From PoC to Industrialisation

In 2026, the issue is no longer accumulating demonstrations but turning useful PoCs into production capabilities that are monitored, governed and genuinely adopted.

• 46% of AI PoCs reach production in IDC-Lenovo’s CIO Playbook 2026.
• BCG’s framework distinguishes 10% algorithms, 20% technology and data, and 70% people and processes.
• This white paper connects the seven last-mile frictions, five change-management levers, AgentOps and a 12-month roadmap.
• It follows Maturity of AI in French SMEs 2025-2026.

1. June 2026: the year the question changed

Rising use is shifting European leaders’ question from experimentation to the ability to deliver repeatable production.

• Eurostat reports that 20.0% of EU27 enterprises used AI in 2025, up from 13.5% in 2024.
• IDC-Lenovo’s CIO Playbook 2026 reports that 46% of AI PoCs reach production.
• HCLTech reports, from 467 leaders of enterprises with revenue above US$1 billion, that 43% of major AI initiatives risk failure, mainly through shortcomings in execution.
• Deloitte finds among 3,235 leaders in 24 countries that 23% of organisations already use agentic AI at least moderately, while 21% report mature governance for those agents.

“We have around ten PoCs. Some work. But nothing has scaled. How do we industrialise?”

The BCG “10-20-70” framework helps allocate effort without reducing transformation to technology alone.

• 10% concerns algorithms.
• 20% concerns technology implementation and data.
• 70% concerns people, processes and changes in ways of working.
• For agentic systems, AgentOps adds the continuous control required by this organisational transformation.

Go deeper: The Junyr Method™ Scale: five levels of AI maturity.

2. The European picture: 20% adoption, a skills wall

European adoption is rising, but its distribution by company size and country prevents direct extrapolation to any given French SME.

• Eurostat measures 55% adoption among large enterprises versus 17% among small enterprises in 2025.
• Eurostat’s national results range, for example, from 42% in Denmark to 5.2% in Romania, with differences that also reflect national economic structures.
• These rates describe reported AI use; they measure neither a Junyr level, deployment quality nor value created.

Enterprises that considered AI without adopting it primarily describe an organisational-capability problem.

• 70.9% cite a lack of relevant expertise in Eurostat’s data.
• 52.5% cite uncertainty about legal consequences.
• 48.8% cite data protection and privacy.
• BCG AI at Work 2026 reports that 72% of surveyed workers already see the skills expected in their role changing and 61% believe agents could perform at least half of their work within three years.

Industrialisation must therefore address skills, trust, coordination and risk control together.

• AI literacy creates a shared vocabulary.
• Process redesign creates the conditions for measurable improvement.
• Governance makes uses visible and open to arbitration.

Go deeper: The end of prompt engineering.

3. The last mile: integrating technology into work

The “last mile” describes the gap between demonstrated technical capability and lasting integration into an organisation.

• Karim Lakhani, Jared Spataro and Jen Stave’s HBR article, The “Last Mile” Problem Slowing AI Transformation, connects the blockage to organisational design.
• Deloitte finds that 30% of surveyed organisations redesign core processes around AI, while 37% add it without fundamental redesign.
• The Microsoft Work Trend Index 2026 documents individual uses that often precede formal adoption by the enterprise.

Stanford and KPMG each illuminate one part of scaling, but neither supports a single statistic about SME maturity.

• The Stanford AI Index 2026 separates broader AI deployment from agent use; scaled agent deployment remains in single digits in almost every function studied.
• A per-function rate does not show that fewer than 10% of organisations have scaled any form of AI in at least one function.
• KPMG classifies 11% of its sample as “AI leaders”; this group represents neither all organisations obtaining value nor a representative sample of French SMEs.
• KPMG’s sample covers organisations above US$100 million in revenue, 75% of which exceed US$1 billion.

Industrialising means converting a fragile local use into a collective process whose outputs, costs and risks are observable.

• A business owner is accountable for the result.
• A reference set of cases checks quality.
• A control plan sets action and escalation rights.
• Indicators track adoption and impact separately.

Go deeper: Five framing mistakes that weaken an SME AI transformation.

