number generators (HSM), you manage the IKEv2 fragmentation induced by the new post-quantum keys (CRYSTALS-Kyber), and you migrate on your own schedule. That is the definition of crypto-agility.

Conclusion: five risks, one calculation

Risk	Status	Deadline
Geopolitical shutdown / cloud outages (CLOUD Act, kill switch)	Already triggered	Immediate
IP leak via AI coding tools (non-Enterprise Copilot)	Effective since April 2026	Immediate
Offensive AI bots / automated zero-day (GTIG)	Documented 11 May 2026	Immediate
Uncontrolled SaaS inflation (Microsoft 365 +5.2 to +16.6%)	Announced	July 2026
ANSSI post-quantum obligation	Timeline set	2027 (qualification)

The question is no longer “can my SME afford a sovereign infrastructure?” It is “can it afford not to have made the call?”

The ROI is calculable and favourable: hybrid on-premise LLM (up to ~90% lower inference costs by routing to compact models, according to InfoWorld, May 2026; break-even in a few months, MATIA estimate), sovereign email (less than €10/month), local AI agents (elimination of iPaaS costs of $48,000-180,000/year), post-quantum compliance (a competitive advantage as early as 2027).

The MATIA Method™ integrates the sovereignty axis from maturity level 3, not as an on-premise dogma, but as a considered trade-off: classified data, SecNumCloud-type sovereign cloud where it suffices, the local option where it is justified. An AI deployed without legal and technical control of its infrastructure remains a risk disguised as a tool.

Paul-Antoine TUAL · AI Transformation Leader Founder of Croissance & Transitions and the MATIA Method™ croissance-transitions.fr | junyr.fr | junyr.app (Junyr Agents™) | junyr-mail.com (Junyr Mail™)

Sources

GTIG/Google report, 11 May 2026
ANSSI cyber.gouv.fr
CERT-FR-2026-CTI-001
GTIG / Mandiant, AI threat reports (February 2026)
Ifri
EU Perspectives, European cloud market (February 2026)
WEF Outlook 2026
NIST FIPS 203/204/205/206
GitHub Copilot data policy, April 2026
Microsoft 365 pricing, July 2026
Microsoft Exchange EOL, 14 October 2025
KYP.ai supply chain security 2026
Ollama GitHub.

From win-win to ethical hard bargaining: why I had to change my negotiation method

Wed, 05 Nov 2025 00:00:00 GMT

For years I preached the cooperative method and trained colleagues in constructive negotiation: separate the person from the problem, focus on interests, build trust. In short, win-win. I am convinced of it: it is the most powerful approach for long-term relationships. But the ground has shifted. And I had to adapt to survive. The reality is that, in a socio-economic context where the inertia of systems (notably the slowness of the justice system) devalues the threat of legal proceedings, the other party is tempted to fill that vacuum with immediate coercion or a total absence of response. The legal threat is remote. Faced with this rise of hard bargaining (tough negotiation), my purely cooperative approach has become, in certain cases, naive.

The Impossible Equation: Firmness Without Destruction

Adopting pure aggression (threats, unjustified extreme anchoring) destroys relational capital and leads to costly deadlocks. A third way must be found: I propose Considered Firmness. I have made the transition to ethical hard bargaining. A method that borrows the tools of the power struggle, but anchors them in rationality, objectivity and respect for principles. Here are the 3 pillars that now guide my strategy:

1. Anchoring Justified by the Reassessed BATNA

The classic hard bargainer uses extreme anchoring (Door-in-the-Face) to throw you off balance. My anchoring, for its part, is unshakeable because it rests on a total BATNA (Best Alternative to a Negotiated Agreement). My BATNA is no longer only the direct cost of the failure of the negotiation. I factor in the opportunity cost linked to time lost, reputational risk and the deterioration of the climate. My initial demand is high, but it reflects a factual assessment of my total risk, and not a mere bluff.

2. Factual and Transparent Pressure

I reject manipulation and personal insults. If I must apply pressure (the ethical equivalent of a public threat), I hold to the transparency of the facts. Pressure consists in setting out the verifiable consequences of a non-agreement, based on impartial data. I do not say: “You are incompetent.” I say: “Your 60-day delay, proven by the performance reports, triggers a contractual penalty of X, a cost that we must factor in right now if we do not find a mutual solution.” I use power to bring people back to reason, not to bring them to their knees.

