How do you find shadow AI?

Three passes, in order of how much they find per hour spent. Look at what the tools you already pay for have switched on, since that is the largest category and the easiest to check. Look at expense claims and card statements for individual subscriptions. Then ask people directly, having made clear that the answer is not a disciplinary matter, because the personal-account cases leave no trace anywhere else.

Is shadow AI a security problem or a governance problem?

Both, and they fail in different ways. The security failure is data reaching a place with no agreement covering it. The governance failure is quieter: an inventory that is confidently wrong, so every downstream check runs against a set that is not the real one and reports a clean result.

Should we just block AI tools?

Blocking moves the same use onto personal devices and personal accounts, which is the version you can see least and control least. It also tends to fail on the embedded category, since the tool being blocked is one you approved for another purpose. The route that works is making the sanctioned option fast and the boundary specific about data instead of about tools.

What is shadow AI?

Shadow AI is AI being used in a company that nobody approved, recorded or reviewed. It arrives in three ways: a person signs up for a tool because it makes their job faster, a vendor switches on an AI feature inside software you already pay for, and somebody uses a personal account for work because the company one does not exist. Only the first looks like a decision, and it is the smallest of the three.

The category nobody counts

Ask a company what AI it uses and you usually get a list of AI products it bought. That list is real and it is the visible part. The larger part is the assistant that appeared in a note-taking app, the summarizer added to a help desk, the writing suggestions in a CRM. Nobody adopted those. They arrived in an update.

This is what makes an inventory wrong rather than merely incomplete. An incomplete list is missing entries you would recognize. A list that omits embedded features names tools you already approved for something else, so every check you run against it passes while describing something other than what is happening.

Shadow AI is almost never defiance

The instinct is to treat unapproved adoption as a discipline problem, and that reading survives about five minutes of asking people why they did it. Almost always the answer is that the sanctioned route was slower than the work, or that no route existed and asking felt like an imposition.

That reframes the fix. A stricter prohibition produces the same adoption with less visibility, because the behavior was never driven by permission in the first place. A faster approval route produces less shadow AI, because using it becomes easier than not. The measure of an intake process is not how thorough it is. It is whether people would rather use it than avoid it.

Why the personal-account case is the worst of the three

A tool on a personal account sits outside everything you can do. You cannot revoke the access, export what went into it, see what it retained, or apply a data agreement to it. When the person leaves, whatever they put in stays where it is under an account you have no relationship with.

It is also the hardest to find, because it leaves no trace in billing, in single sign-on, or in any system you administer. Finding it means asking, and asking works only if the honest answer is not going to be punished. A company that treats disclosure as an admission gets an inventory of what people were willing to admit.

Common questions

How do you find shadow AI?: Three passes, in order of how much they find per hour spent. Look at what the tools you already pay for have switched on, since that is the largest category and the easiest to check. Look at expense claims and card statements for individual subscriptions. Then ask people directly, having made clear that the answer is not a disciplinary matter, because the personal-account cases leave no trace anywhere else.
Is shadow AI a security problem or a governance problem?: Both, and they fail in different ways. The security failure is data reaching a place with no agreement covering it. The governance failure is quieter: an inventory that is confidently wrong, so every downstream check runs against a set that is not the real one and reports a clean result.
Should we just block AI tools?: Blocking moves the same use onto personal devices and personal accounts, which is the version you can see least and control least. It also tends to fail on the embedded category, since the tool being blocked is one you approved for another purpose. The route that works is making the sanctioned option fast and the boundary specific about data instead of about tools.

Find out where you stand

The assessment on this site asks 22 questions about what AI is in use, what rule governs it and who is accountable, and returns a score, a verdict and, where a control is short, the one thing to fix first. It names no framework, and it takes about ten minutes.

It is a self-assessment of what you could show today, not an audit, not a certification, not a conformity assessment and not legal advice.

Run the assessment