Recommendations for using AI well at work

Offer to contribute before the works council (CSE) gives its opinion: the occupational physician advises employee representatives. Fill in the assessment grid in a meeting, without any individual health data, and analyse concrete work situations.

Labour Code R. 4623-1L. 2314-3

Sources: Works council and AI · occupational health · Assess an AI project

Explore recent changes in the work, AI included, without automatically attributing these signs to it. Who approves what, and how many times a day? What has the time saved been used for? Does the tool make demands outside working hours?

Sources: Publication · LLMs and psychosocial risks · Reading · Human in the loop

Move from the individual signal to the collective one: alert the employer and the works council without any individual health data, and propose a workplace action and the recording of the risks in the DUERP (the company’s occupational risk assessment document).

Sources: Economic and social risks · Publication · LLMs and psychosocial risks

Tertiary prevention: an appointment with the occupational physician on request, support for returning to work after exhaustion, prevention of health-related job loss, a skills assessment or training leading to a qualification.

Sources: Psychosocial risks · Claude skill · job retention (in French)

No personal, medical or identifying data. Rather than a blanket ban, map existing uses, define prohibited data with concrete examples and offer authorised alternatives.

Source: AI in OHS services · shadow AI

Assist, never conclude: an authorised environment (certified health-data hosting, HDS), de-identified text, AI as a second reader. Analysis, wording, validation and signature remain in medical hands; an individual score requires a formal analysis before any decision.

Sources: AI in OHS services · uses · Publication · AI and medical recommendations

A good candidate for a limited pilot: internal document search with sources displayed, training administration, routing simple requests with human follow-up.

Source: AI in OHS services · value / risk matrix

Follow the sequence: need, value, data, risk, parallel test, collective decision. Measure the tool (errors caught, missed or invented; total time after correction) and the work (checking load, skills, incidents).

Source: AI in OHS services · building a pilot

Train each profession before deployment, on its real tasks and the errors it cannot accept. Protect the practice of residents and beginners: when you review a summary prepared by AI, you are the human in the loop.

AI Act art. 4

Sources: AI in OHS services · training · Reading · Human in the loop

Describe the task, the people concerned, the expected result and the points where an error would have consequences. “Saving time” is not a task. Name who will be able to launch, change or stop the pilot.

Sources: Law and governance · Understanding AI · five questions

Drafting, summarising, searching

Usage note, data entry rules, human verification protocol and review time built into the workload.

Recruiting, assessing, managing mobility

Sensitive use: AI Act classification (some employment uses are high-risk, with rules applying from 2 December 2027), non-discrimination, informing the works council and candidates in advance, a way to challenge decisions.

Labour Code L. 2312-38L. 1132-1

Planning or allocating tasks

Mapping of work before and after, rules for contesting, attention to autonomy, fairness and work intensification.

Monitoring activity or performance

Purpose, necessity and proportionality test, informing employees in advance, consulting the works council before implementation. No fully algorithmic individual assessment.

Labour Code L. 1222-4L. 2312-38GDPR art. 22

Touching on health or fitness for work

Legal, clinical, technical and organisational analysis before any pilot: health data, professional secrecy, separation of roles.

Source: Law and governance · use cases

Define in advance what it can do on its own, require explicit approval for actions with serious consequences, keep a log of its actions and keep a real means of stopping it. The higher the level, the better equipped human oversight must be.

Sources: Works council and AI · autonomy · Reading · Human in the loop

In companies with at least 50 employees, inform and consult the works council (CSE) early enough for its opinion to still be able to change the project, including for a tool presented as optional (2026 court decisions). Provide a description of the need, the groups of employees concerned, the data and the supplier’s documentation.

Labour Code L. 2312-8L. 2312-14

Sources: Works council and AI · Law and governance · social dialogue

Inform early and precisely: distinguish what has been decided, what remains uncertain and when the next information will come. Include AI in strategic workforce planning (GEPP), open training to everyone and preserve the positions where people learn the job.

Labour Code L. 2242-20

Sources: Psychosocial risks · insecurity · Economic and social risks

Before the pilot (T0), fill in the nine-axis grid: intensity, emotions, autonomy, relationships, values, insecurity, contestability, supervision load, skills. Record the risks identified in the DUERP (the company’s occupational risk assessment document) where the framework requires it.

Labour Code L. 4121-3R. 4121-1

Sources: Assess an AI project · Law and governance · DUERP

Not before measuring the complete task, checking and corrections included, task by task and across different groups of staff: a positive average can hide negative effects. Decide with the team what share of the gains goes to production, quality, learning and recovery.

Sources: Reading · AI and real work · Psychosocial risks · intensity

State in writing that professional judgement takes precedence over the AI output, guarantee the right to correct, refuse or report without sanction, and designate who can review a decision, and within what time.

Source: Assess an AI project · autonomy, contestability

Match the volume of outputs to the real capacity for checking, count supervision in workload and staffing, and train during working hours. Plan periods when no response is expected, even if an agent is blocked.

