There's a scene, early in Analogique, the novel I wrote. André Vidal runs a forty-two-person consulting firm. One Monday morning, he discovers that his last six proposals are indistinguishable. Same tone, same turns of phrase, same absence of signature. He doesn't understand it. His teams have never produced so much, never delivered so fast.
What André can't yet see is that each of them, on their own, found the same assistant, and asks it roughly the same questions. No one told him. It appears in no official tool, no license, no report. It's Shadow AI. And the day he notices, it isn't through a security alert. It's because his proposals all look alike.
Shadow AI is all the AI use unfolding inside your organization outside any framework. In personal accounts, private tabs, subscriptions paid individually and expensed. You don't see it, so you don't steer it. And most of the time, you drastically underestimate its scale: it isn't a handful of the curious, it's already half your teams.
Faced with this, two reflexes, and both fail.
The first: ban it. Block access, publish a memo, restate the rule. The result, the use doesn't disappear, it goes deeper. People switch to their personal phone, copy-paste on the sly, and carry on exactly as before, without the slightest trace. You've turned a visible problem into an invisible one. You've made Shadow AI worse while thinking you removed it.
The second: allow everything. Let it run, in the name of productivity, because “you can't fight it anyway.” The result, sensitive data flowing into tools you didn't choose, deliverable quality that depends entirely on whoever holds the keyboard, and no memory of what worked. Speed, yes. Capital, never.
There's a third way, and it comes down to a simple shift: you don't govern the tool, you govern the practice.
Shadow AI isn't first and foremost a security problem to lock down. It's uncaptured value. When a salesperson finds, alone, the right answer to a hard objection, that isn't a leak: it's a practice that deserves to become the whole team's. When someone improves a framework, perfects a better follow-up message, that isn't a risk to extinguish, it's an asset to recover. The problem is never that they found it with AI. It's that it stays in the shadows, and tomorrow their colleague will walk the same path, from scratch.
Framing the practice means capturing what works, making it visible, having it validated by a referent, and putting it back in everyone's hands. The same use that, left alone, dilutes the firm's singularity becomes, once shared, what reinforces it. Individual use becomes a collective capability. The shadow becomes an asset.
Leaving that capital in the shadows has a cost, and it is now documented. Acemoglu, Kong and Ozdaglar (NBER, working paper 34910, February 2026) formalize it: when everyone delegates to AI without anything aggregating, an organization can execute faster while producing less shared knowledge. Beyond a threshold, the stock of common knowledge stops renewing itself. The researchers call that tipping point knowledge collapseKnowledge collapseThe erosion of institutional memory when AI accelerates execution but nothing gets capitalized.. And their conclusion is unambiguous: what protects an organization is its capacity to aggregate what its members learn.
That's exactly what André is missing. Not one more tool to block. A place where what each person discovers, alone, stops being a secret and becomes a strength.
So the real question, for a leader, isn't: how do I stop my teams from using AI on the quiet.
It's: why do their best discoveries stay in the shadows, instead of becoming ours?