A team wants to work faster and switches on an AI tool, usually to test something concrete: a summary, a first draft, a classification. There is no approval process because it's "just a trial." The trial works, or works well enough, and disappears into daily process without anyone ever making a decision to keep it. There was no moment where someone deliberately chose: we're including this in our AI inventory, this has an owner, this falls under that risk level. There was only a moment of starting, and that moment becomes fixed into a way of working that nobody has touched since.
Stopping a pilot setup requires a decision. Continuing requires nothing. As long as nobody reports an incident or an external party asks for proof of control, there is no reason to revisit something that has long since become ordinary work. The people using the tool have no interest in bringing it to attention: it works, it saves time, and reporting it feels like risking that it gets taken away. So a temporary solution remains structural, not because nobody knows about it, but because nobody asks the question that would lead to shutting it down.
It seems logical to think that an overview of approved software makes this visible. That overview shows what was purchased, not what is used. A pilot setup often runs outside that channel: a free account, a browser plugin, a separate subscription on a team card. The same pattern occurs with a browser extension with access to your email, with a department with its own subscription outside central procurement, and with company data pasted into a free chat window without a contract with data processing terms behind it. None of these situations appear on an IT list, because none of them was ever registered as a project.
The only way to know what is really running is to ask. Not as retrospective control, but as an opening: what do you use, for what, and since when. Whoever asks that question with the threat of sanctions in the background will get no answer, or an incomplete one. Employees using a tool that nobody has approved have something to lose from honesty and nothing to gain. That is precisely the tension addressed in employees using a tool that nobody has approved: without an invitation that is safe to answer, the pilot setup stays underground, and it doesn't disappear, it just becomes quieter.
An inventory without consequences for the reporter is the first step, not the conclusion. After making things visible comes classification: what is this, what role does it play in a process, and what risk level fits it. A summarization tool for internal notes requires a different regime than a tool that helps decide on customer acceptance or personnel evaluation. That classification should align with the risk structure that already exists, not with a new framework alongside it. A second, informal process that gets ignored alongside the existing process is precisely why a second process alongside the existing one gets ignored is a recognizable pattern: people follow the path of least resistance, and that path only becomes visible once someone asks about it.
Then comes the question of who still looks at it. A pilot setup that has been let loose in daily work usually no longer has an owner who assesses outcomes before they are used. That is where what human oversight means in practice becomes relevant: not as an abstract principle, but as a concrete question of who reads the output before it touches a customer, a file, or a decision.
An executive who is questioned about this — by a supervisory authority, an auditor, a supervisory board — has two options: say that the overview is complete, or say that the process exists to make it complete and keep it that way. The second option is more honest and, in the longer term, stronger, because an organization that actively takes inventory has something demonstrable, even when the inventory is not yet finished. The point is not that an unauthorized pilot setup will never arise again. The point is that there is a mechanism that finds it before an incident does that for you.
Once it becomes visible which AI has arisen unplanned, another question often follows: if this team was already doing this, without anyone organizing it, what would happen if we did organize it? That is a different question from governance, and it belongs with the work scan from FTE TO AI, which calculates per task which part of the work can reasonably be taken over by AI. Where this page describes how you gain control over what is already running, that scan describes what, with the same sober eye toward tasks and risks, can still be built.
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.