AI Act art. 4art. 26

Sources: Reading · Human in the loop · Reading · AI agents and disconnecting

Prepare the eight documents of the pilot file: usage note, current and planned work, AI Act classification, data mapping, impact assessments, information and consultation file for the works council, reversible pilot protocol, monitoring and incident plan.

Source: Law and governance · pilot file

Continue the use, subject to conditions, with the next measurement already scheduled.

Correct the organisation, resources or rules; reduce the scope or suspend the use.

Source: Assess an AI project

Request the information and consultation procedure without delay. In companies with at least 50 employees, where consultation is required it must take place before the decision is made, including for a generative AI tool presented as optional (2026 court decisions).

Labour Code L. 2312-8L. 2312-14

Source: Works council and AI · frequently asked questions

Ask for details of the model, supplier, version, hosting, memory and access: emails, files, employee data, the internet, business software. The risk depends on the complete system, not only on the use case presented.

Source: Works council and AI · the ten questions

Ask for proof of human control: who can stop the AI, and does this person have the time, information, skills and authority to do so? The higher the level, the more robust the justification, monitoring and reversibility must be.

Source: Works council and AI · assessing autonomy

Specific rules apply: the works council must be informed in advance about recruitment support methods and automated personnel management systems, and informed and consulted before any means of monitoring employees’ activity is introduced. Ask about the purpose, proportionality and the ways to challenge decisions.

Labour Code L. 2312-38

Sources: Works council and AI · legal framework · Law and governance · use cases

Ask for the mapping of work before and after, and the expected effects for each group of employees concerned. A supplier’s demonstration is not enough to describe future work: examine these points with the employees concerned.

Sources: Law and governance · social dialogue · Works council and AI · psychosocial risks

Call on the occupational physician or the multidisciplinary team of the occupational health service before giving the opinion. The occupational physician advises employee representatives and takes part, in an advisory capacity, in meetings on health and working conditions.

Labour Code R. 4623-1L. 2314-3

Source: Works council and AI · occupational health

Require an identified person able to contest, correct or stop the system, a reporting channel with logging and feedback, written stop thresholds and a tested rollback procedure.

Source: Works council and AI · questions 8 to 10

Ask for this information at each review point, together with the rules for human intervention, changes in objectives and the conditions for immediate suspension.

Source: Law and governance · during the pilot

They must remain collective and transparent, be discussed with the works council and serve to adjust the organisation, never to judge people. A system for monitoring activity requires employees to be informed beforehand.

Labour Code L. 1222-4L. 2312-38

Source: Reading · Human in the loop

Ask for the gaps between planned and real work, the effects on health, autonomy and skills, the corrections made and a reasoned decision to continue or stop. The assessment grid, completed again at T1 and T2, makes it possible to check the commitments.

Sources: Law and governance · after the pilot · Assess an AI project

Ask for a reassessment: these changes, like an incident, must reopen the analysis. Have them listed in the opinion as triggers for reassessment.

Sources: Works council and AI · reassessment · Assess an AI project · outside the schedule

Start from real situations: tasks that have moved, checks, corrections, exceptions, decisions that are influenced. Activity often shifts towards checking outputs and handling the most complex cases.

Sources: Homepage · prevention specialist profile · Law and governance · role of prevention

Analyse supervision as a vigilance task: approvals per hour, time and authority to refuse, session length, breaks, rotation, responsibilities set out in writing. Record it in the DUERP (the company’s occupational risk assessment document) as a situation exposing workers to psychosocial risks.

Source: Reading · Human in the loop

Document what enters and leaves the occupation, the time actually saved, the training needed and who is responsible for checking. Being able to do a task with AI does not automatically mean having the skill.

Source: Reading · Occupational boundaries

Identify the exposed groups and document the differing effects: a positive average can hide negative effects, and an average gain must not become a uniform target.

Sources: Reading · AI and real work · Law and governance · role of prevention

Assess it task by task (omissions, corrections, quality of the action plan), with no identifying data in an unauthorised tool, and do not let an automatic summary replace field analysis.

Sources: AI in OHS services · uses · AI in OHS services · responsibilities

Propose some: hours, sick leave, errors, incidents, workarounds, changes in skills, recorded at T1, T2, T3 and at each change of tool. They serve to adjust the organisation, never to assess people.

Sources: Assess an AI project · Psychosocial risks · early detection

Do not enter any personal, health, confidential or identifying data. Instead, explain your need to your manager to obtain an authorised solution.

Source: AI in OHS services · shadow AI

First check where the data goes, how long it is kept and whether it is used for training; health data must be stored with a certified health-data host (HDS). Share only what is strictly necessary, with nothing identifying.

Source: Understanding AI · from model to system

Do not use the result as it is: ask for a competent review or keep the task out of the tool. Producing a result with AI does not mean being able to judge its quality.

Source: Reading · Occupational boundaries

Check references, figures and legal articles, which may be invented. Be wary of an answer that confirms your hypothesis too quickly, and of instructions hidden in a document the tool reads.

Source: Understanding AI · capabilities and limits

Report it to your manager: checking and correcting are part of the work and must be counted. How the time saved is used should be decided collectively, before it is turned into targets.

