Here's an uncomfortable number. Among organizations that actually have an AI governance policy, only 34% regularly audit for unsanctioned AI, according to IBM's 2025 Cost of a Data Breach Report. That's the subset with a policy. Everyone else isn't looking at all.
So most shadow AI gets discovered one of three ways: by accident, by incident, or by an auditor. All three are worse than finding it yourself. This post is about the fourth way, going and looking. If you want the full risk picture first, the numbers live in our companion piece on the shadow AI compliance gap. If you already know what's out there and want to stop it recurring, that's the prevention playbook. This one falls between them.
Where shadow AI actually hides
The mental image is an employee with ChatGPT open in a browser tab. That's real, but it's only the most visible layer. Personal AI accounts on work machines are the obvious case. Less obvious: browser extensions with AI features that read page content, including whatever internal system is on screen. AI features quietly switched on inside SaaS tools you already approved for something else. A department running its own pilot with a tool nobody submitted for review. A developer calling a public model's API from a script that never went through procurement.
Each of these moves data, and none of them show up if your only detection method is asking IT what's installed.
Five signals that reveal it
- Network and DNS logs. Start here, because it's the cheapest look. Traffic to AI domains from corporate devices gives you a rough census of which tools are in use and how often. You're after patterns: which departments, what volume, what time of day.
- DLP and paste events. If you run data loss prevention tooling, look at large paste or upload events targeting AI domains. This is the difference between knowing people use a chatbot and knowing someone moved four thousand words of something into one.
- Expense reports and procurement records. AI subscriptions bought on personal or department cards are shadow AI with a receipt. Search expense lines for the obvious vendor names. It's tedious and it works.
- Browser extension inventories. Managed browsers can report installed extensions. Flag anything with AI summarization, writing, or page-reading capability, and check what permissions it holds.
- Just ask, with amnesty. The technical signals miss the human layer. KPMG's 2025 global study found 57% of employees hide their AI use. They hide it because they expect punishment. A short survey that promises no consequences for honest answers will surface tools no log file ever will, and it tells your people you're mapping demand rather than hunting offenders. That distinction decides whether they keep talking to you.
A first pass you can run in 30 days
You don't need a program to start. Pull 30 days of DNS or proxy logs and count AI domains by department. While that runs, have finance search a year of expense lines for AI vendors. Send the amnesty survey in the same window. Then put the three lists side by side. The overlap is your confirmed inventory. The gaps between them are just as useful: tools in the logs that nobody admitted to tell you where trust is thin, and tools in the survey that don't appear in logs tell you people are using personal devices for work data, which is its own finding.
Write down what you found, including the data classes likely involved. In a regulated environment that record matters. An organization that found shadow AI, documented it, and acted on it looks very different to an examiner than one that never looked.
What to do with what you find
Resist the urge to fire off a warning memo. The inventory you just built is a map of unmet demand: every unsanctioned tool is a task someone needed help with, sorted by department and frequency. That's exactly the information you need to stand up a sanctioned alternative people will actually use, which is the only fix that has evidence behind it. For regulated organizations, that usually means a private AI environment where the data never leaves your network and every query is logged, so the next audit finds a system you can answer for instead of a browser history you can't.
Run the pass again in a quarter. The tool landscape shifts fast, and the inventory is how you find out what your people are already telling you with their workarounds. Look first. Then build.
The full data picture is in Your People Are Already Using AI. Your Compliance Program Can't See It. For closing the gap after you've found it, see How to Prevent Shadow AI at Work and the AI governance tools buyer's guide.
Replace the shadow tools with one you can audit
Book a short demo and see a private AI assistant working on your own documents, with logging, citations, and nothing leaving your network.
Request a Demo