AI not only transforms tools: it redefines autonomy, expertise and the collective. The challenge is no longer just to govern the models, but their human and organizational impacts.
Over the past two years, large organizations have structured their artificial intelligence governance around models, data and compliance. This is progress. But that is no longer enough. The real test is not played out in the AI committees, it is played out in the actual work of the teams and in the way in which employees will experience this transformation. And in this area, we are already late.
Act or suffer, the new dividing line
Generative AI establishes in organizations a line of demarcation that is more discreet than the digital divide of yesterday, and deeper. It no longer separates those who have access to technology from those who are excluded. It separates those who act with AI from those who experience it.
To act is to maintain control over the question asked, over the decision taken, over the meaning given to what we do. To submit is to simply accept the result that appears, because it arrives quickly, because it seems rigorous, because it eliminates the need for analysis. The boundary between the two is not a boundary of jurisdiction. It is a boundary of intention, and it crosses every team, every job, every day.
This dividing line is new because the technology that draws it is also new. Generative AI is not another tool in the production chain. Innovation economists call it general-purpose technology, like printing or electricity. Such technology does not add to the work. It reconfigures it, and with it the balance of expertise, authority, power to act.
Acting requires thinking, and thinking requires time that the machine shortens
There is no action without thought. And there is no thought without time which AI, by construction, shortens. This tension is the heart of the debate. It is too rarely formulated as such.
When an AI assistant produces in seconds what took hours to build, two movements coexist. Freed up time, that’s a fact. But also a fatigue of analysis which fades, and an expert gesture which no longer needs to be performed. By delegating, we stop practicing. The first work from the MIT Media Lab published in 2025, certainly preliminary and therefore to be handled with caution, points to a measurable drop in brain connectivity among intensive users of language models in writing tasks. Occupational psychologists have observed for a longer time a related phenomenon regarding the feeling of personal effectiveness, which erodes when expert gestures disappear.
The diagnosis is not new. La Boétie, in the 16th century, was surprised that free men voluntarily accepted to be dominated not by force, but by the comfort of no longer having to decide. Erich Fromm, in the 20th century, spoke of the fear of freedom, demanding, uncertain, empowering, in the face of which delegation provides relief. AI presents all the characteristics of an ideal interlocutor for this delegation movement: available, seemingly infallible, never judging, sometimes even flattering. The relief it provides is real. This is what makes it strategically risky.
An organization that wants to act, and not suffer, must therefore protect a space that no one spontaneously protects: the one where we continue to think, sometimes slowly, sometimes seemingly uselessly, because it is in this space that judgment is formed.
The collective that silently fragments
The other shift, even more discreet, concerns the collective. The AI responds quickly, without mood, without agenda, without the need to formally request it. She becomes the natural interlocutor for the question that would have been asked of the colleague, the expert, the manager. This shift is not visible on the production indicators. It is ultimately seen in the living memory of the organization.
When each employee has a universal assistant at their fingertips, the organization runs the risk of transforming into a juxtaposition of brilliant soloists, equipped with the same cognitive prosthesis and working side by side without any longer needing to speak to each other. This is the scenario of an omnipotent individualism, where everyone experiences themselves as an autonomous productive center, where dependence on the collective is perceived as slowness, where deliberation appears as a cost.
However, real expertise, that which keeps an organization going in difficult times, has never resided in the sum of individuals. It lies in the capacity of a collective to debate, to arbitrate, to transmit. The work of Amy Edmondson and Google’s Project Aristotle have been converging on this point for ten years: psychological safety, being able to say what you don’t understand, report an error, express disagreement, remains the primary factor in team performance. It is today the most exposed factor. And it’s rarely what AI project dashboards measure.
AI as a lever for engagement, or as an accelerator for withdrawal
The choice presented to leaders is not a technological choice. It is anthropological, in the literal sense of the term: what becomes of work, and what becomes of the one who works, in this transformation.
AI can be a powerful lever for engagement. It can free up time for the benefit of what is least easily delegated: deciding in the face of uncertainty, taking care of the relationship, passing on a job, debating a course. It can give managers the room for maneuver that operational pressure has deprived them of for ten years. It can raise the level of collective play, on one condition. That we explicitly choose to reinvest the time it frees up in what keeps a collective going, and not in a simple intensification of the pace.
It can also produce the opposite. When AI is introduced as a gross productivity gain, without link to the notion of work, without recognition of the profession, without dialogue with teams, it accelerates withdrawal. Everyone makes arrangements. Uses are developing alongside the operation of the company. The manager improvises. And the Codir discovers the effects when they are already installed. European regulators set a deadline: the AI Act becomes fully applicable on high-risk systems on August 2, 2026; but no regulation will replace, on this subject, a lucid internal governance decision.
Governing impacts, just like models
The four requirements that are imposed today on the governance of AI models (traceability, perimeter, reversibility, measurement) have their exact symmetry on the team side. Trace, profession by profession, what AI supports and what it moves in the chain of expertise. Explicitly decide what you will not delegate, in dialogue with the employees concerned, because the decision is not technical. Preserve the reversibility of the organization, that is to say the spaces of transmission, the collective rituals, the areas where we learn by doing. Measure, over time, what affects commitment, the feeling of usefulness, and the quality of the collective.
These requirements are not the exclusive responsibility of HR managers. They report to the Management Committee in the same way as cybersecurity or compliance, because they touch on an asset that the balance sheets never name and that organizations quickly lose: the collective ability to think together.
The question facing managers is no longer whether AI will be in the company, it is already there whether we have decided for it or not. The question now is what we want to continue doing ourselves, because we consider it important. It is this choice, and it alone, which separates an organization which acts from an organization which suffers.
We know how to govern models. We still have to govern what they do with us.
