There's a sentence you find in almost every AI policy written in the last two years. Humans keep the final word. The ultimate responsibility stays human. The decision is never delegated to the machine.

Pierre Plouzennec, a foresight specialist at Michelin, puts his finger on what often makes that sentence hollow:

Humans have the final word, but if I choose an option against the AI and my manager tells me off, then the idea that humans have the final word is just wishful thinking.

That's the real test. Not what's written in the charter. What actually happens to the person who contradicts the machine.

This quote comes from a study published in May 2026 by Projet Sens, a collective of seventeen major French groups representing 1.5 million employees. Sixty field managers testified, twenty-five HR directors worked on the conclusions. Nothing comparable exists in Mauritius today, and I won't pretend otherwise. But the mechanisms described, I recognise them in the conversations I have here every week.

The automation bias doesn't show up in an org chart

It has a name in cognitive science: the psychological tendency to place excessive trust in the suggestions of automated systems.

Ludovic Demierre, HR director of the Vinci group, describes how it shows up day to day:

The big risk is the impression that what it states is true. That risk will be found in our companies: it seems so right and so true that you lose your critical judgement.

The interface has a lot to do with it. By mimicking the codes of human conversation, it disarms vigilance. An engineer interviewed in the study puts it well: this kind of false friendliness, he sees it as a trap. The legal scholar Alain Supiot speaks of a sophisticated stochastic parrot, useful as long as you don't take it for a human being.

For an HR director, the problem is very concrete. A recruiter who receives an AI-generated candidate scoring isn't arguing with an algorithm. They're arguing with something that writes better than they do, faster than they do, and that never says "I don't know."

No one has ever sued a calculator

On what cannot be entrusted to AI, the four groups of managers in the study converge on two things: the final decision and responsibility.

One manager gives the analogy that closes the debate:

An engineer doing a stress calculation on which railway safety depends, if the calculator he misused gives a wrong result, no one would ask whether it's the calculator's fault. For AI, it's the same.

Laura Chaubard, director general of École Polytechnique, draws from this the organisational requirement: every organisation must keep a clear, functioning chain of human responsibility. Whether the decision is informed by AI or not, a human must own its consequences. She sees it as a safeguard against the dilution of responsibilities and the robotisation of governance.

Translate that to a dismissal file, a performance review, a rejected application. If the tool produced the analysis and no one can explain the reasoning in their own name, the chain is already broken. The problem isn't legal in the first place. It's managerial.

A chain of responsibility only holds if error is allowed

Philippe Bru, HR director of the SNCF group, takes the logic to its end:

If the company wants to mature in that sharing, it will have to accept, despite AI, the right to make mistakes. Because if we don't accept it, we'll kill everything that makes the magic of innovation, of intuition. We'll end up with robotised humans in our jobs, totally formatted, and that's not what we want.

The right to error isn't an HR favour. It's a condition of production. Without it, no one will ever make the unprecedented decision, the one that past data only partially illuminates, the one that precisely cannot come from the machine.

And that's where the subject becomes very directly HR. An employee who knows they'll be punished for getting it wrong while contradicting the tool will follow the tool. Always. You'll have a perfect charter and an organisation where no one decides anything anymore.

Choosing the criteria, not producing the answer

What remains is knowing what the human skill actually covers. A training manager frames it better than any competency framework: learning AI is relatively simple, the real challenge isn't there. What matters is what will make the difference between two people who both master the tool.

An anecdote from the study answers the question. The son of an executive, a military officer, uses an AI to compare weapons. The AI produces an excellent technical comparison. His superior immediately points out the gap: the reliability criterion, the one that should have been prioritised for that specific mission.

The AI had answered the question asked perfectly well. It couldn't know it wasn't the right question.

What matters isn't so much being able to redo a calculation by hand, but knowing which criteria allow you to make a decision.

That's the skill to recruit and develop. Not producing. Not verifying. Choosing the criteria. For a Mauritian HR director revising their evaluation grids or job descriptions in 2026, that's probably the most useful line to add.

Those who write the rules are those who don't use the tool

One last point, almost comical if it weren't so widespread. A manager recounts an internal survey: asked who diligently uses AI daily on a personal level, 50% of employees answer yes. Among executive committees, 25%.

His conclusion fits in one sentence: the rules come from people who don't have the usage.

The gap is measured elsewhere too. According to McKinsey, executives estimate that 4% of their employees use AI for about 30% of their tasks. The reality is 13%. Three times more.

Writing an AI policy without practising AI is legislating on a territory you haven't visited. And in the meantime, clandestine usage continues. The CFDT observes a detail that should interest every HR director: AI doesn't appear as giving additional skills, but as revealing a lack of skills. It's used without saying so because it's a bit shameful.

If your teams hide their use of AI, you don't have a tool problem. You have a psychological safety problem.

Where to start

Three questions to ask in your executive committee, in this order.

Who, in this room, actually uses AI every week? The answer conditions the credibility of everything that follows.

When someone contradicts a recommendation produced by a tool and gets it wrong, what actually happens to them? If the answer is vague, your chain of responsibility is decorative.

On which HR decisions have we decided not to delegate, and is that written somewhere? CV screening, evaluation, compensation, termination. Those four deserve a clear position.

The dividing line on AI doesn't run between those who are for it and those who are against it. It runs between technological determinism and voluntarism. The slope of this technology exists. It's not a fate. It calls for choices, and those choices are yours.

Em 🌱

Sources

• Jean-Baptiste Barfety, "Working with AI, deciding among humans", Projet Sens, 19 May 2026. Study conducted with 63 managers, 25 HR directors and 26 experts, sponsored by Laura Chaubard and Jean-Dominique Senard. projet-sens.fr
• Alain Supiot, Collège de France, interview cited in the Projet Sens study (2026).
• McKinsey & Company, "Superagency in the Workplace: Enabling the AI-powered organization", QuantumBlack / McKinsey Global Institute, 2025.
• CFDT, contributions by Charles Parmentier and Luc Mathieu, cited in the Projet Sens study (2026).
• All manager quotes come from the anonymised focus groups of the Projet Sens study.