3. The Conditional Way Out

Unlike the “take-it-or-leave-it” tactic, my commitment to firmness is always conditional. I say clearly: “My position is firm as long as market conditions do not change” or “We are ready to resume dialogue as soon as deliverable X is completed.” This makes it possible to defuse escalation while maintaining firmness on substance. It is a signal that I am seeking agreement, but only a fair and profitable agreement.

The Bottom Line

Ethical hard bargaining is not a betrayal of the win-win principles. It is its armed version, suited to high-intensity conflicts and low-trust environments.

It makes it possible to be demanding on substance without being destructive on form. It is the only way to preserve credibility and to ensure that the signed agreements will be not only advantageous, but lasting. And you? Have you observed this surge of hard tactics in your negotiations? How do you manage to stay firm while protecting your professional relationships?

#Negotiation #Management #Leadership #ConflictResolution #HardBargaining #BATNA

Paul-Antoine TUAL · AI Transformation Leader · Croissance & Transitions

Alternative Dispute Resolution (ADR) in the Service of Organisational Performance

Mon, 01 Sep 2025 00:00:00 GMT

Alternative dispute resolution (ADR) settles internal and external disputes more quickly, fostering a context of “peace” conducive to the sound development of organisations.

Did you know that internal conflict costs French companies the equivalent of one month of lost work per year for a large majority of employees? This figure, drawn from a 2021 study, highlights a major strategic issue: the speed and quality of dispute resolution are no longer an option, but the very heart of organisational performance.

When it comes to relationships with their partners, organisations must likewise maintain a context of peace in order to secure their development, since every dispute adds to uncertainty and makes decision-making difficult. The search for a swift amicable solution is therefore a priority for the external disputes of organisations as well

Mediation and conciliation emerge as the essential levers for turning these hidden costs into gains in productivity, engagement and competitiveness. Far from being mere legal tools, these amicable dispute-resolution methods (ADR) draw on law, sociology and management to create a serene context for development.

Here is how a structured approach to mediation and conciliation drives your organisation's performance.

1. The Legal Framework: Accelerating Resolution to Control Costs (Law)

The French legal system has strongly encouraged amicable settlement as the preferred route, thereby acknowledging the need to ease congestion in the courts and to offer organisations faster and less destructive solutions.

Since 1 January 2020, a prior attempt at amicable resolution (mediation or conciliation) has been made mandatory for certain disputes, on pain of the legal claim being ruled inadmissible.

The impact on performance is measured in time saved and legal certainty:

Reduced timeframes: Mediation and conciliation offer markedly shorter resolution windows than litigation, allowing managers and teams to refocus quickly on productive objectives.
Securing agreements: Mediation agreements relating to the employment contract may be approved by the employment tribunal (Conseil des prud'hommes), giving them enforceable status on a par with a judgment. Moreover, resorting to these amicable procedures suspends the limitation period, safeguarding the right to bring legal action should the process fail.

The speed of amicable settlement prevents escalation and limits the financial and human costs associated with lengthy proceedings.

2. The Sociology of Organisations: Restoring Trust and Engagement

From a sociological standpoint, conflict is the expression of the interplay of actors within a system, going beyond the façade of unanimity that organisations so often seek. Mediation intervenes here as a powerful tool for restoring cohesion and corporate culture.

The mechanism for improving performance runs through engagement:

Perceived Organisational Support (OST): When the organisation invests in mediation to resolve disputes in a fair and humane manner, employees perceive strong Organisational Support. This socio-emotional support positively reinforces the employer-employee relationship.
Restoring Engagement: This sense of recognition and worth leads to a heightened level of employee engagement, as employees become more motivated to commit to collective success. This engagement acts as a key mediator between calmer working relationships and greater organisational performance.
Constructive culture: Mediation restores deep communication, fosters listening and mutual understanding, and enables the parties to arrive at lasting solutions on their own. This consolidates a positive corporate culture, essential to the continuous improvement of results.

By turning crisis into a learning opportunity, mediation strengthens the company's social capital.

Further reading:

L'analyse de pratiques en médiation : Méthodes, outils et réflexions (2020) by Carine Bernardi.
Conflits au travail : Passer de la crise à l'opportunité en 4 étapes (2020) by Jean-François Thiriet.
Mediation Effect of Employee Engagement on the Relationship Between Employee Relations and Organizational Performance (2024), Journal of Business Management Review.

3. Managerial and Strategic Lever: Measuring Return on Investment (Management)

Strategic management regards mediation as an investment, not an expense. It is a tool for managing risk (disputes, resignations, sick leave) and for creating value (performance, innovation).