4. The seven last-mile frictions, seen from the field

The framework drawn from the HBR article connects each organisational blockage to a production decision.

• Portfolio frictions require a priority, sponsor and stop criterion.
• Process frictions require redesign, qualified data and a business owner.
• Agentic frictions require governed architecture, bounded rights and an explicit value outcome.

Three frictions up close

Pilot proliferation is resolved by turning every demonstration into an explicit allocation choice.

• Link the PoC to an annual priority.
• Name a sponsor, business owner and indicator.
• Stop trials without a mandate so budget and attention can focus on the others.

Process debt requires redesigning the work before automating it.

• Map the decision, inputs and exceptions.
• Remove steps that add no value and clarify reference data.
• Then automate the target process, as developed in the first white paper.

Identity tied to tacit knowledge is addressed by giving the expert a guarantor role rather than bypassing their expertise.

• Give them ownership of the reference base and edge cases.
• Formalise their validation and stop rights.
• Turn their corrections into a reusable asset for the team.

The five forms of human resistance, one by one

Human resistance has different causes and therefore requires five distinct responses.

• Fear of substitution: state the AI’s scope, the decisions that remain human and the intended use of released capacity.
• Distrust of outputs: install native verification, evidence and a readable log.
• Skills deficit: train according to role and exposure to the system.
• Lack of ownership: name a business champion accountable for quality and adoption.
• Regulatory uncertainty: classify the system, use case and enterprise role before deriving obligations.

Go deeper: Junyr Agents™: delegating to AI without losing control.

5. What a controlled trial establishes, and what it does not

The randomised controlled trial by Tinelli and co-authors establishes an effect on a management-coaching practice and provides a cautious foundation for thinking about change support.

• Impacts of adopting a new management practice: Operational Coaching™ randomly allocated managers from English SMEs between 40 trained firms and 22 controls.
• The intervention concerned an enquiring, facilitative posture under the public Business Basics programme.
• Time spent coaching teams increased by 13.8 percentage points, with a statistically significant difference on that behaviour.
• Productivity trends do not constitute proof of improved business performance; the authors call for further work on this relationship.

Applying that finding to AI industrialisation remains a methodological inference that is useful but was not tested by this trial.

• A change in management behaviour can be measured and developed.
• This study concerned neither an AI deployment, AgentOps nor the ROI of a digital programme.
• It supports testing coaching behaviours rather than assuming their economic effect.

6. The European framework as a lever, not a brake

The European framework becomes usable when an SME connects each obligation to the relevant system, use and actor role.

• Applicable obligations depend on the system’s classification and whether the company is a provider, deployer, importer or distributor.
• Rules for Annex III high-risk systems generally apply from 2 December 2027, and those for Annex I systems from 2 August 2028, subject to the current transition conditions.
• Amended Article 4 requires relevant providers and deployers to support a sufficient level of AI literacy; it does not by itself guarantee programme-wide compliance.
• The reference source remains Regulation (EU) 2024/1689, supplemented by applicable legislation and guidance.

Governance turns those requirements into adoption levers without promising universal compliance.

• An internal inventory connects systems, purposes, data and accountable owners.
• A hosting policy addresses data location and access.
• Reversibility limits dependency on a supplier.
• Use-case analysis avoids assuming that every SME use falls outside the high-risk categories.

Digital sovereignty becomes practical when the enterprise combines technology choices with an exit capability.

• Mistral Large 3 is released under the Apache 2.0 licence and can run on infrastructure controlled by the enterprise.
• Local deployment still requires secure access, managed updates and qualified data.
• Choosing between cloud, a European provider and owned infrastructure requires an analysis of risk, cost and reversibility.

Go deeper: The death of “all-cloud” · Cryptography 2026.

7. AgentOps: industrialising agents, not just deploying them

An agent should be managed as a probabilistic system acting on a business process, with requirements proportionate to the consequences of error.

• BCG AI at Work 2026 reports that 47% of surveyed workers spend more time directing and supervising AI than performing the relevant work themselves.
• A Sinch survey of 2,527 decision-makers in 10 countries reports that 74% had taken at least one customer-communication agent out of production.
• Gartner’s Hype Cycle for Agentic AI forecasts that more than 40% of agentic projects will be cancelled by 2027, including for cost and risk reasons.

