Thinking about the enterprise in the face of AI is neither an HR matter nor an IT matter. It is a strategic architectural work that falls to the leader.
Published on 2026-05-15 by Nathalie Lamborghini Dumas.
Thinking about the enterprise in the face of AI is neither an HR matter nor an IT matter. It is a strategic architectural work that falls to the leader.
The AI conversation in executive committees almost always forms around the same question: which tasks to automate, which tasks to augment? The grid is reassuring. It carves work into measurable units and feeds decision files. But it never says why a role exists.
David Autor, an economist at MIT, shows in his work on automation that jobs are assemblies of tasks, and that what determines their value lies in what remains in the assembly once certain tasks are removed. An accountant whose data entry is automated refocuses on tax advice and gains value. A taxi driver whose knowledge of the streets has been absorbed by GPS sees the job open up to everyone: more drivers, lower income. The same tool enriches a role or impoverishes it, depending on the task it removes.
But this reading applies to a job taken in isolation. A leader does not steer a job, they steer an entire enterprise, made of roles that depend on one another. At that scale, reasoning job by job is no longer enough.
Sangeet Paul Choudary, a specialist in platform models and author of Reshuffle, starts from an observation: a role in an organization was not designed to execute a list of tasks, but to resolve a constraint. The procurement department exists to manage supplier risk, not to place orders. The compliance function exists to absorb regulatory complexity, not to fill in forms.
Choudary gives an example: the anesthesiologist. In the operating room, almost every gesture they supervise is performed by a machine. Their role remains, and stays well paid, because it carries a constraint that no machine absorbs: managing the vital risk while the patient is asleep. Automating the visible gestures of a role does not make the constraint that gave rise to it disappear.
These constraints are not equal in the face of AI. Three families stand out.
Lasting value lodges in the last two.
At the scale of an entire enterprise, these constraints form a map. Take a medical device manufacturer. Its structuring constraints are neither to design nor to assemble, but for example:
These four constraints hold the enterprise together. Assembling the devices is only a task, and AI will touch it well before the constraints. It is this map, not the list of positions, that the leader must keep in mind.
Klarna learned this the hard way. The Swedish fintech reduced its headcount from 5,500 to 3,400 between 2022 and 2024, mostly in customer service, relying on a chatbot meant to do the work of 700 people. In 2025, its CEO Sebastian Siemiatkowski admitted he had gone too far: degraded quality, eroded customer trust, rehires. The usual reading of the episode, AI does not replace human empathy, is flat.
The chatbot did work: it handled the majority of requests quickly and well, at lower cost. The error lies elsewhere. Klarna confused the function of customer service, measured in volumes handled and turnaround times, with its constraint: maintaining enough trust for a customer to agree to defer a payment of several hundred euros. That trust is decided in moments of friction, not in smooth interactions, and those moments are precisely the ones the chatbot handled badly. Klarna's commercial control point, its capacity to inspire trust, eroded silently before appearing in the indicators. Siemiatkowski was probably right to believe most of customer service was automatable. He was wrong about the nature of what remained once the essential was removed.
The Bpifrance Le Lab study of June 2025, conducted among 1,209 leaders of SMEs and mid-caps, is revealing. 58% consider AI important or very important for the survival of their company over 3 to 5 years. But 57% have not formalized an AI strategy, and among those that have adopted an AI, 94% do so to optimize the existing rather than to develop their activity. Lucidity about the danger coexists with action that has stayed operational.
This reflex reflects a problem of reading grid, not a lack of tools or skills. And it is a missed opportunity, because European SMEs and mid-caps hold assets the giants cannot replicate: long relationships, fine knowledge of their markets, accumulated trust capital. So many structuring constraints they already hold, without always naming them.
The work starts from the constraints, not from the positions. It is about identifying the few constraints that hold the enterprise together, like the map of the medical device manufacturer, then putting them through a three-step filter.
For each one, does AI dissolve it, displace it or harden it? The scarcity of a knowledge dissolves, coordination shifts toward the governance of agents, risk and trust harden.
For the constraints that shift or harden, how to reorganize the roles around their new form, and with which human-machine assemblies?
This work belongs neither to skills (human resources) nor to tools (information systems). It precedes them. Only the leader sees the enterprise as a system of constraints articulated with one another. It is their responsibility, and AI makes it explicit.
Do we prefer to take the enterprise back through its constraints, knowing that this work disrupts perimeters and takes time, or do we keep optimizing task by task, betting that overall coherence will restore itself?