A real task
Describe the current workflow, pain point, expected benefit and what will stop being done if the system is introduced.
Thomas Cole, The Arcadian or Pastoral State, 1836
II. Uses & fieldUse AI to support prevention.
Use AI to support prevention.
Not to replace professional responsibility.
AI systems, work organisation and occupational health.
Defensible uses start from a precise professional need and a clearly assigned human responsibility. Their value must be tested in real work, not inferred from a technical demonstration.
In 30 seconds
Essential points
- A useful system addresses a real task, leaves a visible audit trail and allows a competent professional to reject its output.
- Risk rises with sensitive data, automated influence over people and the difficulty of detecting an error.
- Deployment evidence should compare the stated objective with actual effects on work, staffing, quality and skills.
EstablishedEmergingAnalysis
On this page
The French edition contains the complete study description, use cases and source list.
Selection
Criteria for a defensible use
An AI use is easier to justify when the task is clearly delimited, the output is advisory, the professional remains able to inspect the evidence, and the consequences of an error are limited. A vague promise to “increase productivity” is not a sufficient use case.
Verifiability
The user must have enough information and time to check the output. Verification time belongs in the workload calculation.
Reversibility
A pilot should allow withdrawal without losing essential skills, records or organisational capacity.
Published work
AI as a second reader rather than an autonomous decision-maker
The author’s published work examined the use of a language model as a second reader of occupational-health recommendations. The relevant lesson is not that a model can replace a physician. It is that a narrowly defined review task may reveal inconsistencies or prompts for reconsideration when the physician keeps the final decision and checks every alert.
Bounded role
The model flags possible issues; it does not issue the medical recommendation.
No silent automation
The professional must see what the model identified and why the final wording was retained or changed.
Local validation
Performance measured on one corpus should not be assumed to transfer to another service, speciality or model version.
Risk-based use
Levels of vigilance
Uses can be grouped pragmatically according to the data involved, the degree of autonomy and the consequences of an error. This is an operational orientation, not a substitute for legal classification.
Lower vigilance
Drafting from public information, formatting, brainstorming and non-sensitive summarisation with systematic review.
Intermediate vigilance
Internal document analysis, prioritisation, knowledge retrieval or coding support where confidentiality and error propagation matter.
High vigilance
Recruitment, worker scoring, health recommendations, surveillance, work allocation or decisions with significant effects on individuals.
Field evidence
What early workplace deployments show
Public announcements frequently describe AI as a route to rapid productivity gains. Real deployments also reveal rework, quality problems, employee resistance, unplanned verification burdens, skill erosion and strategic reversals. The unit of analysis should therefore be the whole work system, not the isolated model output.
Promise versus use
Measure who actually uses the system, for which tasks and how often outputs are overridden.
Employment effects
Distinguish positions removed because of AI, positions left unfilled, work transferred to contractors and strategic narratives used to justify restructuring.
Hidden work
Prompting, checking, correcting, documenting and resolving errors may transfer effort rather than eliminate it.
Deployment
Conditions that should exist before scaling
A responsible pilot needs a named owner, a precise purpose, approved data, user training, incident reporting, stop criteria, worker involvement and an agreed review date. Occupational health should be consulted when tasks, pace, monitoring, autonomy or psychosocial exposures may change.
Before
Map tasks, populations, data, expected benefits, failure modes and consultation duties.
During
Track errors, overrides, verification time, workarounds, workload, user confidence and unequal effects between groups.
After
Decide whether to scale, redesign or stop. Preserve the evidence, including negative findings.
Key references
- INRS, TF 335 — use of a large language model as a second reader.
- Full field cases and references are available in the French reference edition.