PST 5 calls for evidence on AI as a tool for preventing occupational risks and improving working conditions, while also examining AI systems and their uses as possible sources of occupational risk.
OHS services · Professional uses · Human oversight
AI can support an OHS service when the task is precise.
Reviewing a document, finding a source, structuring observations or preparing training: value does not come from “AI” in general, but from a limited use that can be tested and corrected.
An evaluated example. A study published by INRS in June 2026 tested a language model as a second reader of occupational physician recommendations. It shows genuine potential, but not an ability to replace the occupational physician.
Limited taskHuman referenceDecision retained
Between two sides
The Oxbow · 1836
Between the storm and the valley, the landscape rests in an uncertain balance. The work reminds us that every promising use must be considered from both sides: what it improves and what it may make fragile if its effects are not observed. View the painting at The Met ↗
PST 5 · 2026–2030
France’s national plan makes AI at work an occupational health issue.
Under Objective 3, France’s Occupational Health Plan 2026–2030 states that AI-related transformations should be introduced through consultation, protect working conditions and be assessed through their real effects. Sub-action 3.1.1 provides for a national Observatory for AI at Work, led by Anact.
The plan provides for quantitative and qualitative studies, experiments, training, webinars, and practical tools and methods that can turn findings into prevention measures.
Annex 3 calls them “pivotal actors”. It asks them to contribute evidence on AI with Présanse and Anact, while adopting a measured use of AI in their own activities.
For an inter-company occupational health service (SPSTI), the task is therefore twofold: test clearly bounded professional uses with care, and observe what AI deployment changes in client organisations — workload, autonomy, skills, teams and occupational risks.
Source: France’s Occupational Health Plan 2026–2030, Action 3.1.1, pp. 44–45, and Annex 3, pp. 154–156. Read the official document (French) ↗
01 / Where AI may help
Six concrete uses with clearly defined human responsibility.
Except for the review case presented below, these uses are possibilities to experiment with and assess in each service.
Review the wording of a recommendation.
Identify ambiguous wording, an implicit obligation, unnecessary information or a disclosure risk before transmission.
Humans retain: work-post analysis, final wording, decision and signature.Query internal documentation.
Retrieve a protocol, risk sheet or procedure from a controlled repository, with sources shown in the answer.
Humans retain: verification of the source, its date and its applicability to the case.Structure field observations.
Group notes by theme, prepare a synthesis and surface missing information before collective analysis.
Humans retain: observation, interpretation, prioritisation and proposed measures.Prepare cases and materials.
Produce variants of a situation, discussion questions or an initial outline adapted to participants’ occupations.
Humans retain: objectives, content validity and teaching.Identify incomplete files.
Check for expected sections or flag a formal inconsistency without automatically assigning a risk level to the worker.
Humans retain: access to the file, interpretation and any resulting action.Reduce repetitive tasks.
Prepare a message, classify a non-medical request or rephrase practical information in more accessible language.
Humans retain: exception handling and validation before sending.What must not be automated
Fitness for work, clinical direction, individual worker assessment or a prevention decision cannot be inferred from a model output without professional judgement and the work context.
02 / An example published by INRS
AI-assisted review, evaluated on one precise task.
I authored this study, published by INRS in June 2026. It evaluates support from a language model without delegating the consultation, the drafting of the recommendation or the medical decision.
Références en Santé au Travail · TF 335 · June 2026
The model acts only as a second reader.
The o1 model reviewed 385 medical recommendations that had already been written and contained no identifying information. It looked for five predefined defects: imprecision, ambiguity about obligation, information outside Annex 4, an implicit job change and disclosure of medical or personal information.
Essential limitation. The study measures detection of wording defects. It does not assess consultation quality or effects on health and job retention.
Read the full article on the INRS website ↗Overall agreement
The model and the multidisciplinary consensus reach the same assessment.
Missed errors
The consensus identifies a defect that the model did not flag.
Observed hallucinations
In this sample and under this precise protocol; the result is not a general guarantee.
Cautious interpretation
These results are promising for tightly controlled review support. They concern one model and one task, with no measurement of clinical effects. The occupational physician remains responsible for analysis, wording and validation.
03 / Build a pilot in an OHS service
Evaluate the use before trying to scale it.
A useful pilot begins with a measurable question and a human reference. It must also allow a return to the previous way of working.
Choose a narrow task.
A recurring, correctable problem distinct from the clinical decision.
Establish the reference.
A set of cases assessed by professionals using explicit criteria.
Protect the data.
Minimisation, appropriate hosting, limited access and no unnecessary identifiers.
Test in parallel.
The model acts as a second reader without changing the usual process.
Measure errors.
Agreement by criterion, over-detections, missed defects, time saved and new review workload.
Decide collectively.
Continue, correct or stop based on documented results and effects on work.
04 / Allocate responsibilities
A service-wide project, not a tool left to a few users.
Professionals must be able to define the criteria, limits and situations in which the system must not be used.
ContributionDefine cases, the professional reference and unacceptable errors.
ResponsibilityRetain clinical interpretation and validation of every output used.
ContributionDescribe work situations and the pilot’s organisational effects.
ResponsibilityPrevent an automated synthesis from replacing field analysis.
ContributionIdentify repetitive tasks, exceptions and information needs.
ResponsibilityReport the actual correction workload and errors affecting service users.
ContributionOrganise the contract, data, access, security and monitoring.
ResponsibilityProvide the means to stop use and inform the relevant bodies.
ContributionExamine consequences for occupations, workload, autonomy and skills.
ResponsibilityRequest a documented review before any wider rollout.
Use in OHS services · Training · Research
Would you like to discuss a pilot or the study?
Describe the intended task, professionals concerned, data used and what you want to measure. The discussion may cover use-case selection, the evaluation protocol or effects on work.
Professional contact on LinkedIn · article available on the INRS website