GenAI produces gains. It establishes two invisible debts that no one measures: a phantom quality and a cognitive surrender. Untreated, they pay for themselves later, at the worst time.
They are right, and this is precisely where we must look
Bpifrance Conseil and Siparex have just published the most comprehensive overview of the year on AI in SMEs: twenty use cases, twenty documented successes, all functions combined. On the winnings, nothing to complain about. The demonstration is solid.
The most interesting, however, is in the first pages. The book cites a McKinsey study from 2025: more than 80% of organizations that have invested in generative AI see no tangible financial impact. MIT figures the same gap at 95% in The GenAI Divide. A book that celebrates twenty successes admits in the preamble that the rule is failure.
So the right question is not whether generative AI works. It works. The question is what it installs while it’s running.
The paradox: what is counted, and what accumulates
What is measured is controlled. What cannot be measured accumulates. The documented use cases measure return on investment project by project. None measure what each use deposits, silently, in the way teams work.
This deposit has a name: practice debt. It takes two forms, invisible because they do not appear in any dashboard. And that’s precisely why they run.
First debt: ghost quality
We trust an AI output because its form is convincing, without verifying its substance. The sentence is fluid, the structure is clean, the tone is assured. We validate.
The problem is not the one-time error. This is what I call the amplifying loop. Plausible but false output is not corrected: it is reused. It feeds the following production, which takes it as an acquired knowledge. Error does not dissolve, it spreads. And as it circulates, it gains authority. What was generated once becomes, three documents later, an internal reference that no one thinks of questioning anymore.
This is not a risk of hallucination, which the white paper correctly addresses. It is a regime of production where verification recedes at the exact pace that fluidity advances.
Second debt: cognitive surrender
By delegating reasoning, judgment atrophies. The team moves from producer to receiver. She rereads, she validates, she adjusts at the margin. Little by little, she herself ceases to produce the reasoning that she is now content to validate.
It doesn’t show until everything is going well. This can be seen the day you have to arbitrate a borderline case, resolve a situation that the AI has not seen, or detect that a convincing answer is actually off topic. That day, the skill is no longer there. No one has decided to part with it. She faded away while we were delegating.
It is the more serious of the two debts. The ghost quality can be corrected with method. Cognitive surrender erodes the very ability to correct.
These are not tool issues
Neither of these two debts can be resolved by changing the model. A more powerful model produces even more convincing, therefore even less verified, outputs. It makes the problem worse instead of solving it.
These debts live in actions, at three levels: the individual, the team, the organization. Training individuals improves individual outputs, not what the team produces collectively. Acting at one level is not enough.
We must make a distinction here that the white paper does not have to make, but that a manager has an interest in seeing. The work grades the ambition of the projects, from individual use to structuring projects. This is a project scale axis. The two debts lie on a completely different axis: the dissemination of a practice within the collective. A project can show an excellent return on investment and install both debts at the same time. The project measures what it produces. Practice decides what remains.
Interrupt loops
Dealing with these debts is ongoing work, not a deliverable. Structure your thoughts before acting. Make each statement verifiable. Spread the right actions rather than concentrating them on a few experts. Evolve them as tools change.
This job needs a manager. Not just another AI project pilot, but a role that takes on the practice itself: its quality, its rules of use, its consistency, its dissemination. He is to generative AI what the Data Steward is to data. Its lever is not technical.
And unlike a permanent AI manager, this role is intended to fade into the background. He sets up the practice, trains internally, then retires. What then remains is not another dependency. It’s an ability.
These debts are not intuition. They are measured, as is the gap between what management aims for in three years and what the teams actually do every day.
Generative AI does not replace judgment. She puts him to the test. Humans must remain the referee, not the spectator. The determinant has never been the tool. It is in practice that we install around.
