AI deployment in France: consultation before construction
A French deployment has an extra step near the beginning, and projects that discover it near the end lose a quarter. Where a system changes working conditions or the organisation of work, employee representatives have consultation rights, and a system that alters how a team does its job usually qualifies. This is a consultation rather than an approval in most cases, so it is a step rather than a veto, and treating it as a formality is the reliable way to turn it into a fight. Beyond that, the ordering of constraints is not what teams expect. The General Data Protection Regulation binds today and governs the personal data the system touches; the AI Act adds documentation whose weight scales with the risk category of the use, and most internal automation sits well below the heaviest tier. Language sits alongside both and is treated as a nicety far too often.
France, in short
Binding AI law| Instrument that binds | Regulation (EU) 2024/1689, the AI Act, alongside the GDPR |
|---|---|
| Population adoption | The AI Index does not publish a figure for this country. That is an absence of data, not a low number. |
| What differs here | Works councils have consultation rights over changes to working conditions, so a deployment that changes how people work has a procedural step that has no equivalent in most markets. |
| Working language | French, including for any output an employee has to act on |
Legal position checked 2026-09-24. This is a starting point for a question to a local lawyer, not an answer. No compensation figures: see methodology.
The consultation, and how to make it useful
The formal description makes this sound like an administrative gate. In practice it is the earliest structured access a deployment gets to the people whose work is about to change, which is the population every project struggles to reach and reaches too late.
Teams that present a finished system get the predictable outcome: a body with no room to influence anything, exercising the only form of influence remaining to it. Teams that present a described project get questions, and the questions are frequently the same ones a user would have raised in month four, when answering them would have meant rebuilding.
The practical rule is that the consultation should start when the project is describable, not when it is demonstrable. That timing costs nothing, because the process then runs in parallel with stages one and two rather than after stage three.
What binds, in the order it bites
First, personal data. The GDPR is in force, enforced, and applies to the employee and customer data most internal systems touch. Establishing the lawful basis before the build is ordinary practice and remains the single most consequential piece of preparatory work on a French project.
Second, employment law. The consultation described above, plus the general principle that a system which evaluates or monitors people attracts more scrutiny than one which does not. As in Germany, the capability matters and not only the intention, so a project that produces per-person logs should assume it is in scope.
Third, the AI Act. Regulation (EU) 2024/1689 attaches obligations to the use rather than the technology. For a large share of workplace automation the practical effect is being able to state what the system does, on what data, with what human oversight, in writing, before anybody asks. Organisations that already document their systems find this unremarkable.
Language, and why it is a quality problem
Output an employee has to act on has to be right in French, including register and the domain's own vocabulary. This is not politeness. A system producing French that reads as machine-translated is treated as unreliable in general, including on the parts it gets right.
The design consequence is concrete: the evaluation set must be in French and labelled by French-speaking domain experts. An evaluation set assembled in English measures a system that nobody in the organisation is going to use, and it will report improving numbers while adoption falls.
The same applies to failure messages, which are the output users read most often when they are already annoyed, and which are the most commonly left in the original language.
What this changes about who you need
A French deployment needs someone who can hold the consultation conversation, and that is a different person from the one who builds. Sending an engineer alone into a works council meeting is unfair to the engineer and unproductive for the project.
It also needs a domain expert who works in French to own the evaluation set. This person is the scarce input on most French projects, more so than engineering capacity, and identifying them in week one is worth more than any amount of planning.
What does not change is the method. The five stages hold; stage two stretches or contracts on the same variable as everywhere, which is how long a read credential takes; and stage four simply acquires a formal calendar at its front rather than a different character.
