AI-assisted ergonomics: when practice automation raises the question of professional value

Accepted as an oral paper by the scientific committee of the 59th congress of the French-speaking Ergonomics Society (SELF, Nantes, September 2026). Not presented, not published in the proceedings. PDF version, in French.

If a machine can propose a pre-diagnosis whose structure meets mine, what grounds the ergonomist’s value? This text recounts a practice. It does not generalize it. Since 2022, across about forty missions, I record, I have the interviews transcribed, a language model proposes a structure and, on about twenty of them, I compare. Processing no longer fills the diary. The time given back goes to the relationship and to debate, when deliverable pressure no longer forces me to cut it. As soon as the situation is relational, the model does not see what organizes speech. The value is not in processing the data. It is in the time that processing used to occupy, and in what the transcript still does not contain.

The question

It would be tempting to pose the question as a technical problem: can AI replace the ergonomist? But that wording sets aside precisely what deserves to be examined.

Since 2022, I have integrated artificial intelligence across my whole intervention process, from demand analysis to support for implementation. What this experience has taught me is not about the tool. It is about the profession: about what occupied time without being its core, and about what made up that core without ever receiving the time it would have deserved.

What I can recount is narrow: my own practice, about forty missions between 2022 and 2026. The largest is a mission for a local authority. I conducted it alone. There is no other reading of the material. Nothing that follows transfers as is to another intervention.

On about twenty of these missions, I compared the structures proposed by AI with my own analyses, conducted on my side. Structure, identification of determinants, ranking of issues: the two converge, in my reading, except where the situation is first relational. I come back to this below. I made this comparison alone. That is its main limit: it says nothing about what another practitioner would have concluded from the same traces.

If a structure can be proposed without my writing it, the question is no longer whether the machine does the job. It becomes: what grounds the value, once processing no longer fills the diary?

What I do

The constraint does not disappear. It shifts.

I record the interviews, systematically. Photographs and video too. Freed from the pen, I listen differently. Transcription is automatic, by Whisper (Radford et al., 2022). Optical document recognition and image analysis follow, when the material is there. What training teaches little, and what independent practice endures, is this passage from raw material to something usable. It consumed a considerable share of the time.

Consent is collected in two steps. From the framing of the mission, with the client, who knows that the interviews will be recorded and processed with an AI tool. Then orally, before each interview, with the person themselves. Before anything is sent to a remote model, the transcripts are pseudonymized locally, on my own machine.

A language model then proposes a pre-diagnosis: determinants, a ranking of issues. They are the models anyone can use, taken as they are. They have changed several times since 2022. I set their proposal beside the analysis I conduct on my side.

On part of the missions, I keep an unassisted diagnosis. This is not a habit kept by inertia. It is a deliberate choice, the condition for supervision to remain supervision and not blind delegation.

The instrumental approach (Rabardel, 1995) describes what happens better than a split of tasks between the human and the machine. The model is only an artefact. It becomes an instrument through the utilization schemes I have built around it: record systematically, compare, keep a share of diagnosis without assistance. This instrument reorganizes the activity: what I look at during the interview, what I can defer, what I hold from one meeting to the next. It does not take the place of judgment. It changes the moment when that judgment takes place. This way of instrumenting the intervention is what I call AI-assisted ergonomics. The term names a practice, not a model.

When the diagnosis holds, it does not hold in the proposal. It takes place between the traces, that proposal, what I keep of it, and what the interlocutors do with it when it is put to debate (Hutchins, 1995). I describe a course of events. I do not draw a law from it.

What has shifted

A single occupational risk assessment document used to take me six to ten hours. It now takes me two to four. I can now offer it to very small companies. The figure is mine. I do not give it as an average for the profession.

The mission for a local authority would classically have mobilized several ergonomists. I conducted it alone. This is not proof that the diagnosis gains from it. It is an organizational fact: the volume to transcribe no longer decided the size of the team.

On missions with a strong psychosocial component, the pre-diagnoses lack nuance on relational dynamics. Spotting the unsaid, implicit alliances, tensions that are never stated directly but that nevertheless organize everything that is said: the recording does not have them. Neither does the model.

An employee avoids a subject. They circle it, come back to it, stop short. In the transcript, detours and unfinished sentences remain: the model can interpret them as it wishes, nothing obliges it to read an avoidance there. During the interview, what alerted me was what the employee was showing without saying it. I asked the question again later, differently. The subject was sensitive, and it deserved to enter the mission.

The gap between prescribed and real work, turned back toward the practitioner, is not closed. Intervention as training describes it assumes time for collection, processing and analysis that the reality of independent practice often does not allow. The instrument empties the processing slot. It does not see, in my place, what is singular about the situation.

The other person’s time

A department head needed three to four meetings before saying what really mattered. Whisper did not reduce that number. It could not have: that time belonged to the department head.

The usual pressure of the deliverable would have made this temporality impossible to hold. It no longer forbids it. One more meeting costs me less, because processing the previous one no longer blocks the calendar. The tool does not make people speak.

In the feedback session, attention concentrates on facilitation, group dynamics, what is politically at stake in putting work to debate. I do not know in advance what will need to be done in that hour. That is why it cannot be delegated.

What I cannot conclude

The ironies of automation (Bainbridge, 1983) are the warning. The supervisor freed from part of the execution must still maintain what that part made it possible to exercise. Automating part of one’s practice is to risk losing what that part sustained. I have not measured that loss.

Two readings remain open. One says that judgment finally receives the time that processing used to take from it. The other says that this time was also the place where the profession maintained itself. My missions lean toward the first. They do not demonstrate it.

What is missing, for this question to receive an answer that goes beyond testimony, is a comparison between ergonomists who instrument and ergonomists who do not, on missions of the same nature and the same complexity. I have not done it.

The value

Evaluating work, Hubault and du Tertre (2008) remind us, means “measuring, sometimes; judging, always”. Applied to my practice, this gives a narrow answer. The value of the intervention is not in processing the data. That processing was indispensable, time-consuming, little taught. It filled the diary. Once that slot is empty, what becomes visible is the time it occupied, and what the transcript still does not contain.

A director needed three to four meetings. No transcript would have sped that up. Whisper made it tenable not to cancel the next one.

References

• Bainbridge, L. (1983). Ironies of automation. Automatica, 19(6), 775-779.
• Hubault, F., & du Tertre, C. (2008). Le travail d’évaluation. In F. Hubault (Ed.), Évaluation du travail, travail d’évaluation (pp. 95-114). Octarès.
• Hutchins, E. (1995). Cognition in the wild. MIT Press.
• Rabardel, P. (1995). Les hommes et les technologies : approche cognitive des instruments contemporains. Armand Colin.
• Radford, A., Kim, J. W., Xu, T., Brockman, G., McLeavey, C. and Sutskever, I. (2022). Robust speech recognition via large-scale weak supervision. arXiv. https://arxiv.org/abs/2212.04356

I will present this text in a webinar. To receive the link, get in touch. It extends my opening note, What the AI tool displaces.