Shadow AI is any AI tool your staff use for work without the company knowing about it or approving it. Usually it’s a personal ChatGPT account, a free meeting note-taker, or a browser extension that rewrites emails.
The people using these tools aren’t doing anything sinister. They found something that saves them an hour a day. The problem is that company and client data now lives in accounts you don’t control, under terms you never read.
Finding it doesn’t require a new product. Most of the evidence is already sitting in systems you pay for.
Finding shadow AI
Five places your existing systems already show where data goes
- 1Connected app consentsEntra ID or Google Workspace: AI apps granted access to mail and files
- 2DNS and firewall logsWhich AI services your network talks to, and how often
- 3Browser extensionsManaged browser reports list AI add-ons that read pages
- 4Meeting attendeesNote-taker bots joining Teams and Zoom calls
- 5Ask your staffAn anonymous survey finds what logs can't, including personal accounts
Hover a step to see what it reveals. Logs show tools and traffic, not what anyone typed.
Check which apps people have connected to their work accounts
This is the highest-value place to look, because these apps can read data directly.
When someone signs into an AI tool with “Sign in with Microsoft” or “Sign in with Google” and clicks Accept, that tool gets ongoing access to whatever they approved. That can include mail, calendar, and files.
- Microsoft 365: in the Microsoft Entra admin center, open Enterprise applications and review the apps users have consented to, along with the permissions each one holds.
- Google Workspace: in the Admin console, go to Security > Access and data control > API controls and review third-party app access.
Look for note-takers, writing assistants, and “AI inbox” tools with mail or file access. Then decide whether users should be able to grant that kind of access on their own. Both platforms let you require admin approval instead.
Read your DNS and firewall logs
Your firewall or DNS filter logs every domain your network talks to. Search those logs for the big AI services: chatgpt.com, claude.ai, gemini.google.com, perplexity.ai, copilot.microsoft.com, and the note-taker apps your team is likely to use.
This tells you how much traffic goes to each service and from which devices. It doesn’t tell you what anyone typed. It also misses anything done on a personal phone off the office network, which is a real gap for remote teams.
Look at browser extensions
AI extensions are easy to miss and often have broad permissions, including the ability to read every page the user opens. If your browsers are managed through Intune or Chrome Enterprise, you can pull a list of installed extensions across devices. If they’re not managed, that’s a finding in itself.
Check who’s showing up to your meetings
AI note-takers join calls as participants, often named something like “[Name]’s Notetaker.” Review recent Teams or Zoom meetings, especially client calls. A bot recording a client conversation through someone’s personal account is exactly the exposure you’re trying to find.
Ask your staff
Logs show you tools. They don’t show you why. An anonymous survey does. Ask what AI tools people use, what tasks they use them for, and what they wish they had.
Tell them up front that the point is to give them better tools, not to punish anyone. You’ll get honest answers, and you’ll learn which tasks are worth automating properly.
Decide what to keep, replace, or block
For each tool you find, pick one of three answers:
- Sanction it. Move people onto a business account with proper data terms and admin control.
- Replace it. Point people to an approved tool that does the same job.
- Block it. Reserve this for tools with no business version or unacceptable data terms.
Then write down the rules in a short AI acceptable-use policy: which tools are approved, and what kinds of data can never go into any AI tool.
Our take: blocking everything doesn’t stop shadow AI. It moves it to personal phones, where you can’t see it at all. The fix that works is giving people a sanctioned option that’s at least as good as what they found on their own. Locking down app consents and access is part of how we run cybersecurity for our clients, and we’re happy to walk through what we’d check first.