Calculating the Return on Investment (ROI):

Avoided costs: One of the main values of mediation lies in avoiding the hidden costs of conflict, which include the loss of managerial time, absenteeism, reduced quality of work, and the cost of any potential litigation. An amicable approach bears no comparison with the costs of adversarial confrontation or protracted disputes.
Relational gains: Beyond direct savings, mediation makes it possible to restore or improve informal relationships between partners (business-to-business) or staff (internal), stimulating creativity and fostering lasting relational contracts.
Mediating management: The development of “managerial mediation” is a strategic issue. Managers trained to regulate tensions and to adopt the posture of a mediator establish an enduring climate of trust, essential to collective performance.

Mediation is an instrument for the early detection of, and support for, companies in relational or financial difficulty.

Further reading:

Press release from the cost-of-conflict observatory (2021), highlighting the financial losses caused by conflict.
Report from the Médiateur national du crédit and the Médiateur des entreprises (April 2025), on the early detection of difficulties.
La médiation managériale : Un levier stratégique pour la résolution des conflits en entreprise (2025).

Conclusion: Mediation, a Strategic Imperative

Embedding mediation and conciliation within organisations must become a first-order reflex; it means adopting the stance of a mature, high-performing company. It means choosing speed of process, durability of solutions, and improved human engagement.

It is a skill that should now be built into every managerial function.

By making amicable settlement a factor of internal and external peace and of cohesion, your organisation does not merely resolve its conflicts; it transcends them, turning them into the source of its resilience and future success.

Paul-Antoine TUAL · AI Transformation Leader · Croissance & Transitions

AI in micro-businesses: a revolution similar to 1980s computing?

Tue, 19 Nov 2024 00:00:00 GMT

How to implement AI in a micro-business: a 4-step strategy for a successful transition, based on the experience of adopting computing in the 1980s.

Artificial intelligence (AI) is often perceived as a technology reserved for large companies, but it can bring considerable benefits to very small businesses (micro-businesses) as well. The key to successful adoption lies in a gradual and well-structured approach. Here is a 4-step strategy for integrating AI into a micro-business, together with examples of concrete applications and suggested tools.

The adoption of computing in the 1980s and the implementation of AI today share many similarities in terms of progression and the reluctance shown by companies. Here is how a parallel can be drawn between these two technological revolutions through the four-step AI adoption strategy.

Introduction: one-off computing vs. one-off AI

1980s - One-off computing

In the 1980s, computing was first introduced into companies on a one-off basis, often for specific tasks. For example, the first computers were used to replace typewriters in secretarial departments or for precise financial calculations in accounting departments.

Today - One-off AI

Similarly, AI is today used for specific, one-off tasks. For example, companies begin by using AI to draft content, create automatic replies, or generate marketing ideas. This first approach makes it possible to understand how AI can ease simple tasks without disrupting the organisation.

Regular use: systematic computing vs. systematic AI

1980s - Systematic computing

Once computing had proven its usefulness, it became more systematic. Management software such as spreadsheets (e.g. Lotus 1-2-3) and rudimentary databases began to be integrated into the daily routines of companies, automating processes such as inventory management or accounting.

Today - Systematic AI

For AI, this stage is equivalent to creating routines in which teams use AI on a regular basis. For example, chatbots automate interactions with customers, or text-analysis tools are used to extract key information from reports. Companies recognise the usefulness of AI and begin to give it a more central role.

Optimising secondary processes: No-Code in the 1980s vs. No-Code AI today

1980s - No-Code computing

In the 1980s and then the 1990s, more accessible interfaces and “No-Code” software (even if the term did not yet exist) began to emerge, allowing non-programmers to create simple applications. Software such as dBase, then Microsoft Access (1992), made it possible to build databases without coding, and users were able to automate tasks without going through the IT departments.

Today - No-Code AI

Today, the same logic applies with AI. Tools such as Zapier, Make, or Airtable make it possible to automate secondary processes without writing any code. These tools embed AI to handle tasks such as creating workflows or analysing customer data, and even teams without technical skills can benefit from the advantages of AI.

Total transformation: complete integration of computing vs. complete integration of AI

1980s-90s - Complete integration of computing

Over time, computing became integrated into the key processes of every company, revolutionising entire sectors. ERP (Enterprise Resource Planning) software automated the management of company resources, while the internet transformed communication and business in the following years.

Today - Complete integration of AI

We are entering an era in which AI is becoming integrated into the core processes of companies. This can include optimising the supply chain, forecasting sales, or personalising customer experiences through predictive-analytics systems. AI is becoming a genuine strategic asset, just as computing was in the past, making companies more competitive and efficient.

