A consultant hid his best work
That morning, he sends me his deliverable at 7:30. 50 slides, clean, well argued. I ask him how long it took.
Silence. Then: "I had help."
Not fraud. Not a mistake. His best work, concealed by the person who produced it.
That is credibility debt. It is not contracted when someone cheats, but when someone succeeds and would rather not say so.
Massive usage, zero visibility
The phenomenon is documented: massive and growing use of generative AI inside organizations, and near-zero visibility over that use.
The consequences are very concrete. Good practices die inside chat histories. The prompts that change the game stay on the machines of those who built them. Every week, teams reinvent the wheel because nobody capitalizes on what already works three desks away.
The mechanism is simple: making your AI use visible remains a social risk. You expose yourself to having your work reread differently, to having your added value questioned, to a manager wondering aloud why the role still exists. So nobody does it in the open. So the organization learns nothing.
This is not a tool problem and it is not a generational problem. It is a safety problem: as long as showing costs more than hiding, usage will stay clandestine, whatever the quality of the platform deployed.
What it actually costs
An organization in this state pays three times.
It pays for the licences of a tool whose real usage and produced value it cannot measure, since the usage hides. It pays for repetition: the same problem solved twenty times in parallel, without any of those solutions becoming an asset. And it pays for the risk, because invisible usage is also unframed usage — the very definition of Shadow AI.
The third cost is the one that reaches the executive committee, because it touches compliance. The first two are heavier and appear nowhere.
The antidote is not complicated
It requires no budget, no platform, no additional committee.
Prompts shared in team spaces, on the same footing as a document template. Usage feedback documented in retrospectives, with what worked and what produced noise. And above all usage named and valued by managers, out loud, in front of others.
That last point does all the work. As long as the first person to show their usage takes the risk alone, nobody moves. The moment a manager names someone in a meeting for the way they built their assistant, the social cost inverts.
Not much, on paper. Except that it turns AI from an individual secret into a collective advantage.
The chaos that comes next
Once usage is out of clandestinity, a second problem appears — and it is better anticipated than discovered.
We all saw documents everywhere when collaborative platforms arrived. Organized panic: everyone with their own folder tree, their own conventions, their own copy of the copy. With AI, the risk is of the same nature, only faster. Everyone creates their agent, their task, their prompt, and the organization ends up with inconsistencies that slow everything down.
The reflex answer is to put an AI on top to control it all. It is the wrong one. What is needed is a culture of sharing and a common map: who built what, for which use, and where to find it.
Before building the future, you have to make sure everyone knows where the plans are. In an organization deploying AI at scale, the deciding question is not "which model do we choose" but "who takes care of the digital living-together".
Two indicators that tell the truth
If you want to know where you stand without launching an audit, two signals are enough.
The first: in a team meeting, ask who used AI on the deliverable in progress. Count the seconds of silence before the first hand. That delay measures your credibility debt better than any anonymous survey.
The second: ask to see the last useful prompt written in the team. If it exists somewhere other than in a private conversation, you have started to capitalize. If not, you are funding, every month, a learning process that evaporates.
AI governance does not start with a 200-page charter. It starts by making usage speakable — the condition for everything else, and I detail the minimum setup in the missing link in AI governance.