The AI you didn't deploy is still yours#
Ask a compliance lead where AI touches their business. You will usually get a short list: the chatbot, maybe a copilot in the productivity suite. Ask again a week later, after they have looked properly, and the list is much longer. A model summarising contracts in one team. An enrichment step calling an API in another. A spreadsheet macro routing customer data through a service nobody reviewed.
None of that was a decision. It accumulated. That is why it is hard to govern: you cannot put rules around what you cannot see.
Find it first, then decide#
The instinct is to reach for a control. But a control you place before you understand the workflow blocks the wrong thing, or nothing at all. The better first move is quieter. Watch. See where AI actually sits in the work, what data it touches, and what would happen if it were wrong — before you change anything.
That is the idea behind starting an Inspector in Watch mode. It observes a real workflow at zero risk and shows you what it would have caught. No autonomy handed over. Nothing rewired. Just a clear picture of where you stand.
Three places to look first#
Most of what people find sits in one of three places, and none of them is where the org chart says.
- Inside a tool you already pay for. A copilot in the productivity suite reads whatever the user can read. That is a permissions question, not a licensing one.
- In an integration nobody calls AI. An enrichment step, a classifier, a routing rule. It was bought as a feature, so it never went through an AI review.
- In a workaround someone built to save time. A script, a macro, a browser extension. It works, which is exactly why nobody mentions it.
Write down what each one reads, what it writes, and who would notice if it were wrong. That list is the start of a governance plan. Until it exists, everything else is guessing.
Run it as an interview, not a survey#
A survey asks "where do we use AI?" and gets the answer people can recall. An interview asks what a team actually did last Tuesday, and finds the rest.
Sit with one team for an hour. Walk a single piece of work end to end — a ticket, an invoice, a report — and ask at each step what produced this, and where did it come from. You are not looking for AI. You are looking for steps where the output arrived faster than a person could have produced it, or where nobody can quite say who wrote something.
That framing matters, because people do not think of their tools as AI. They think of them as the button that summarises the thread. Ask about the button and you get an answer. Ask about AI and you get a shrug.
What you will find, roughly in this order#
Things nobody hid. Someone will mention a tool in the first ten minutes that was never on any list. It was not concealed — it just never came up, because it works.
Things that came with a renewal. A feature that appeared in a release note. Nobody chose it, nobody reviewed it, and it now reads more of your data than the pilot everyone argued about.
Things built to route around a delay. A script or a personal workflow that exists because the official path took three weeks. These are the most sensitive to handle, because the person who built one is solving a real problem and will reasonably resist having it taken away.
Do not lead with a ban#
The instinct after a first pass is to shut things down. It is the fastest way to make the second pass impossible, because the next team will remember what happened to the last one.
The more useful response is boring: write it down, ask what would happen if it were wrong, and only then decide. Most of what you find will turn out to be fine. A small number will be genuinely important, and you will only hear about those if people believe an honest answer is safe.
Turning the list into a decision#
When the list exists, sort it once — by what a wrong answer costs, not by how much AI is involved. The two rankings look nothing alike, and the second one is the reason so many programmes spend their first year governing a chatbot while an enrichment step quietly moves customer data.
Then take the top one and measure it. Six weeks, read-only, on your own systems. You are not trying to prove the tool works. You are trying to find out what already happens, in enough detail that the next decision is not a guess.
Related