Decide with your data, not your habits: 360° analysis now costs next to nothing

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

Where this sits. Third instalment in our “AI Agents” series, after engineering agentic systems and LLM context optimisation. This article is written for the leader of an SME or mid-cap who decides every week with two tables and an intuition. It documents, prices in hand, the collapse in the cost of analysing one’s own data, what near-instant scenarios change about the quality and the serenity of decisions, and what this near-zero cost displaces: verification, the state of the data, the responsibility of deciding. Prices and leaderboards were read on 18 August 2026; every survey cited is dated in the text and in the sources.

---

1. One leadership decision, two eras

Take an ordinary executive committee decision: raise prices by 4% on 1 January, discontinue a fading product line, hire a second key-account salesperson. In 2019, that dossier had three possible paths. The leader’s intuition, fed by two tables from the management controller. The business-intelligence project, with integrator, data warehouse and dashboards, billed in tens of thousands of euros. Or the study entrusted to a consultancy, in thousands of euros and in weeks. Most SMEs stuck to the first path, for lack of budget for the other two: decisions were made on gut feeling, with whatever figures were to hand.

In 2026, the same dossier is built in a morning. An agent connected to the ERP and the CRM pulls three years of sales, recomputes margins by customer and by product line once logistics costs are reallocated, isolates seasonality, then returns three quantified projection scenarios, with their assumptions written out in full. Marginal cost: a few euros of compute on a subscription already paid for, with a turnaround measured in minutes or hours.

Leaders are already drawing the consequences. In the IBM Institute for Business Value survey published in May 2026 (2,000 CEOs and executives interviewed between February and April, 33 geographies), 64% say they are comfortable making major strategic decisions based on AI-generated analysis [9]. The SME owner’s gut feeling remains an asset; it stops being alone. 360° analysis and scenarios, yesterday reserved for those who could pay for them, become a consumable of decision-making. This article documents the shift, then what it demands in return: verification, data in working order, and a leader who decides.

2. What an analysis used to cost, and what it costs now

Let us lay out the figures of both eras. On the traditional side, the 2026 price grid of a French data consultancy puts an SME’s foundational project (pipeline, modelling, multi-source centralisation) between €10,000 and €30,000, and from €30,000 to €60,000 and beyond once AI is integrated; its recent client examples: €18,000 and six weeks for the data foundation of a 40-employee company, €12,000 for an 80-site restaurant chain [1]. Then comes the running cost: €300 to €1,000 per month of data stack for a small structure, €1,500 to €5,000 per month for a structured SME [1]. The human line item follows the same scale: 2026 salary surveys place the data analyst at around €45,000 gross per year at the national median, €35,000 to €42,000 at junior level [2]. A one-off outsourced analysis bills at €8,000 to €12,000 per study [3], and access to off-the-shelf analysts remains a large-company product: an entry-level Gartner subscription negotiates between $25,000 and $60,000 per year, an advisory hour between $3,000 and $7,000 [4].

On the other side, the prices of August 2026, read on the vendors’ official pages. Microsoft 365 Copilot: €26.00 excl. VAT per user per month [5]. Claude Pro: $20 per month, $17 on an annual commitment, deep research included [6]. OpenAI slotted in a $100-per-month tier in April between its $20 and $200 offerings [8]. And at the unit level, Google’s official documentation prices a full task of its Deep Research agent between $1 and $3, with most research completing in under twenty minutes [7]. Between yesterday’s study and today’s agent task, the gap reaches three orders of magnitude (see Figure 1).

The shift hangs on the word “agentic”. The chatbot answered questions; the agent performs work. The Claude Cowork product page documents office use cases that were, yesterday, full analyst days: spreadsheet reconciliation, period-on-period financial analysis with variance flagging, the example instruction given being “Flag any line where the variance is over 5% or over $50k” [6]. On the connection side, the open MCP protocol has established itself as the standard socket between assistants and internal systems, and the data platforms (BigQuery, Snowflake, Databricks) now sell agents that converse directly with the warehouse [6][7]. The “next to” in “next to nothing” remains honest: the subscription is paid for, so is compute at high volume (we have published a token budget guide), and getting the data into shape has a real cost, to which we return. The quote for services, though, is gone.

3. Projection scenarios alongside intuition

What the agent produces for a few euros does not look like a dashboard. A dashboard answers the questions asked at the time of the BI project; a decision asks new ones. The form useful to a leader is the scenario dossier: three quantified trajectories built from the company’s real data, assumptions written out and open to challenge, sensitivity analysis (what must be true for scenario B to beat scenario A), and the premortem, that classic technique of projecting oneself into failure to list its causes before signing [27]. The management press documents the same acceleration at group level: MIT Sloan Management Review described in late 2025 two scenario-planning exercises assisted by generative AI (at Fazer and at Unum), where the classic route means “waiting six months or more for useful insights”; no quantified gain is claimed, which we note honestly [20].

