Must you consult the works council before deploying AI?

In France, yes, as soon as the tool touches work organisation, and 2026 turned that theoretical answer into an immediate operational risk.

Two decisions established it. On 29 January 2026 the Nanterre court suspended the rollout of two HR management tools with built-in AI features, pilot phase included, under a penalty of €500 per day. On 21 May 2026 the Paris Court of Appeal held that making AI tools available to employees to assist them in their work constitutes the introduction of new technology, and suspended their use under a penalty of €1,000 per day, with damages.

The legal basis is old; its application is new

Article L. 2312-8 of the French Labour Code has long required informing and consulting the works council before introducing new technology or making any significant change to working conditions. Article L. 2312-37 adds prior consultation on any major project introducing new technology.

Those texts served for twenty years for CCTV, vehicle tracking and activity monitoring software. What 2026 added is their plain application to AI tools, including consumer tools simply made available to teams.

The Paris Court of Appeal’s formulation explains why: no office software could previously claim to reproduce intellectual capabilities, as AI tools now do. A spreadsheet automates a calculation; an agent takes over part of a job. Those are not the same thing in terms of working conditions, and the courts have now said so.

What triggers the obligation, and what does not

The criterion is not the technology, it is the effect. A language model inside a spellchecker changes nothing about how work is organised; an agent that writes the write-ups, screens applications or prepares follow-ups changes a great deal.

Three situations fall clearly within scope.

The tool takes over part of a team’s work. That is the case for an agent that performs tasks rather than suggesting text. The distribution of work changes, so working conditions change.

The tool is used to assess people. Annual reviews, assignment allocation, identifying training needs, precisely the scope of the software suspended at Nanterre.

The tool observes activity. Even without monitoring intent, a system that records what each person does falls under already settled case law.

Conversely, individual, occasional use with no effect on work distribution or assessment triggers nothing. The genuine grey area sits between the two, and it narrows as the tool becomes useful.

The point that makes the risk serious

Failure to consult is treated as a manifestly unlawful disturbance. The consequence is procedural and heavy: a judge can order suspension without the claimant having to demonstrate urgency.

In other words, a works council going to court does not have to prove the tool is harmful. It only has to establish that it was not consulted. That is a fast, cheap action with a predictable outcome, which explains why it is now being used.

And the sanction does not address the past: it stops the project in progress. For a company that has paid for licences, trained its teams and wired the tool into its processes, suspension costs vastly more than the consultation it tried to avoid.

It is, incidentally, a cause of project failure nobody files under technical causes, although it belongs to the same family as the three usual reasons for abandonment: a deployment stopped by a judge is a deployment that skipped the people it affected.

What to bring to the works council

Consultation is not a formality, but it is not a trial either. Articles L. 2312-14 and L. 2312-15 set the terms: sufficient time, precise written information, a reasoned response from the employer.

In practice a useful file fits into five points, each of which representatives will ask about anyway.

What the tool does, task by task. Not the vendor’s brochure: the list of actions it will actually perform, and on what data.

What remains decided by a human. This is the central question, and it is the same one as the approval line. A file that answers clearly here defuses most of the concern.

The expected effect on roles. Task distribution, workload, expected skills. Saying “no impact on employment” without showing it is the surest way not to be believed.

What is logged, and who can see it. Fear of surveillance is legitimate and is addressed with detail: what is recorded, for how long, who can consult it.

Reversibility. What happens if the trial does not convince.

Two edge cases staffing firms hit

The consultant working at a client site. The tool is deployed by you, but the person works elsewhere. The consultation is still yours: the employer is the one introducing technology to its own employees, wherever the work is performed. If the agent reads client data, however, steering committee notes, exchanges with their teams: you are in a different, contractual subject, and it is better handled before a project director discovers it.

The reverse also happens, and it is trickier: a client deploying an AI tool that affects your consultants has no obligation to consult your works council. Protection there comes from the services contract, not from labour law.

The company with no works council. Below eleven employees there is no consultation obligation, and many young firms are in that position. Two important caveats: the threshold is assessed over twelve consecutive months, so rapid growth creates the obligation mid-project; and the absence of a works council waives none of the other duties, informing employees about processing that concerns them, transparency on what is logged, compliance with Article 22 GDPR on automated decisions.

In both cases the practical recommendation is the same and it costs nothing: write the file anyway. The five points above are the documentation you will need for your large-account clients regardless, since their procurement teams now ask the same questions, and for the compliance obligations coming.

The timeline, and how not to lose it

Allow one to three months between submitting the file and the opinion, longer if representatives commission an expert review. That period is not negotiable, but it can run in parallel.

The classic mistake is to choose the tool, sign, then open the consultation, and discover you must wait three months with a licence already running. The sequence that works is the reverse: open the information process when you start evaluating solutions, and let the consultation run while you test.

That has a useful side effect. A works council informed early asks questions that improve the requirements, and concerns raised before purchase are far cheaper than the same concerns raised after.

What it changes when choosing a vendor

Three things become criteria, and none appear on a conventional scorecard.

Can the vendor supply the exact list of actions and data touched? If their documentation cannot, you will be writing the file blind.

Can the scope be narrowed? A tool that activates only on certain tasks lets you consult on a reduced scope, then widen it. An all-or-nothing tool forces the broadest consultation from the outset.

Are the logs exportable and readable? They serve compliance, but first they serve answering elected representatives, and they are also what the AI Act will require in December 2027.

One last, less legal point. Consultation is a real constraint, and it is also the one moment when the people who will run the tool are asked what they think. Projects that pass that stage well are, empirically, the ones still in use a year later.