METR horizons measure the duration of baseline human tasks that an agent can complete with a stated probability, not a continuous period of autonomous production work.

• The 50% horizon corresponds to tasks the agent succeeds at roughly one time in two under benchmark conditions.
• The 80% horizon corresponds to success roughly four times in five and is shorter than the 50% horizon.
• The baseline is the time a competent human would take to complete the task; it does not measure how long the agent operates without supervision.
• METR’s published limitations establish no universal production threshold: 80% may be tolerable for an easily checked task, while an irreversible action may require much more.

From DevOps to AgentOps: what actually changes

AgentOps extends software operations to observe and govern the lifecycle of production agents.

• Observability: replay every step and tool call.
• Evaluation: rerun a business reference set before every change.
• Governance: document action, validation and escalation rights.
• Security: restrict accessible data, tools and identities.
• Resilience: plan the stop, human takeover and restoration of a safe state.

The Junyr triptych: Plan → Execute → Verify

The Junyr triptych separates intent, action and control so that every trajectory can be tested and audited.

• Plan: produce a plan bounded by the use case’s rules.
• Execute: call tools with minimum rights and select a model for each step that is proportionate to risk and cost.
• Verify: compare the result with acceptance criteria before continuing.

The minimum control plan for an SME

The minimum control plan should cover interruption, sensitive actions, evidence, costs and continuous improvement.

• Session interruption: a person can stop the agent and the agent stops itself outside its perimeter.
• Human validation: every irreversible or high-impact action awaits approval.
• Auditable log: every Plan → Execute → Verify step remains readable.
• Capped budget: a consumption limit triggers an alert and then a controlled stop.
• Periodic review: the team analyses deviations, corrects instructions and records refusals.

Go deeper: Junyr Agents™: delegating to AI without losing control.

8. Change management in five levers (the Junyr Method™)

The move from Artisan to Orchestra organises five complementary levers that should be adapted to the risk level and chosen scope.

• Mandate and narrative: give released capacity a purpose, sponsor and destination.
• Business champions: bring decisions closer to the field and create lasting accountability.
• Three-tier training: align skills with the roles people perform.
• Minimum viable governance: apply rules light enough to follow and precise enough to audit.
• Adoption measurement: decide in advance what justifies continuing, redesigning or stopping.

Lever 1: Mandate and narrative

The leader makes the transformation understandable by connecting strategy, system scope and expected effects on work.

• Explain why this use case is a priority now.
• State which decisions remain human.
• Describe the intended use of released capacity.
• Repeat a short, stable narrative in governance meetings.

Lever 2: Business champions

A champion respected by their peers creates operational ownership of the process and brings unfiltered deviations to the surface.

• Support users day to day.
• Classify objections, errors and edge cases.
• Propose changes to the reference base.
• Report adoption, quality and incidents to the steering committee.

Lever 3: Three-tier training

Training addresses the skills gap at the depth useful for each audience.

• Build AI literacy for all staff exposed to enterprise uses.
• Train practitioners in templates, acceptance criteria and checks.
• Train champions and leaders in case selection, ROI, AgentOps and the applicable framework.

Lever 4: Minimum viable governance + AgentOps

Minimum viable governance brings essential rules together in materials that teams can actually apply.

• A short, approved usage charter.
• Data classification and access rights.
• An agentic control plan linked to consequence level.
• An owner and escalation route for every use case.

Lever 5: Adoption measurement

Measurement enables action before a use case degrades silently.

• Establish a baseline before deployment.
• Separate use, quality, released capacity and economic effect.
• Review thresholds and incidents periodically.
• Document the decision to continue, correct or stop.

Progression through levels remains a sequencing principle rather than a statistical guarantee of success or compliance.

• Check data, accountability and security foundations before extending.
• Commit to an agent platform only when a process, owner and tests are ready.
• Use the Junyr Method™ Scale as a discussion and prioritisation framework.

Go deeper: The Junyr Method™ Scale.