This dossier supports the gut feeling, never replaces it. An SME owner’s intuition encodes years of ground truth; it is what senses that a product line is fading or that a customer is preparing to leave, and it is what formulates the starting hypothesis. What decision psychology has documented since Kahneman and Tversky is that this same intuition, left alone, carries systematic biases: confirmation (we look for what validates), anchoring (the first figure heard frames all the others), overconfidence, attachment to the status quo [26]. Until now, countering those biases cost precisely the price of a study: it was almost never paid for, and the bias won by default. The decision then fell back on habits (“we have always done it this way”) or on the generic management recipe, the fashionable matrix applied as is. Those shortcuts were substitution heuristics, used when the specific analysis cost too much. Their economic justification has just disappeared: the counter-examination of an intuition now costs three minutes of formulation and a few euros of compute. Before a committing decision, the rational reflex becomes to ask first for the analysis of one’s own data, and to call on the recipe only afterwards, if any need remains.

The serenity gained is a benefit in its own right. Deciding on a dossier, with explicit assumptions and a quantified cost of turning back, weighs less than arbitrating on instinct under total uncertainty; our measurements and sources on this point are detailed in the article on mental load, whose conclusion applies here: AI’s most reproducible gain is in perceived effort, more than on the clock. The French adoption figures show where the gap lies: in the Bpifrance Le Lab barometer of mid-caps published in June 2026 (534 responses, fieldwork from March to May), 77% of leaders report that their teams or they themselves use generative AI, but only 17% observe concrete time savings [10]. The two figures do not contradict each other: one measures the spread of a tool, the other the existence of a method. Rewriting e-mails does not change a decision; connecting an agent to your margins does.

4. The real price: verification

The counterpart can be measured too. On the official leaderboards read on 18 August 2026, the best natural-language querying systems plateau at 81.95% execution accuracy on the BIRD benchmark (real databases, analysis questions), where human reference performance is measured at 92.96%; the best 2026 submissions (Ant Group in June, Xiaomi in July) crowd between 80.8 and 81.7% without breaking that ceiling [11]. The Spider 2.0 benchmark, built on real enterprise workflows, sharpens the picture: 96.7% success on targeted Snowflake queries, 76.2% on its lighter track, and 65.6% on full analytics workflows, the track that most resembles an analyst’s actual work [12]. Performance depends on the playing field (see Figure 2): on the task closest to the job, one failure in three. A piquant detail noted in January by a team from the University of Illinois: the answer keys of these benchmarks themselves contain annotation errors; even the exam gets things wrong [13].

Long-running research agents call for the same caution. DeepResearch Bench II, published in January 2026 (132 tasks, 9,430 evaluation criteria drawn from more than 400 hours of human expertise), measures that the best models satisfy fewer than 50% of the criteria [14]; a late-January study on hallucination propagation found no agent robust end to end and describes two recurring machine biases, temporal anchoring and source homogeneity [15]. AI therefore has its biases, as intuition has its own; they do not overlap, and that is precisely what makes the pairing effective, provided there is a referee.

A lived example of what control catches. In April, press write-ups headlined that BCG “reports 25% of revenue from AI”. The firm’s official release says two different things: AI- and technology-centred services account for more than 40% of 2025 revenue, and AI services are growing at 25% per year [22]. Both figures are true; the headline that merges them is false. It is the kind of error a hurried agent copies with aplomb, and that a ten-minute check catches. Hence the rules to impose on any agent-produced dossier: require the queries and sources with every result, run a second engine independently on the same question, recount a sample by hand, reconcile with the accounts. The general ledger does not lie. Control is the new cost.

5. The remaining bottleneck: the state of your data

An agent analyses what it can reach, and it only reaches what is in working order. The 2026 surveys converge on this bottleneck. At Cloudera and Harvard Business Review Analytic Services (published March 2026, fieldwork October 2025 among more than 230 data decision-makers), only 7% of organisations judge their data “completely ready” for AI, and 56% cite silos as the first obstacle [16]. The Nasuni survey of May 2026 (1,000 decision-makers, fieldwork in March) measures the other face: 97% of organisations are deploying or piloting agents, 57% of AI projects fail to meet their objectives, and 46% report that their AI initiatives mainly revealed data quality and governance problems they did not know they had [17]. The CIO barometer published by Efalia on 11 August 2026 (methodology not detailed on the page, which we flag) drives the point home for France: 85% of CIOs state they do not control the quality of their document estate, and 62% devote less than 1% of the IT budget to it [18].

