France’s “Osez l’IA” plan is entering a new phase: public support is shifting from awareness to implementation. For an SME leader, however, the relevant question is not whether to “adopt AI,” but which process deserves investment, which support scheme fits the project and when an unconvincing pilot should be stopped.
What has just happened
On 4 September 2026, the French Ministry for the Economy published the first annual review of the “Osez l’IA” plan. It reports more than 35,000 businesses reached, 615 AI Ambassadors, 88 listed providers and 54 published use cases. The government plans to roll out 1,000 Data AI Diagnostics for SMEs and mid-sized companies with 10 to 2,000 employees, delivered as an eight-person-day engagement with 40% of the cost covered by France 2030. New AI Accelerator cohorts are also planned, with an 18-month programme receiving 46% support. The figures and current terms appear in the Ministry’s annual review of the “Osez l’IA” plan.
Why this may—or may not—concern you
The announcement matters if your company has identified a costly or difficult business task: analysing supplier files, classifying customer requests, preparing commercial answers, extracting information from documents or assisting employees with a specific procedure. A subsidised diagnostic may then reduce the cost of deciding whether integration is justified.
It also matters when several departments are running disconnected experiments. A structured diagnostic can compare use cases and produce a common roadmap, provided that the expected deliverable goes beyond a generic catalogue of ideas.
It matters less if your only objective is to give a few employees access to a general-purpose AI assistant. Such usage can often be tested through existing tools without immediately committing to a full diagnostic or a long transformation programme.
Most importantly, adoption is not the same as value. Bpifrance’s 2026 survey reports that 77% of responding mid-sized company leaders use generative AI professionally, while only 17% of user companies observe time savings. The analysis is based on 534 responses from French group headquarters. It cannot be applied mechanically to every SME, but it shows why access to a tool is not an operational result. The data is available in Bpifrance Le Lab’s sixteenth annual survey of mid-sized companies.
This article extends our earlier analysis of the expanded “Osez l’IA” plan. The September update is more useful operationally because it describes the support formats and their current funding rates.
What this changes in practice
What works: a bounded and measurable workflow
Defensible AI projects do not start with a model selection. They start with a task that has a trigger, identifiable input data, an expected output and a business owner able to review exceptions.
Supplier onboarding offers a documented example. Bpifrance reports that Activ Inside connected its document repository to an AI agent used by the quality department. The task fell from approximately two hours to one hour and the tool was deployed to two or three employees. Both implementation and annual recurring costs are reported as below €50,000, with implementation completed in less than a year. These company-specific figures, published in Activ Inside’s supplier onboarding case study, should not be treated as market pricing.
The transferable pattern is not merely “using an AI agent.” It is the combination of a narrow scope, available documentation, identified users and a before-and-after comparison performed on the same task.
A useful project brief should specify:
- the department that owns the process;
- the event that triggers the workflow;
- the documents and data the agent may access;
- the actions it may perform without approval;
- the situations requiring human review;
- the measure used to continue, correct or stop the project.
Our guide to validating a first AI workflow explains how to turn that brief into a controlled pilot.
What fails: funding a vague intention
A diagnostic will not repair a process whose rules are unknown. When every employee handles requests differently, the AI will reproduce the variation or generate excessive exceptions. The priority is then to stabilise the procedure, not deploy an agent.
A general assistant disconnected from the CRM, ERP or document repository also creates limited shared value. Users must copy the data, explain the context and manually enter the result back into the business system. A new interface has been added, but no process step has been removed.
Another common dead end is a proof of concept with no operational owner. A prototype may perform well during a demonstration and remain impossible to run because there is no support process, API cost control, action log or error-handling procedure.
Finally, a subsidy does not make a weak project economically sound. It reduces part of an eligible expense; it does not guarantee usable data, employee adoption or a positive return.
What it actually costs
The September announcement does not publish an absolute price. It states that the eight-person-day Data AI Diagnostic receives 40% support and that the 18-month Accelerator receives 46% support. The net amount should therefore be calculated only after obtaining a quote and written confirmation of eligibility under the terms described in the official programme review.
| Budget item | What it should cover | Hidden cost to disclose |
|---|---|---|
| Scoping | Target process, data, risks and acceptance criteria | Business owner and IT time |
| Subsidised diagnostic | Analysis and roadmap defined in the quote | Net contribution and out-of-scope work |
| Pilot | Workflow, agent, integrations, interface and tests | Document preparation and exception handling |
| Production | Security, permissions, logging and monitoring | Acceptance testing, support and change management |
| Operations | Licences, APIs, hosting and maintenance | Model changes, controls and reversibility |
This prevents management from comparing a monthly licence price with the cost of an operational system. The useful figure is the cost of running the process, not the price of accessing the model.
What to do by 31 January 2027
- By 18 September — CEO and business owner: select a process and approve an opportunity brief describing its trigger, inputs, expected output, exceptions and decision owner.
- By 30 September — CFO and IT lead: produce a total-cost sheet separating the diagnostic quote, expected support, integration, licences, consumption, internal time, security, support and exit costs.
- By 15 October — business owner and IT: prepare a representative test set and an acceptance grid defining errors, human approvals and situations that automatically stop the workflow.
- By 30 November — CEO: choose between a direct pilot, a Data AI Diagnostic and an Accelerator. Record the authorised budget, business owner and stop condition.
- By 31 January 2027 — project lead: close the pilot with a documented decision: stop, make a limited correction or move into production with an operating owner, backlog and recurring budget.
If the underlying process remains unclear, use our process mapping playbook. D1 Consulting can formalise it through a business process audit, supported where appropriate by BPMN Studio, our process modelling application.
What you can ignore for now
The government’s target for 80% of SMEs and mid-sized companies to adopt AI by 2030, stated in the “Osez l’IA” annual review, is not a legal obligation for your company. It does not justify an immediate purchase or a company-wide programme.
You can also ignore the number of listed vendors until the target process is documented. A catalogue can support market research, but it cannot replace an assessment of integrations, hosting, operating costs and reversibility.
A long programme is not automatically required for an isolated task. When the data is available, the process is stable and the impact can be verified, a directly integrated workflow pilot may be more appropriate.
Do not postpone the correction of an obviously defective procedure while waiting for a subsidy. Diagnostic funding should not become funding for indecision.
Our view at D1 Consulting
The new “Osez l’IA” review is helpful because it makes structured scoping more accessible. Its limitation is equally important: awareness and diagnostic volumes do not yet show how many workflows remain in production or the net value they generate.
Our position is to use public support when it helps fund a decision the company genuinely needs—not to manufacture a project solely because support exists.
Our method is to:
- describe the business task and rules before selecting technology;
- calculate total cost with and without public support;
- build the workflow or agent with human approval proportionate to the risk;
- organise acceptance, production and ongoing operations;
- stop the project when the before-and-after evidence does not justify continuation.
Implementation can be delivered through our Automation & Process Optimisation service, covering workflows and AI agents. Our IT Project Management service handles scoping, acceptance, supplier coordination and change management when the initiative affects several teams or systems.
As a France Num Activator, D1 Consulting can also help connect the operational project with available public schemes—without confusing an application for support with an investment decision.
👉 Book your free 30-minute diagnostic to select an AI workflow, calculate its total cost and decide whether a public support scheme fits the project.