9. The three-tier training plan & the augmented leader

Tier 1: AI literacy, dispelling fear

Shared AI literacy gives staff the reference points needed to understand authorised uses, system limitations and their rights.

• Explain what AI does and what it does not allow the organisation to infer.
• Name authorised data, prohibited uses and the reporting channel.
• Connect the session to the provider or deployer role and the relevant system under Article 4’s framework.
• Gather concrete questions that will feed the charter and training plan.

Tier 2: Practitioners, commanding AI, not talking to it

Practitioners turn an improvised exchange into a repeatable procedure controlled by acceptance criteria.

• Structure the context, roles and input data.
• Break down tasks and specify the output format.
• Test results on representative and edge cases.
• Version the templates and process rules.

Tier 3: Architects & the augmented leader

Champions and the executive committee learn to arbitrate the portfolio, risks and economics of the system without delegating their judgement.

• Select use cases and read an ROI calculation.
• Manage AgentOps, action rights and incidents.
• Connect the legal framework to the system and the enterprise’s role.
• Use AI to prepare decisions while retaining human decision-making and accountability.
• Use France Num and the “Osez l’IA” plan when their conditions apply.

Go deeper: The end of prompt engineering.

10. Measuring adoption: KPIs and the agentic control plan

The three families of KPI, in practice

The dashboard separates three indicator families so that use frequency, business result and operational control are not conflated.

• Adoption: active users, frequency by team and depth of use within the process.
• Impact: cycle time, quality, released capacity, attributable incremental margin and avoided costs.
• Control: inference cost, human-validation rate, incidents, refusals, takeovers and compliance with internal rules.
• Method: a documented baseline, period, population, total cost and attribution rule.

ROI should be calculated over a defined period as (attributable benefits − total costs) / total costs, without double-counting released hours and any revenue those hours may help generate.

• Distinguish cash savings from released capacity.
• Use incremental margin rather than revenue when measuring an economic effect.
• Document deployment, operation, supervision and correction costs.
• Present avoided risks as expected loss with explicit assumptions.

When to stop a use case

A use case should be redesigned or stopped when repeated measures show it is not crossing the thresholds decided at the outset.

• Review at least two consecutive periods to avoid deciding from an isolated signal.
• Investigate usefulness, quality, process and training separately.
• Document the decision and retain reusable lessons.
• Return to the five-phase method in the first white paper when a new framing is required.

Go deeper: Token budgets and AI APIs: the FinOps guide for SMEs in 2026.

11. The deflation of expertise: why value is shifting

Falling model-access costs shift value towards organisational assets that a competitor cannot simply rent or download.

• Open-weight models reduce supplier dependency for some uses while leaving operation to the enterprise.
• Artificial Analysis illustrates the narrowing of some performance gaps in April 2026.
• A study revised in March 2026 estimates a rapid fall in inference cost at constant performance, with findings dependent on benchmark and period.
• Judgement, proprietary data, processes, customer relationships and governance capacity remain sources of differentiation.

The strategic response is to invest in the quality of the work system as much as in the model’s raw capability.

• Document processes and decision criteria.
• Build maintained business reference sets.
• Reserve human intervention for ambiguous or high-consequence decisions.
• Measure final quality rather than cost per request alone.

Go deeper: Where is AI’s value heading? Three signals of commoditisation.

12. INDUSTEC: organising the evaluation of a quotation assistant

The illustrative INDUSTEC example connects the operation of a quotation assistant to the team’s responsibilities, so that a first use case can be checked for control and accountability before any expansion is considered.

• The leader connects the use case to the intended business outcome and defines the decisions that remain human.
• Business champions receive training before deployment, with a defined remit and backup for each role.
• The team reviews adoption, corrections and errors every week.
• A monthly committee decides on expansion and corrective action from the evidence gathered.

The champions’ weekly work

A weekly routine can turn individual corrections into collective improvement when the champions have a sample of quotations, a rejection log and an approval process for changes.

• Review a sample of quotations produced with the assistant.
• Record rejected proposals and the reason for rejection.
• Raise problems in the product reference base.
• Test corrections before rolling them out.
• Summarise adoption, quality and incidents on one page for the monthly committee.

