PST 5 aims to document AI both as a tool for prevention and for improving working conditions, and as a possible source of occupational risks through AI systems and their uses.
Practical guide · OHS services · Sources 2025–2026
AI in OHS services: uses, risks and deployment method.
In an occupational health service, artificial intelligence can assist with documentation, administration, prevention, member relations and certain strictly supervised medical uses. It replaces neither the analysis of work nor professional decision-making.
The common principle. Start with a specific task, use the minimum necessary data, measure errors and retain an identified person who can check the output. Présanse Charter, 2025 ↗
Sources throughoutRisks clearly differentiatedHuman decision-making
PST 5 · 2026–2030
What does PST 5 provide for AI at work?
The national plan makes AI at work a subject for prevention. In Objective 3, the Occupational Health Plan 2026–2030 requires AI-related transformations to be developed in consultation, respect working conditions and be evaluated through their real effects. Sub-action 3.1.1 provides for the creation of a national observatory for AI at work, led by the National Agency for the Improvement of Working Conditions (Anact).
The plan provides for quantitative and qualitative studies, experiments, training, webinars as well as tools and methods to move from findings to action.
Appendix 3 describes prevention and occupational health services as “pivotal players”. It asks them to contribute to knowledge about AI with Présanse and Anact, while developing a considered use of AI in their own activities.
For an occupational health service, the goal is therefore twofold: cautiously test well-defined professional uses and observe what AI deployment changes in the companies it supports — workload, autonomy, skills, work groups and occupational risks.
Source: Occupational health plan 2026–2030, action 3.1.1, p. 44–45, and annex 3, p. 154–156. Consult the official document ↗
01 / What uses of AI in an OHS service?
Five families. Different levels of evidence and risk.
Présanse groups the cases reported in occupational health services into five families. This classification describes a field of uses; it does not demonstrate the effectiveness of each tool. Présanse presentation ↗
Assist, never conclude.
Transcription, synthesis, documentary research or second reading: the level of vigilance increases as soon as health data or an individual decision is concerned.
Uses cited by Présanse ↗Structuring the field.
Company files, exploration of DUERP, grouping of observations or preparation of action plans: each task must be evaluated separately.
Uses cited by Présanse ↗Reduce duplicate data entry.
Internal research, CRM tickets, reporting, planning and assistance with administrative or quality processes.
Uses cited by Présanse ↗Direct requests.
Callbots, appointment scheduling and simple responses, with human response for exceptions and complex situations.
Uses cited by Présanse ↗Monitor developments and train teams.
Scientific and regulatory monitoring, training materials, content and controlled exploration of new services.
Uses cited by Présanse ↗| Usage | Main profession | Status and source | Data and risk | Useful indicator | Supervision required |
|---|---|---|---|---|---|
| Callbot for simple requests | Member relations | Documented field useSPST 19-24 ↗ | Contact data · moderate risk depending on the scenario | Routing errors, drop-offs and human takeovers | Escalation to a human team |
| Training administration | Assistants · support functions | Documented field useMT71 ↗ | Administrative data · low to moderate risk | Corrections, exceptions, end-to-end time | Checking exceptions and outgoing communications |
| Second reading of recommendations | Occupational physician | Published studyINRS TF 335 ↗ | De-identified text · high risk | Defects spotted, missed defects, hallucinations | Medical validation and signature |
| Help with DUERP or company file | IPRP · prevention professionals | To be evaluated by taskUse mentioned ↗ | Field data · variable risk | Omissions, corrections, quality of the action plan | Analysis of the work by the prevention officers |
| Synthesis of medical information | Occupational physicians · occupational-health nurses | Sensitive useCharte Présanse ↗ | Health data · very high risk | Errors, omissions, accesses and incidents | Authorised environment and clinical validation |
| Scoring or individual prioritisation | Management · medical teams | Maximum vigilanceAnalyse the framework ↗ | Individual data · potentially serious consequences | False positives, false negatives, differences between groups | Formal analysis before any decision |
Keep this limitation visible
The presence of a use in a sector presentation establishes neither its performance nor its compliance. The task, data, provider and decision-making process must be examined in each service.
02 / What already exists in France
Three examples from the field, three operational problems.
Présanse presented these experiences in November 2025. They document real approaches but do not, on their own, constitute evidence of generalisable effectiveness. Présanse article ↗
Absorb peaks in call volumes.
A callbot handles simple requests, responds or redirects. The human team remains available for complex situations and requests outside the defined scenario.
To measure: transfers, orientation errors, abandonments, satisfaction and rework burden.
Présanse field account ↗Streamline training administration.
An assistant handles pre-registrations, changes, communications with member companies and duplicate entry of information related to awareness sessions.
To measure: corrections, exceptions, end-to-end time and quality of exchanges.
Présanse field account ↗Create a dedicated AI function.
A cross-functional role drives the transformation, links strategy to the organisation and coordinates compliance, security and prioritisation.
To measure: documented decisions, training, incidents and real capacity to suspend a project.
