Many companies have deployed AI without seeing their margins move. The brake rarely lies in the tool, but in where it is plugged in: an organization designed before AI.
Published on 2026-05-29 by Nathalie Lamborghini Dumas.
Many companies have deployed AI across several departments without seeing their margins move. The brake rarely lies in the tool. It lies in where it is plugged in: into an organization designed before AI, whose structure absorbs the gains before they reach the margins.
The scenario recurs in many companies. One department adopts an assistant to draft its reports, another automates its data entry, a third tests a copilot on its code. Each team saves time, measures it, presents it. And at the level of the income statement, nothing moves. Margins stay where they were.
This gap between local progress and an unchanged income statement is not a measurement problem. It separates two gestures often confused: equipping one's organization with AI, and designing one's organization around AI.
A company does not only produce. It coordinates continuously. A good part of its teams' time goes into circulating information, waiting for a validation, rephrasing a request for the neighboring department, arbitrating between two priorities, checking that what the upstream delivered is usable downstream. This hidden cost, let us call it the coordination tax: everything an organization spends to synchronize with itself, and which appears on no budget line.
This tax is set by the structure: the carving into departments, the decision circuits, the validation levels, the software that does not talk to each other. It was calibrated for a world where each task cost human time. Accelerating tasks with AI without touching this structure leaves the tax intact. You produce more quickly documents that still wait for the same validation, transit through the same circuits, stumble on the same arbitrations. The tool's gain exists, but it dilutes in an organization that keeps paying its old synchronization cost.
To spot where AI really acts, a grid helps. Analytic-predictive AI reads: it classifies, detects, anticipates. Generative AI responds: it drafts, summarizes, translates. Agentic AI acts: it chains steps and carries out a task end to end. The current wave of rollouts is mostly generative, so it accelerates individual gestures. A faster gesture in an unchanged circuit does not move the result. What touches the coordination tax belongs to agentic AI, when an entire chain is entrusted to it. But this still requires leaving it room, which means reviewing who validates what, and why.
The field confirms this gap. A study by MIT (NANDA initiative, 2025), built on 150 executive interviews and the analysis of 300 deployments, quantifies the gap: despite 30 to 40 billion dollars invested, nearly 95% of generative AI projects in companies produce no measurable effect on the income statement. The authors rule out the technological explanation: at fault is the grafting of models onto processes and an organization that do not adapt to them, more than their intrinsic quality. Generic tools accelerate everyone's work without entering the company's circuits. An industrial leader interviewed sums up the paradox: much agitation in the discourse, nothing moved in operations.
This observation extends AI in the enterprise: the architectural work that falls to the leader, where a role exists to resolve a problem of scarcity, of risk or of coordination, not to execute gestures. As long as AI lands on the gestures without touching the constraint that structures the organization, the cost shifts instead of disappearing.
At the other end, some companies place AI at the center of their organization from the start. These companies "born with AI" design their roles, circuits and validations starting from what AI can do. They therefore never carry the old coordination tax.
Lovable, a Swedish company launched in late 2024, illustrates this point. Its platform allows applications to be created in natural language. Its team remained small for a long time, a few dozen people. Yet its revenue per employee far exceeds the software norm: around two million dollars of annual recurring revenue per employee, against a few hundred thousand elsewhere. What explains this ratio is not so much the quality of its AI as its organization. Synchronization between functions never needed to be heavy there, because it was conceived from the outset around automation.
An SME or a mid-cap will not be reborn. It has a history, customers, regulatory obligations, assets the pure players do not have. The lesson is therefore not to copy Lovable. It is to look its own coordination tax in the face, then to choose one process at a time, the one where it weighs the most. Where information waits, where validations pile up, where what already exists elsewhere is re-entered. Reorganizing this process around AI, rather than laying a tool on top, is what inscribes the effect over time.
It is at the moment of reorganizing a process that the strengths of an established company count. Its knowledge of the trade tells it where synchronization really costs. Its accumulated data feeds the models. Its trust relationship gives it the right to redraw a circuit without breaking everything. A company born with AI starts from a blank sheet; an established company starts from a terrain that it alone knows.
Two uses of AI coexist in a company, and they do not produce the same effects. The first accelerates gestures in an unchanged organization and reads mostly in activity reports. The second reorganizes a circuit around AI and ends up reading in the margins. Both have their place, provided you know which one you are deploying, and where.
On which processes do we accept to equip the existing for quick but capped results, and on which are we ready to redraw the organization so that the effect lasts over time?