Source: Psychosocial risks · intensity

Keep doing some tasks without AI assistance, especially early in your career, and ask for a fallback procedure for days when the tool is unavailable or makes mistakes.

Sources: Assess an AI project · Reading · Human in the loop

Ask for clear information on the purpose, data, criteria and known limits. A system for monitoring activity requires employees to be informed beforehand and the works council to be consulted.

Labour Code L. 1222-4L. 2312-38

Source: Reading · Human in the loop

Ask for a human review, with facts to support it, and get help from your employee representatives. The GDPR regulates decisions based solely on automated processing.

GDPR art. 22

Sources: Psychosocial risks · legal framework · Assess an AI project

Your professional judgement must be able to prevail over the AI output. Report this pressure to management and the works council: the right to correct, refuse or report without penalty should be written down.

Source: Psychosocial risks · autonomy

Flag that the difficult cases are being concentrated on you. The split between people and AI must take the emotional load into account: quick access to a colleague, alternating with less exposed tasks, debriefing after an incident.

Source: Psychosocial risks · emotional demands

Emotion recognition in the workplace is prohibited by the European AI Act, except for medical or safety reasons. Alert your employee representatives.

AI Act art. 5

Source: Psychosocial risks · legal framework

Do not approve what you cannot check. Describe the gap between what is asked and what is feasible: volume of approvals, missing information, responsibility not written down.

Source: Reading · Human in the loop

This is a sign of approval fatigue, which comes from the organisation more than from you. Ask for breaks, rotation, grouped approvals and shorter checking sequences.

Source: Reading · Human in the loop

Turn off notifications outside working time: an agent blocked for a few hours can wait. Ask for the right-to-disconnect rules to cover requests from AI too.

Labour Code L. 2242-17

Source: Reading · AI agents and disconnecting

Put these difficulties into words and talk about them with colleagues, your manager, employee representatives or the occupational health service. You can ask to see the occupational physician.

Source: Psychosocial risks · six dimensions

Keep reporting errors and incidents and taking part in feedback sessions: your observations help adjust the tool and the organisation.

Source: Homepage · employee profile

De-identify data before anything is sent to a model, and work in an authorised environment (for health data, certified health-data hosting, HDS). To test an online tool, use fictitious texts, like the 540 workplace recommendations in the Préconisations test bench.

Sources: Publication · AI and medical recommendations · Préconisations tool (in French)

Model performance

Test on cases representative of the real task: frequent errors and serious errors, stability from one answer to the next, resistance to misleading inputs, cost and turnaround time.

Effects on work

Evaluate the work system: time saved including review, final quality, exceptions, responsibilities, workload, autonomy, skills and cooperation, before and after implementation.

Source: Understanding AI · back to real work

Check that it resembles the target work; describe the instructions, tools, number of attempts and version; specify what the measure captures (accuracy, preference, cost); look for data seen during training and for errors hidden by the average.

Source: Understanding AI · reading benchmarks

Build a reference (multidisciplinary consensus, expected rating), tune the instructions on a first sample, then test on a fresh sample. Record the model version, the prompt and the date: each version is a dated object.

Sources: Publication · AI and medical recommendations · Understanding AI · capabilities and limits

Classify each discrepancy: justified alert, excessive flagging, missed defect or invention. A disagreement can also reveal a defect the reference had not spotted.

Source: Publication · AI and medical recommendations

Compare them with an observed measure. In a randomised trial by METR, experienced developers believed they were 20% faster with AI; they were measured as 19% slower.

Source: Understanding AI · perceived and measured gains

Break them down: the capability frontier is jagged, and a positive average can hide negative effects. Also look for the organisational changes that explain a favourable or unfavourable effect.

Sources: Reading · AI and real work · Understanding AI · jagged frontier

Treat them as jobs in their own right and measure the control workload: volume to review per person, time actually available before a release, staff mobilised during an investigation, on-call duty.

Source: Reading · Occupational health and AI safety

Study the state of the supervisors alongside that of the system: a tired supervisor approves faster, and the system can learn to produce what is easily approved. Their health becomes a reliability variable.

Source: Reading · Human in the loop

Include it: safety requires that someone can say they are not sure or that things need to wait. An unspoken doubt is a lost safety signal.

Source: Reading · being able to say no

Include them: their judgements are used to train and evaluate the models, but they often fall outside the occupational health monitoring and protections that employees have.

Source: Reading · Occupational health and AI safety

Code the testimonies using an explicit grid, such as the six Gollac dimensions in the October 2026 preprint, and draw no causal inference from them. Independent access to deployment data has yet to be obtained.

Sources: Publication · preprint · Psychosocial risks · state of the evidence

Bring it back to what is measured: spotting wording defects does not demonstrate a benefit for health or job retention. State what the study does not measure, and for which model, prompt and corpus it holds.

Source: Publication · AI and medical recommendations

Plan longitudinal studies or intervention trials, with independent access to deployment data. Testimonies cannot estimate a prevalence, and neither testimonies nor a cross-sectional study can establish causality.

Sources: Psychosocial risks · state of the evidence · Reading · AI workers