Présanse field account ↗03 / What professionals say
Interest, with a clear demand for evidence and safeguards.
These two surveys published in 2026 describe declared attitudes. They do not establish the effectiveness of a tool, but shed light on the conditions of acceptance.
Survey of 237 occupational physicians in the Ile-de-France region.
Study 2026 ↗Ahead of the reliability of results (84%) and data protection (81%).
Study 2026 ↗Separate survey of 80 occupational health professionals.
Study 2026 ↗04 / An example published by INRS
An assisted review evaluated on a clearly defined task.
This study, published by INRS in June 2026, evaluates the help provided by a language model without entrusting it with the consultation, the drafting of the opinion or the medical decision. INRS article ↗
In Charles Broutin’s study, “Artificial intelligence in occupational health: evaluation of its potential to strengthen the precision and clarity of medical recommendations”, the model achieved 74.6% overall agreement with the multidisciplinary consensus on 385 recommendations. Références en Santé au Travail, TF 335 ↗ Read the detailed study summary →
Références en Santé au Travail · TF 335 · June 2026
The model only acts as a second reader.
The o1 model reviewed 385 medical recommendations that had already been drafted and contained no identifying information. It looked for five predefined defects: imprecision, ambiguity about an obligation, information outside Annex 4, an implicit job change, and disclosure of medical or personal data.
Essential limit. The study measures the detection of wording defects. It does not measure the quality of a consultation, nor the effect on health or continued employment.
Read the full article on the INRS website ↗
Overall agreement
The model and the multidisciplinary consensus lead to the same evaluation.
Result published ↗
Missed errors
The consensus identifies a defect that the model did not report.
Result published ↗
No hallucinations observed
In this sample and with this precise protocol; this result is not a general guarantee.
Result published ↗
Read with caution
These results are promising for a tightly controlled second-reader use. They concern a single model and a single task, without measuring any clinical effect. The occupational physician remains responsible for the analysis, wording and validation.
05 / Présanse AI Charter
Ten principles to translate into daily decisions.
The charter is presented as an adaptable model that can be integrated into the service’s IT charter. The formulations below give an operational translation. Official page ↗
Identify who oversees and validates the use, who can challenge it and who remains accountable for it.
Principle 1 ↗Do not transmit personal or health data to an unauthorised tool.
Principle 2 ↗Document the tool, its purpose, its limits and the exact role of AI.
Principle 3 ↗Look for biases and differing effects across groups and populations.
Principle 4 ↗Document uses, validations, errors, incidents and corrections.
Principle 5 ↗Define those responsible, the decision process and the stopping conditions.
Principle 6 ↗Start from a real need and choose the least intrusive system possible.
Principle 7 ↗Evaluate the environment, access, errors and dependencies.
Principle 8 ↗Train before use and maintain professional skills.
Principle 9 ↗Do not use AI unless its value has been demonstrated.
Principle 10 ↗06 / Shadow AI in OHS services
The risk exists before official deployment.
Présanse uses the term “Shadow AI” to describe uses that are neither validated nor supervised, including the risk of entering sensitive information into public chatbots. Présanse resource ↗
Weak signal, immediate risk
“I just wanted to save five minutes.”
A vague instruction, a personal tool and a copy and paste are enough for sensitive data to leave the intended framework. The Présanse charter prohibits the introduction of personal, medical or identifying data into a tool not validated by the service. Présanse good practices ↗
Map existing uses.
Understand practices, needs and workarounds without driving them underground.
Define prohibited data.
Give concrete examples for personal and health data.
Provide authorised alternatives.
Meet real needs with a framework that teams can actually use.
Train and document incidents.
Improve the framework based on observed errors and difficulties.
07 / How to prioritise uses?
Occupational value × complexity × data × consequence of an error.
Présanse proposes combining occupational value and complexity. This matrix adds data sensitivity and the consequences of an error. Présanse matrix ↗
Value / risk matrixPriority candidate
High value
Controllable riskFormal assessment
before any pilotSecondary actionSeemingly good idea
Controllable riskFormal assessment
before any pilotSecondary actionSeemingly good idea
All professions · documentation
Internal documentation assistant
Find a protocol or risk sheet in a controlled knowledge base, with sources displayed in the answer.
Occupational value
4
Complexity
2
Data sensitivity
2
Consequence of an error
2
OrientationA good candidate for a limited pilot on a controlled knowledge base, with source citations and human verification.
Illustrative rating out of 5. The actual profile depends on the data, provider, integration, groups concerned and decision process.
08 / Building a pilot in an OHS service
Need → value → data → risk → pilot → decision.
Présanse recommends starting small, measuring impacts, adjusting, then expanding gradually. Présanse approach ↗
Describe the need.
A recurring task, an observable problem and the professionals involved.
Estimate the value.
Expected quality, end-to-end time and benefit for the service's mission.
Audit the data.
Nature, necessity, access, conservation, hosting and minimisation.
Anticipate the error.
What can go wrong, who can see it and who can interrupt use.
Test in parallel.
Compare to a human reference without immediately modifying the usual circuit.