The real investment project therefore becomes the agent-ready information system: clean reference data (one customer equals one identifier), data reachable through APIs or connectors rather than locked in manual exports, defined access rights (the sales agent does not read salaries), history preserved. This project still costs money, the ranges from the start of this article [1] still apply, with one difference that changes the calculation: it compounds. Yesterday’s BI project bought answers to frozen questions; getting the data into shape buys the capacity to build the case for every subsequent decision at near-zero marginal cost. Confidentiality belongs to the same project: your margins, salaries and cash position do not travel through just any cloud. The April 2026 French inter-inspectorate report on AI in the public sector, analysed by the specialist press in July, notes up to 40% unregulated AI use among local government staff and conversational assistants rarely certified under SecNumCloud [25]; private companies have the symmetrical problem, and the same answers: enterprise offerings with contractual guarantees, minimal access scopes, controlled hosting for sensitive data. Our five risks of the all-cloud and the eight rules of agentic systems apply to analysis agents as to any other.

6. The risk of the average strategy

One trap remains, the subtlest: free analysis is the same for everyone. A team led by an Esade researcher measured it for Harvard Business Review in March 2026: seven large models put through 15,000 simulations covering seven classic strategic tensions (differentiate or play volume, automate or augment, short term or long term). On six tensions out of seven, the models converge on the socially fashionable option of the moment, whatever company context is supplied, and better prompting corrects almost nothing [19]. The authors have a word for the product of this convergence: “trendslop”. If your three competitors ask the same model “what strategy for our market?”, they receive the same plan three times, with different logos. Yesterday’s generic management recipe returns in industrialised form, served with greater confidence.

The remedy lies in the nature of the problem: the average strategy comes out of average public data. The same model, connected to your company’s data (ten years of customer history, real margins, conversion rates by segment), returns an analysis nobody else can obtain, for the simple reason that nobody else has that data. This is the direct extension of the end of the model race: when intelligence becomes a commodity, the moat migrates to the data and the processes. Your data cannot be copied.

The consulting sector, the historical seller of analysis, understood this before its clients. The Syntec Conseil barometer published in June 2026 measures a 1.5% decline in strategy consulting in 2025, the sixth-worst year since 2002, and notes that clients now use consultancies’ AI usage as a negotiation argument [21]. BCG already generates more than 40% of its revenue from AI- and technology-centred services [22]; McKinsey has candidates work a case study with its internal assistant Lilli, assessing prompt quality and critical thinking in the face of the outputs [23]. When the sellers of analysis reorganise their recruitment around verification and judgement, it is because collection and synthesis, on their own, will not sell for much longer.

7. The leader’s protocol

Concretely, to go from this article to the executive committee, five moves suffice, in this order:

• Write the decision and its turning-back criterion. One sentence: “Should we do X before Y, given Z”. Then the reversibility criterion: what would make us say no, and how much turning back costs. Your intuition sets the hypothesis; that is its place, and decision-gated development makes it the first guardrail.
• Connect the agent to the systems. A connector with minimal access rights to the ERP, the CRM or the warehouse, rather than three-week-old manual exports that freeze figures already wrong.
• Require scenarios, their assumptions and their refutation. Three quantified trajectories, written assumptions, sensitivity analysis, and the premortem of the preferred scenario: what kills it, with what probability, detectable through which signal.
• Control before believing. The rules of section 4: queries and sources required, a second independent engine, a hand-recounted sample, accounting reconciliation. Any unexplained gap is an alert, not a detail.
• Decide in session, with a cool head. Group the trade-offs into a dedicated decision session, as we detail in the personal effectiveness method, and keep the human signature on everything that commits the company.

The temptation to go one step further, letting the agent execute the decision itself, already has its cautionary tale. Klarna had entrusted its customer support to an agent (2.3 million conversations in the first month, resolution time down from 11 to 2 minutes) and cut its staff from 5,000 to 3,500 people; Forbes reports that its CEO Sebastian Siemiatkowski then acknowledged the company had “cut too far, too quickly” and lost valuable human expertise, and it rehired around a hundred specialists [24]. The 2,000 CEOs of the IBM study draw the boundary themselves: they anticipate that 48% of codifiable operational decisions will be taken by AI without human intervention by 2030, against 25% today, bounded operational work being delegated, the committing decision being kept [9]. Figure 3 sums up the full circuit and what each step now costs.

