An employee gets stuck on a text, a summary, a piece of code or an email that just isn't working out. There is a chat window, free, immediately available, without waiting time and without a form. The text goes in, including the customer's name, the contract amount, the personnel file passage or the source code that belongs with it. The answer comes back, the work is done, the window is closed. No one has asked permission for this, and no one has refused, because no one knew.
This is not an incident. It is a predictable pattern that arises at every point where workload is greater than the effort of asking. A deadline, an unanswered question that the company neither explicitly allows nor forbids, and a free tool that is faster than the internal process: that combination leads to this behavior on its own, regardless of the sector or the size of the organization.
A ban in a policy document changes nothing about the moment when the deadline pinches. The chat window remains accessible, the employee remains under pressure, and the trade-off between "done quickly" and "by the book" more often falls in favor of done quickly. That is not unwillingness. It is an organization that has made it easier to use the window than to not start using it.
The pattern strongly resembles what happens with a department that has taken out its own subscription: someone sees a problem, finds a solution that is available within budget and without an approval process, and acts. The difference is that a chat window does not need a purchase order. There is no invoice that stands out, no contract that has to go through legal affairs, no subscription that appears in an overview. The use leaves no trace except in the browser history of one employee, and sometimes not even that.
As a result, it continues to exist alongside every official process that the organization has set up. Whoever builds an approved alternative but does not block the free window or make it unattractive sees exactly the effect that also occurs when a second process alongside the existing one is ignored: the new system is not used, because the old one feels faster or is simply the habit.
An inventory of approved software tells you what has been purchased. It tells you nothing about what is used. A chat window in a browser tab, a browser extension with access to mail, an AI function that has been added somewhere deep in an existing subscription: this kind of use arises outside every purchasing process and therefore does not appear on the list that IT manages. The list is not wrong, it is simply not complete, and the difference between those two is exactly where shadow AI lives.
The only way to know what is actually happening is to ask. Not as a check, not as a harbinger of a sanction, but as a question to which an honest answer is possible. Whoever feels that an answer to "which tools do you use besides the approved systems" leads to a conversation with a manager will not give that answer. Then the organization remains dependent on what happens to stand out: a bill that is just slightly off, a colleague who accidentally mentions it, or a complaint from a customer whose data has turned up somewhere it should not have ended up.
An inventory that starts with accusations does not produce an inventory, but silence. Employees who know that an honest answer will create a problem will simply start opening the window in incognito mode. The goal is not to punish yesterday's behavior; the goal is to get a current picture of what is running now, who uses it and for what, so that this use can be given a place in the existing risk framework instead of continuing to exist alongside it.
This also applies to AI that does not arrive via a chat window but via a supplier. When a supplier has built AI into a product that was already in use, the nature of the data processing changes without a new contract having been signed. And a pilot setup that worked well is rarely formally concluded: it keeps running, even after a pilot setup is never switched off while no one knows anymore who is responsible for it. In all three cases, it is not a matter of ill intent, but of a gap between what has been approved and what works. Where oversight takes shape in practice in such places is addressed on the page about what human oversight means in practice; the content of the regulation itself lies elsewhere and that does not change.
Once it is clear which AI is actually being used, a different kind of question arises: which part of the work that now runs through a free chat window actually belongs to a task that AI can take over in a structured way, and which part does not. The workscan from FTE TO AI makes that distinction visible: it calculates per task which part of the work can be taken over by AI, so that the use that is now hidden in a browser window can be given a place that fits what the work actually requires.
Vraag maar. Governance begint bij weten wat er draait — ook wat niemand heeft goedgekeurd.
Answers come from this site’s knowledge base. Not tailored advice, and not a scan of your company.