Decide collectively.
Continue, correct or stop based on documented results and effects.
What the tool does.
- spotted errors, missed errors and invented responses
- stability after a change of model or instruction
- time saved after verification and correction
- differences between cases, occupations and groups
How the tool affects teams.
- checking workload, interruptions and exceptions
- autonomy, the ability to challenge and real accountability
- changes in skills and the work group
- incidents, workarounds and inequalities in use
To prepare for this test, use the assessment tool before and after deployment. To work through these concepts in order, follow the AI & Occupational Health pathway.
09 / Assign responsibilities
A service project, not a tool left to a few users.
The Présanse charter requires clear governance and the definition of who controls, validates and monitors uses. Principle of governance ↗
InputDefine cases, professional reference and unacceptable errors.
ResponsibilityRetain the clinical interpretation and validation of any output used.
InputDescribe the work situations and organisational effects of the pilot.
ResponsibilityEnsure that an automated summary does not replace field analysis.
InputIdentify repetitive tasks, exceptions and information needs.
ResponsibilityReport the real correction load and the errors that reach users.
InputOrganize the contract, data, access, security and monitoring.
ResponsibilityProvide the means to stop the use and inform the relevant bodies.
InputExamine the consequences on jobs, workload, autonomy and skills.
ResponsibilityRequest documented feedback before any wider roll-out.
The role of the CSE and the organisation of decision-making are detailed in the law and governance benchmarks for AI at work.
10 / AI training and literacy
Train each profession before deployment.
Useful training is not limited to operating a tool. It connects capabilities, limitations, data, verification, accountability and effects on real work.
Understand before use.
How models work, plausible errors, information protection, recourse, and human accountability.
Train on real tasks.
Use separate cases for occupational physicians, nurses, prevention professionals, assistants, management, the DPO and IT, with the unacceptable errors for each profession.
Maintain the ability to control.
Verification time, right to challenge, reporting of incidents, updating of skills and shutdown conditions.
Article 4 of the AI Act requires providers and deployers to take measures to ensure a sufficient level of AI literacy. The expected level depends in particular on knowledge, experience, training, the context of use and the people concerned. Read the European regulation ↗
11 / Direct answers
Six frequently asked questions about AI in OHS services.
These short answers summarize the decision points. Each statement refers to the corresponding source or detailed section.
What does “AI in OHS services” mean?
This means the use of artificial intelligence systems in prevention and occupational health services. AI can assist with administrative, documentary, prevention and member-relations tasks, as well as certain strictly supervised medical uses. It replaces neither the analysis of work nor professional decision-making. Read the full answer
Which uses of AI are relevant in an occupational health service?
The first candidates are precise and verifiable tasks: internal documentary research, administration of training, guidance of simple requests or preparation of documents. Medical and individual uses require greater vigilance, because the consequences of an error and the sensitivity of the data increase. Compare uses · Présanse ↗
Can we use consumer AI with health data?
No, not if the tool has not been authorised by the service. The Présanse charter requires that no personal, medical or identifiable data be entered into an unvalidated tool. The framework must specify hosting, access, purpose and responsibilities before any use. See the risk of Shadow AI · Présanse Charter ↗
How should an occupational health service choose its first AI pilot?
Choose a task with high occupational value and manageable risk. The pilot must start from an observable need, use the minimum amount of data, make errors visible and be capable of being stopped without disrupting the usual process. Test the matrix · Follow the six steps
Can AI decide on behalf of the occupational physician?
No. In the study published by INRS, the model acts as a second reader of an existing text; it does not conduct the consultation, make the decision or sign the recommendation. The occupational physician remains responsible for analysis, formulation and validation. Read the results and their limits · INRS TF 335 ↗
What should be measured during a pilot?
Measure both the system and the work. Monitor detected and missed errors, stability and end-to-end time after correction, as well as checking workload, autonomy, skills, incidents and inequalities in use. See the indicators · Consult the assessment method
12 / Reference sources
Go back to the original document before deciding.
The links below complete the references placed throughout the page. Official documents and scientific articles remain the basis for any legal, medical or technical decision.
Présanse AI Charter
The ten principles and good practices proposed to occupational health services.
Official page ↗Complete PDF ↗Présanse, November 2025
SPST 19-24, MT71, AST74, use families and the prioritisation matrix.
Article ↗PDF summary ↗Présanse material
Awareness raising, use cases, data, pilot and unsupervised uses.
Consult support ↗INRS — TF 335
Protocol and results on the second reading of medical recommendations.
Consult the article ↗Occupational physicians in the Ile-de-France region
237 respondents, clinical complementarity, validation and data protection.
Consult the study ↗KAP survey
80 occupational health professionals and declared interest in training.
Consult the study ↗Occupational health plan 2026–2030
Objective 3, sub-action 3.1.1 and role of prevention services.
Official document ↗CNIL
Official benchmarks on artificial intelligence and the GDPR.
Consult ↗HDS certified hosting
Repository and list of certified hosts.
Digital Health Agency ↗