The price has moved

Let us recap what the sections establish. 360° analysis and projection scenarios have gone from thousands of euros and weeks to a few euros and a few hours [1][5][6][7]. This abundance supports the leader’s intuition and counterweighs its biases, provided the new price is paid, and it is a price of attention: verifying the outputs [11][12], getting the data into shape [16][17], and refusing the average strategy that models produce when left alone with public data [19]. The researchers behind the seven-tensions study conclude, and the sentence deserves a place on the executive committee wall: “Leadership is ultimately about making hard choices in conditions of uncertainty and taking responsibility for them. AI cannot and should not be a substitute” [19].

The competitive advantage is therefore no longer access to analysis: it is having your own data in working order, a verification loop that holds, and the courage to decide fast. That is the “Orchestre” then “Architecte” rung of our method: tooled, governed, measured processes, where agents such as Junyr’s reach what they need, and nothing but what they need. The opening move fits in a week: pick a real decision from the next committee, build the case twice, your usual circuit on one side, an agent connected to your data on the other, then compare the two dossiers, the euros and the hours. Analysis now costs next to nothing. Being wrong still costs as much as ever.

---

Sources

[1] Fenxi, Combien coûte un projet data en 2026 : prix et délais (fenxi.fr, published 30 March 2026, updated 10 August 2026): foundational SME project €10,000 to €30,000, advanced project with AI €30,000 to €60,000 and above; client examples (€18,000 and 6 weeks for 40 employees, €12,000 for 80 sites); recurring data stack from €300 to €1,000/month (small structure) to €1,500 to €5,000/month (structured SME).

[2] Data analyst salary surveys, France 2026, convergent secondary sources (Glassdoor and specialist schools, consulted 18 August 2026): national median around €45,000 gross per year, €35,000 to €42,000 at junior level, more in Paris. Orders of magnitude not verified against a single primary source, cited as reference points.

[3] IntoTheMinds, Combien coûte une étude de marché ? (published 26 February 2025, updated 12 April 2026): most projects between €8,000 and €12,000; competitive analysis €5,000 to €10,000.

[4] Vendr, Gartner and Forrester pricing pages (updates dated February 2026, SaaS purchasing platform, intermediary reading): Gartner subscription $25,000 to $60,000 per year for targeted access, $100,000 to more than $500,000 for extended access, advisory hours $3,000 to $7,000; Forrester, observed median $65,755 per year.

[5] Microsoft, French pricing page for Microsoft 365 Copilot, Enterprise plan (read 18 August 2026): €26.00 excl. VAT per user per month on annual billing, in addition to an eligible Microsoft 365 licence; the Business plan is billed lower.

[6] Anthropic, claude.com/pricing and claude.com/product/cowork (read 18 August 2026): Claude Pro $20 per month ($17 per month on annual commitment), Max from $100, Team plans; documented Claude Cowork use cases: spreadsheet reconciliation, financial analysis with variance flagging (“Flag any line where the variance is over 5% or over $50k”), Microsoft 365, Google Drive, Slack and Amplitude connectors.

[7] Google, official Gemini API documentation, Deep Research agent (ai.google.dev, read 18 August 2026) and official Deep Research Max blog post (21 April 2026): 60 minutes maximum research time, most tasks under 20 minutes; estimated cost of $1 to $3 per task ($3 to $7 for the Max version), preview pricing.

[8] TechCrunch and MacRumors (9 April 2026): launch of a $100-per-month ChatGPT tier, slotted between the $20 and $200 plans.

[9] IBM Institute for Business Value, CEO study (newsroom.ibm.com, 4 May 2026; 2,000 CEOs and executives, 33 geographies, 21 sectors, fieldwork February to April 2026): 64% comfortable making major strategic decisions based on AI-generated analysis; 48% of codifiable operational decisions anticipated to be taken by AI without human intervention by 2030, against 25% today.

[10] Bpifrance Le Lab, 16th mid-cap barometer (release of 23 June 2026; 534 responses, fieldwork 11 March to 18 May 2026): 77% of leaders report generative AI use by their teams or themselves (+19 points in one year), 31% regular use, 17% observe concrete time savings.

[11] BIRD, official leaderboard (bird-bench.github.io, read 18 August 2026): best score 81.95% execution accuracy (September 2025 submission); best 2026 submissions between 80.83 and 81.67% (Ant Group 19 June, Sber 27 May, Xiaomi 14 July); human reference performance 92.96%.

