AI Is Not Integrated Into Work. It Reconfigures It.

2026-04-24 · Work Systems

Artificial intelligence is being discussed as a tool to be adopted.

But one fundamental question is being forgotten:

What happens to real work when AI enters activity?

Not the work described in procedures. Not the work imagined in strategic presentations.

Real work — the work that takes place amid constraints, trade-offs, interruptions, emergencies, grey areas, errors, rework, arrangements, invisible cooperation, and room to act.

That is where AI truly transforms organizations.

And that is where activity ergonomics becomes indispensable.

When an AI agent is introduced, it does not simply add capacity. It changes the prescription of work.

It changes what is expected, what can be delegated, what must be verified, what becomes invisible, and who is considered responsible when the system produces a result.

This creates a new gap: the gap between the prescribed agentic activity and the effective activity that actually happens in the runtime and in the workplace.

The prescription may say: “the agent drafts, the human validates.”

The real activity may be: the human reformulates the request, checks the agent’s assumptions, investigates tool calls, compares outputs with reality, corrects errors, explains the result to others, and carries the risk when something is wrong.

That gap is not a detail. It is the transformation.

This is why I argue for Agentic Activity Ergonomics: an ergonomics capable of studying both sides of the system at once.

On the human side, we need to understand attention, cooperation, responsibility, interruptions, room to act, and the concrete possibility of intervening.

On the agent side, we need to understand action loops, tool use, runtime constraints, evidence, recovery, and the truthfulness of final claims.

The practical question is not whether AI is “adopted”.

The practical question is whether the reconfigured activity remains understandable, discussable, correctable, and governable.

For laboratories, this means evaluating more than final answers. It means evaluating whether agents can connect claims to actions and observable evidence.

For organizations, it means refusing fake supervision. A human in the loop is meaningful only if that human can see enough, understand enough, and act early enough.

For designers, it means making trajectories inspectable instead of hiding the work behind polished outputs.

The central sentence:

AI is not integrated into an organization: it reconfigures an activity. And every reconfigured activity must be understood, discussed, and regulated from the standpoint of real work.

This is where the quality of future transformations will be decided.

— Julien Talbot