Shadow AI: How to Find Which AI Tools Your Staff Are Using

What counts as shadow AI, and why it matters

Most shadow AI isn't malicious. Staff use AI because it saves time. The risk is what goes in and what the tool can reach: client data pasted into a consumer chatbot, a note-taker recording meetings without consent, or an app granted read access to every email and file.

The PDPA still applies whichever tool staff use. Our ChatGPT at work policy guide covers what data should never go into AI tools.

Microsoft 365: check Enterprise applications and user consent

Any app a user signs into with their Microsoft work account, and allows to read mail, files or calendars, appears in Microsoft Entra. This is the quickest free check.

1. 1

   Open Enterprise applications

   In the Microsoft Entra admin centre, go to Enterprise apps, All applications. Filter by application type and sort by creation date to see recent additions.

2. 2

   Review permissions on anything AI-related

   Open each unfamiliar app and check Permissions. Mail.Read, Files.Read.All or Calendars.ReadWrite on an AI note-taker or assistant means it can read that data for the user who consented.

3. 3

   Check sign-in logs

   Under Monitoring, Sign-in logs, filter by application to see who uses each app and how often.

4. 4

   Restrict user consent

   Under Enterprise apps, Consent and permissions, limit user consent to verified publishers and low-risk permissions, or turn it off. Microsoft's user consent guide explains the options.

5. 5

   Turn on the admin consent workflow

   With the admin consent workflow, staff who hit a blocked app can request approval instead of giving up or using a personal account.

Removing an app's access (delete the enterprise app or revoke its permissions) stops future data access. It doesn't delete data the vendor already copied, so ask the vendor to delete it if needed.

Google Workspace: check Accessed apps and API controls

Google's equivalent sits in the Admin console under Security, Access and data control, API controls. Google's app access help page explains how to view accessed apps and set each one to trusted, limited or blocked.

• Open Manage third-party app access and view Accessed apps: every app users signed into with Google and the number of users.
• Look for AI assistants, note-takers and "chat with your Drive" tools with access to Gmail, Drive or Calendar.
• Set approved apps to Trusted, and block apps you don't want accessing Workspace data.
• Consider restricting access to unconfigured third-party apps so new apps need admin approval.
• Check the OAuth log events in the audit and investigation tool to see when an app was first authorised and by whom.

If staff connect ChatGPT, Claude or Copilot to Drive or SharePoint, read our AI connectors security guide before approving them.

Browser extensions: the blind spot

AI browser extensions can read the pages staff view, including web email, CRM records and banking portals. They don't show up in OAuth lists. Google says Chrome Enterprise Core provides browser management and reporting at no additional cost, including installed extensions, once browsers are enrolled. Microsoft Edge can be managed similarly through Intune.

Paid discovery tools and what licence they need

Microsoft's guide to managing generative AI apps describes using Defender for Cloud Apps to discover, monitor and block AI apps across devices, and Purview DSPM for AI to see what data goes into them. These find website-based AI use that OAuth lists miss.

Other options include DNS filtering services that report AI domains visited from the office network, and secure web gateways. Whatever you use, tell staff what is monitored; covert monitoring damages trust and may raise PDPA notification questions.

Turn findings into an approved AI tools list

Discovery only helps if the result is a clear list staff can follow. Banning everything pushes AI use onto personal phones, where you have no visibility at all.

1. 1

   Ask staff what they use and why

   A five-question anonymous survey finds the free accounts no console shows. Ask which tools, for which tasks, and what data goes in.

2. 2

   Pick approved tools for the common tasks

   Usually one business chatbot with training off (such as Copilot Chat with your work account, ChatGPT Business or Gemini in Workspace), one note-taker and one writing aid.

3. 3

   Check training and retention settings

   Confirm the approved tools don't train on your data. Our AI training opt-out guide lists the settings per tool.

4. 4

   Publish the list with simple rules

   Approved tools, what data may go in, how to request a new tool, and who to ask. One page is enough.

5. 5

   Then restrict the rest

   Require admin consent for new OAuth apps, block high-risk extensions, and revoke tools you've replaced.

6. 6

   Review quarterly

   Re-run the OAuth and extension reports every quarter and update the list.

Do it yourself or get help?

An admin can run the OAuth and extension checks in an afternoon. Help is useful when the lists are long, when you need consent settings and extension policies changed without breaking tools people rely on, or when staff need training on what not to paste into AI.

Our Microsoft 365 / Google Security Lockdown (S$590) reviews connected apps, sets admin consent and extension policies, and documents the result. Security Awareness Training (S$390) covers safe AI use alongside phishing.