Conclusion

Just as computing transformed the business world in several stages, AI is following a similar path. The key for companies is not to rush, but to adopt these new technologies in a gradual and considered way. The analogy clearly shows that, just as it was essential to invest in computing in the 1980s to remain competitive, integrating AI today is becoming a necessity for companies of all sizes.

Paul-Antoine TUAL · AI Transformation Leader · Croissance & Transitions

Supporting the internal promotion of a sales manager in a communications agency

Sun, 16 Jun 2024 00:00:00 GMT

Promoting a sales manager within a communications agency is an important step that requires structured support to ensure their success in this new role. Here is a detailed plan to support this promotion:

1. Preparation before the promotion

Skills assessment: Identify the sales manager's current skills and those required for the new position.
Training plan: Put in place a training programme to close the gaps identified. This may include training in leadership, project management and sales strategy.

2. Announcing the promotion

Internal communication: Inform the entire team of the promotion officially, via an email, a team meeting or an announcement at an internal event. Highlight the sales manager's achievements and the reasons for their promotion.
External communication: Announce the promotion on the company's social media, the website and in the newsletters intended for clients and partners.

3. Onboarding and initial support

Mentoring: Assign a mentor or coach who can guide the new manager in their new responsibilities.
Onboarding plan: Create an onboarding plan for the first 30, 60 and 90 days to help the sales manager adjust to their new duties. This should include clear objectives and regular checkpoints.

4. Defining responsibilities and objectives

Role clarity: Clearly define the sales manager's new responsibilities and ensure they are well understood both by them and by their team.
SMART objectives: Set specific, measurable, achievable, realistic and time-bound (SMART) objectives for the new role.

5. Ongoing support and assessment

Regular feedback: Organise regular meetings to discuss progress, the challenges encountered and opportunities for improvement.
Performance reviews: Put in place quarterly reviews to measure the sales manager's performance against the objectives set.

6. Ongoing development

Continuing training: Offer continuing training opportunities to develop the sales manager's skills further, such as workshops, seminars and online courses.
Attending conferences and events: Encourage attendance at conferences, trade shows and networking events to stay up to date with industry trends and build a professional network.

7. Strengthening the company culture

Alignment with the vision and values: Ensure that the sales manager understands and embodies the agency's values and vision.
Team engagement: Encourage the sales manager to foster a collaborative and motivating work environment for their team.

8. Recognition and rewards

Recognising successes: Celebrate the sales manager's successes and achievements to sustain their motivation and commitment.
Incentives: Put in place incentives such as bonuses, future promotions or career development opportunities to reward exceptional performance.

9. 360° feedback

360-degree review: Gather feedback from various stakeholders, including direct reports, colleagues and superiors, to provide a complete perspective on the sales manager's performance and identify areas for improvement.

By following this structured plan, the communications agency can ensure that the sales manager is well prepared, supported and motivated to succeed in their new role, thereby contributing to the company's growth and success.

Paul-Antoine TUAL · AI Transformation Leader · Croissance & Transitions

Implementing artificial intelligence to grow a communications agency

Sun, 09 Jun 2024 00:00:00 GMT

Implementing artificial intelligence (AI) in a communications company opens up a wealth of opportunities to improve efficiency, personalisation, and innovation. Here are a few key areas where AI can help:

1. Automating repetitive tasks

Chatbots and virtual assistants: using chatbots to handle client requests, answer frequently asked questions, and provide 24/7 support.
Content management: automating social media publishing, comment moderation, and editorial calendar management.

2. Data analysis and insights

Sentiment analysis: using AI to analyse the sentiment expressed on social media and other platforms, to understand client and consumer opinions.
Audience segmentation: analysing client data to build more precise, more personalised audience segments.

3. Content creation

Text generation: using AI tools to create articles, social media posts, video scripts, and more.
Content personalisation: adapting content in real time based on users' preferences and behaviour.

4. Targeted advertising

Campaign optimisation: using machine-learning algorithms to optimise advertising campaigns in real time, maximising return on investment.
Personalised recommendations: offering product or service recommendations based on users' history and preferences.

5. Improving the client relationship

Intelligent CRM: integrating AI into customer relationship management systems to provide proactive support, anticipate client needs, and personalise interactions.
Feedback and analysis: automated collection and analysis of client feedback to improve services and products.

6. Optimising internal processes

Human resources management: automating recruitment, talent management, and employee performance analysis.
Forecasting and planning: using AI to predict market trends and plan communication strategies accordingly.

