10 questions before automating a task with AI

This is not a deployment checklist. It is what you need to be able to answer before connecting a tool to a task. I had announced 12 questions. 10 remain. The other two had nothing to do here once the talk was reread.

The question that carries the others: what was this task preventing from becoming a problem?

Source: talk at Préventica Rennes, 18 June 2026, hosted by CINOV Ouest. How to use it: one real task, 20 minutes, written answers. Only then: move forward, reduce scope, or cancel.

1. Map of the work

1. Did we look at the real work, or only at the job description?

The job description describes what the organisation has managed to make administrable. The work that matters is in what resists: the hazards, what people have to take on themselves, what was not planned. Without observing the activity, you automate a poor map. Not the job.

2. What was this task preventing from becoming a problem?

A company replaces its executives’ assistants with AI agents. A few weeks later, the executives are overloaded. It thought it was removing admin work. It had removed a layer of regulation. Invisible work does not disappear. It bills elsewhere.

2. Solution looking for a problem

3. Is the problem named without AI, or is AI already the answer?

As soon as a tool is highly visible, it falls into the solution-first trap: you have the solution before you have defined the problem. Same pattern as an exoskeleton ordered too soon. Sometimes the assessment does not contain the word AI. That is not a failure. It means the lever was not there.

3. Displacement

4. Who picks up the hazards, the exceptions, the real work the tool will not do?

Working means bending the prescribed task. AI takes over the prescribed task. The real work remains. Often with the operators. Often the most complex part. If no one is named for it, someone will do it anyway. Without it being in their job description.

5. Does the remaining job become continuous-attention monitoring? Who carries that?

An AI that acts does not detect its own error on its own. Someone has to watch. Maintaining that attention is not the same activity as being alerted from time to time. It costs much more. If the remaining job is constant surveillance, say who holds it, for how long, and what happens the night nobody is watching.

4. Meaning / skill

6. Is the interest of the job in what you automate, or in what you leave?

On a foundry line, the interest of the work was not the chisel. It was in defect control. Automating the control, leaving the rework: you remove the meaning, you increase the strain. Same pattern in a knowledge job: you remove “what’s simple”, what’s left is the hassle — without always knowing where it comes from.

7. Is this task still what lets people learn the job?

We automate “no added value” tasks so people can focus on what matters. Sometimes we remove the material that let them grow into the role. A junior and a senior can write the same prompt. Only one still has the eye to see it is hollow.

5. Proof / regulation

8. How will we know it is really done — and who still has the eye to see it is hollow?

The tool gives an illusion of expertise. If you are not from the trade, you do not see the gap. “It’s done” is not proof. It is a sentence. Name the observable proof, and the person still able to read it.

9. Can people still work things out when it drifts, or does no one negotiate with the machine any more?

With a human, you work it out. With a station that only executes the prescribed task, you don’t. What five workstations used to regulate now has to hold on two. If there is no longer a way to deviate, correct, or catch up: don’t automate yet.

10. If everyone uses it, what are we accepting to lose — and where is that discussed?

Consulting, marketing: the output becomes smooth, everything starts to look alike. You don’t notice it at your own desk. You see it across ten. The projects that showed a return were not the ones that had stuck a trade-specific AI onto an org chart. They were the ones that let people make the tool their own, with a framework to say what works, what doesn’t, and what we refuse to lose.

After the 10 answers

Move forward, reduce the scope, or cancel. If you cannot answer 2, 4 or 8 in writing, it is not yet the time to automate.

---

Julien Talbot · Ergonomia · Préventica Rennes · 18 June 2026 · CINOV Ouest