Two difficulties to address before expansion

Preparing the reference data and limiting the number of simultaneous use cases address two distinct risks that should be examined before the process is reproduced across other teams.

• Incomplete or contradictory reference data require correction and validation before production.
• Work spread across several cases weakens correction capacity; the committee can make expansion conditional on stabilising one end-to-end process first.

Go deeper: the INDUSTEC evaluation protocol.

13. A 12-month roadmap & European funding

The following roadmap is a reference scenario for an SME or mid-cap with 50 to 500 staff, to be costed according to scope, starting point and risk level.

• Horizon: 12 months to frame, stabilise one use case, extend and consolidate.
• Indicative budget: €30,000 to €80,000, excluding costs specific to an unusually complex system or data estate.
• Control condition: validate every phase through deliverables and measurements, not through the calendar alone.

Months 0-1: Framing and narrative

The first phase turns a portfolio of ideas into sponsored, measurable decisions.

• Diagnose the maturity relevant to the scope.
• Choose three to five cases linked to business priorities.
• Name the sponsor, owner and success indicator.
• Secure executive approval for a framing note.

Months 2-3: Human foundations

The second phase prepares the people, rules and data for the first use case.

• Appoint and train champions.
• Build AI literacy among the staff concerned.
• Adopt the usage charter and data classification.
• Audit and clean the first case’s reference data.

Months 4-7: Industrialising the first use case

The third phase puts a redesigned process into production with verifiable quality and adoption thresholds.

• Redesign the target process from a blank page.
• Train practitioners and test edge cases.
• Deploy with logging, supervision and human takeover.
• Require two consecutive months above threshold before extending.

Months 8-10: Extending and orchestrating

The fourth phase reuses validated foundations to open new cases under the same control plan.

• Prioritise the second and third cases.
• Introduce agents only where their action rights can be bounded.
• Confirm that critical events are logged.
• Compare costs, quality and supervision load with initial assumptions.

Months 11-12: Measuring and consolidating

The final phase turns observed results into portfolio decisions for the following year.

• Consolidate benefits with documented formulae and periods.
• Review incidents, refusals and human takeovers.
• Establish a quarterly AI committee with defined responsibilities.
• Build the year-two roadmap.

Funding the journey

Public schemes may reduce the cost of an eligible project, but rates, dates and conditions must be checked at the time of application.

• The “Osez l’IA” plan structures national support.
• Diag Data IA and IA Booster France 2030 apply their own company-size, service and timetable conditions.
• European Digital Innovation Hubs provide services whose funding depends on the hub and applicable state-aid regime.
• Funding covers an eligible scope; it does not replace the economic case or the internal capacity to manage delivery.

14. Conclusion: to industrialise is to transmit

Industrialising AI means transmitting a production capability to the whole organisation, with accountability, quality criteria and an explicit stop right.

• Technology and data make up 30% of BCG’s framework, split between 10% algorithms and 20% implementation.
• The 70% assigned to people and processes requires a mandate, champions, training and adoption measurement.
• The cited coaching trial shows that management behaviour can change after an intervention; it does not by itself validate AI industrialisation or its ROI.
• AgentOps makes agent actions observable, testable and reversible.

Agent interoperability is already an architecture decision, even though fully automated commerce between agents remains a prospect.

• The A2A protocol organises exchanges between agents.
• The Model Context Protocol standardises exposure of context and tools.
• An SME can prepare its APIs, identities and logs without assuming its ecosystem’s adoption timetable.
• Every opening should retain the same authentication, authorisation and traceability rules.

Take action

The AI Maturity Audit offered by the Junyr.eu network is a calibrated first step for positioning the enterprise and selecting its next project.

• Format: a free 30-minute video call with no commitment.
• Facilitation: a network practitioner follows a shared structure.
• Deliverable: one page showing the position on the Junyr™ Scale, the main obstacle and the first relevant project.
• Access: book the audit on Junyr.eu or read the blog.

Paul-Antoine TUAL, AI Transformation Leader · Engagements delivered through the SASU Croissance et Transitions.

First published in June 2026; revised on 6 September 2026.