[12] Spider 2.0, official leaderboard (spider2-sql.github.io, xlang-ai, read 18 August 2026): 96.70% on the Snow track, 76.23% on the Lite track, 65.6% on the DBT track (full analytics workflows).

[13] Jin, Choi, Zhu and Kang (University of Illinois Urbana-Champaign), arXiv:2601.08778, v3 of 19 January 2026: systematic annotation errors identified in the reference answers of text-to-SQL benchmarks, including BIRD and Spider 2.0.

[14] Li, Du, Xu, Zhu, Wang and Mao, DeepResearch Bench II, arXiv:2601.08536 (January 2026): 132 research tasks, 9,430 criteria drawn from more than 400 hours of expertise; the strongest models satisfy fewer than 50% of the criteria.

[15] Zhan, Fan, Huang, Guo and Huang (Zhejiang University, University of Hong Kong), arXiv:2601.22984 (30 January 2026): hallucination propagation along the trajectory of deep research agents; no agent robust end to end; temporal anchoring and source homogeneity biases.

[16] Cloudera and Harvard Business Review Analytic Services, release of 5 March 2026 (October 2025 survey, more than 230 data and AI decision-makers): 7% of data judged “completely ready” for AI; 56% cite data silos, 44% the lack of a data strategy, 41% quality or bias.

[17] Nasuni, Sapio Research survey, release of 18 May 2026 (1,000 decision-makers, companies of more than 1,000 employees, United States, United Kingdom, France, Germany-Austria-Switzerland, fieldwork March 2026): 97% deploying or piloting AI agents; 57% of AI projects fail to meet their objectives; 18% at scale; 46% saw their AI initiatives reveal data quality and governance problems.

[18] Efalia, Baromètre DSI Data et IA 2026 (efalia.com, 11 August 2026; methodology not detailed on the page consulted): 96% of organisations judged not ready for AI; 85% of CIOs state they do not control the quality of their document estate; 62% devote less than 1% of the IT budget to it.

[19] Romasanta (Esade), Thomas (Sydney and Imperial) and Levina (NYU Stern), Researchers Asked LLMs for Strategic Advice. They Got Trendslop in Return, Harvard Business Review (16 March 2026); methodology and quotations relayed by Fortune (10 April 2026): 7 models, 15,000 simulations, 7 strategic tensions, convergence on 6 tensions out of 7 towards the socially valued option; closing quotation on leadership cited verbatim.

[20] Ramírez, Lang, Köhler and Mennell, A Faster Way to Build Future Scenarios, MIT Sloan Management Review (9 December 2025, Winter 2026 issue): Fazer and Unum cases; the classic approach is described as “waiting six months or more for useful insights”; no quantified gain claimed.

[21] Syntec Conseil, 30th strategy and management consulting barometer, via La Lettre du Conseil (11 June 2026; 140 firms surveyed from February to June 2026): 1.5% decline in 2025, sixth-worst year since 2002; Consultor, 2025 review and 2026 outlook for strategy consulting (17 June 2026): +2% anticipated in 2026 by the same study, consultancies’ AI usage now a client negotiation argument.

[22] BCG, 2025 results release (23 April 2026): $14.4 billion in revenue (+7%, 22nd consecutive year of growth); AI- and technology-centred services account for more than 40% of total revenue, driven by 25% year-on-year growth in AI services; 33,500 employees.

[23] Consultor, Chez McKinsey, l’IA s’invite aux entretiens de recrutement (27 January 2026): case study worked with the internal assistant Lilli; assessment of prompt quality, critical thinking in the face of outputs, and the ability to contextualise them.

[24] Bernard Marr, How Klarna’s AI Agent Strategy Backfired But Became a Useful Lesson, Forbes (16 July 2026): 2.3 million conversations in the first month, resolution time from 11 to 2 minutes, support staff from 5,000 to 3,500 people, then rehiring of around a hundred specialists; CEO Sebastian Siemiatkowski’s admission as reported by Forbes (“cut too far, too quickly”, loss of valuable human expertise).

[25] Journal du Net (7 July 2026), analysis of the April 2026 French inter-inspectorate report (IGF, IGAS, IGA) on AI in the public sector: up to 40% unregulated AI use among local government staff; conversational assistants rarely SecNumCloud-certified; companies’ strategic data addressed only marginally.

[26] Daniel Kahneman, Thinking, Fast and Slow (2011): confirmation, anchoring, overconfidence and status quo biases in intuitive decision-making.

[27] Gary Klein, Performing a Project Premortem, Harvard Business Review (September 2007): the premortem technique, projecting oneself into the project’s failure to list its causes before deciding.