7. Innovation and new service development

Immersive experiences: creating augmented reality (AR) and virtual reality (VR) content for innovative marketing campaigns.
AI-based product design: developing new AI-based products or services to deliver added value to clients.

Implementation and practical advice

To implement AI effectively in a communications company, here are a few steps and tips:

Needs assessment: identify the areas where AI can bring the most value, based on the company's strategic objectives.
Technology choice: select the AI tools and technologies suited to the needs identified.
Training and skills development: train internal teams so they can use and manage the new AI technologies.
Piloting and experimentation: start with pilot projects to test the effectiveness of AI solutions before rolling them out at scale.
Partnerships and collaboration: work with AI experts or startups to benefit from their expertise and speed up implementation.
Ethics and regulation: make sure AI use complies with ethical and regulatory standards, particularly around personal data protection.

By integrating AI strategically, a communications company can not only improve its operational efficiency but also offer richer, more personalised client experiences, setting itself apart from the competition.

Paul-Antoine TUAL · AI Transformation Leader · Croissance & Transitions

Recruiting Generation Z: innovative strategies to attract tomorrow's talent

Sun, 02 Jun 2024 00:00:00 GMT

Generation Z, made up of individuals born between 1997 and 2012, represents a new wave of talent entering the labour market. Having grown up with technology at their fingertips, these young professionals bring a unique perspective and advanced digital skills. Recruiting Generation Z, however, requires a distinct approach, tailored to their expectations and values. In this article, we will explore effective strategies to attract and engage this promising generation.

Understanding Generation Z

To recruit Generation Z effectively, it is crucial to understand their characteristics and what they are looking for in a career:

Tech-savvy: They are extremely comfortable with digital technologies and social media.
Social commitment: They place great importance on ethics, diversity and corporate social responsibility.
Flexibility and work-life balance: They prefer flexible working environments that allow a good balance between professional and personal life.
Learning and development: They look for opportunities for ongoing personal and professional development.

Strategies to Attract Generation Z

1. Optimising Social Media

Generation Z spends a large part of its time on social media. To attract them, it is essential to have an active and authentic presence on these platforms.

Use the right platforms: Concentrate your efforts on Instagram, TikTok and LinkedIn, where this generation is particularly active.
Authentic content: Publish content that reflects your company culture, your values and your social commitments.
Interactive engagement: Use stories, live videos and interactive posts to build an authentic relationship with potential candidates.

2. Transparency and Company Values

Generation Z is drawn to transparent and ethical companies.

Communicating your values: Highlight your values, your sustainability initiatives and your social responsibility actions.
Transparency in the recruitment process: Clearly explain the recruitment process, the selection criteria and the development opportunities.

3. An Exceptional Candidate Experience

Creating a positive candidate experience is essential to attract Generation Z.

Streamlined process: Simplify the application process with short forms and mobile application options.
Fast feedback: Provide rapid, constructive feedback at each stage of the process.
Personalised engagement: Personalise interactions with candidates, for example with welcome videos or direct messages.

Strategies to Recruit Generation Z

1. Using Technology

Technology plays a key role in recruiting Generation Z.

Modern recruitment platforms: Use intuitive recruitment platforms that make applying and communicating easier.
Artificial intelligence: Integrate AI tools for CV screening and skills analysis to speed up the recruitment process.

2. Internship and Apprenticeship Programmes

Internships and apprenticeship programmes are effective ways to attract Generation Z.

Internship opportunities: Offer well-structured internships that allow young talent to discover your company and gain experience.
Apprenticeship programmes: Offer apprenticeship programmes that combine theoretical training with hands-on practice in the field.

3. Company Culture and Working Environment

Creating a company culture that appeals to Generation Z is essential.

Flexibility: Offer flexible working options, such as remote working and flexible hours.
Collaborative working environment: Foster a collaborative and inclusive working environment.
Personal development: Put in place professional development programmes and mentoring opportunities.

Conclusion

Recruiting Generation Z requires a thorough understanding of their expectations and an adaptation of recruitment strategies. By emphasising the use of social media, transparency and a positive candidate experience, and by integrating modern technologies, companies can attract and recruit this new generation of talent effectively. By investing in their development and creating a flexible, collaborative working environment, companies can not only recruit Generation Z but also retain and motivate them over the long term.

By following these strategies, your company will be well positioned to attract and recruit Generation Z, ensuring a prosperous and innovative future.

Paul-Antoine TUAL · AI Transformation Leader · Croissance & Transitions