Shadow AI: how to find it and govern it without banning it
Some of your people probably use AI tools nobody approved. Here is a practical way to discover that use, set rules people can follow, and give them approved tools instead.
AI tools are easy to start using. Anyone with a browser can open a chatbot, paste in a document and get a summary in seconds. That convenience is exactly why “shadow AI” appears inside organizations: AI tools used for work without the organization’s knowledge, approval or controls.
This is rarely malicious. People use these tools because they help. The goal is not to stop that. It is to make sure it happens with tools and rules the business has chosen.
What shadow AI looks like
- Personal accounts on public AI chat services, used for work documents.
- Browser extensions that read page content or email.
- AI features switched on inside software you already use.
- Meeting assistants that join calls to record and transcribe.
- Third-party apps that people have granted access to their mailbox, calendar or files.
Why it matters
- Data handling. Whatever is pasted into a tool leaves your control. Depending on the service and the account type, it may be stored, reviewed or used to improve the service.
- Access. An app granted access to a mailbox or file share can read far more than the one document someone meant to share.
- Confidentiality and compliance. Client contracts, privacy obligations and industry regulations often restrict where certain data can go.
- Accuracy. AI output that goes to a client without review can carry mistakes under your name.
Step 1: Find out what is in use
Use several sources. Each one shows part of the picture.
Ask people. A short, no-blame survey is often the fastest source. Ask which tools they use, for what tasks, and what kind of information they put in. Make clear the goal is to provide better tools, not to punish anyone.
Review app consents and sign-ins. Your identity platform keeps a list of third-party applications users have connected to their work accounts, and a record of sign-ins. Review both for AI services and for apps with broad access to mail or files.
Check network and DNS logs. Firewall, web filter and DNS logs show which AI services devices on your network reach, including ones used with personal accounts.
Use cloud app discovery if you have it. Some security platforms identify AI apps automatically. Microsoft describes, for example, how Defender for Cloud Apps offers a discovery engine that identifies which generative AI apps are in use, with a catalog of risk scores for those apps. Teams can set policies that alert on new AI app use or mark specific apps as unsanctioned. That setup requires both Defender for Cloud Apps and Defender for Endpoint.
Inventory browser extensions and meeting bots on managed devices and in calendars.
Step 2: Decide what is approved
Sort the tools you found, and the ones people ask for, into three groups:
- Approved. Available to everyone, under company accounts.
- Allowed with limits. For example, no client or personal data.
- Not allowed. Blocked where possible, with an explanation.
To decide, look at each tool’s terms on data use and retention, whether it supports company sign-in and administrative controls, what logging it offers, and where data is stored.
Step 3: Write rules people can follow
Keep the acceptable-use policy short enough to read in five minutes. It should cover:
- Data that never goes into AI tools, such as client confidential information, personal data, credentials and regulated data.
- Review. Any AI-assisted work that leaves the organization is checked by a person.
- The approved list, and how to request a new tool.
- Who to ask when someone is unsure.
Step 4: Give people a better option
Policies fail when the approved option is worse than the unapproved one. If people need summarization, drafting or transcription, provide a sanctioned tool under company accounts, with enterprise controls switched on. Then the policy is pointing people toward something useful, not just away from something.
Step 5: Train, then review regularly
Short, practical training works best: what is allowed, with examples from your own work. Then repeat discovery on a schedule. Quarterly is a reasonable starting point for most small and midsize organizations. Look for new tools, review requests, and update the approved list.
Organizations that want a fuller structure can look at NIST’s AI Risk Management Framework. It is voluntary, released in January 2023, and NIST has since published a Generative AI Profile to help organizations identify the specific risks of generative AI.
Why this conversation should happen now
On CRN’s The Path to Growth, Paul Redding of NinjaOne made a point that applies to anyone responsible for an organization’s technology: if your technology partner does not raise AI use with you, someone else eventually will. The business owner will then ask why it was not raised sooner. Better to find out on your own terms.
A starting checklist
- Run a short, no-blame survey of AI tool use
- Review third-party app consents and sign-ins
- Check network and DNS logs for AI services
- Turn on cloud app discovery if your platform supports it
- Sort tools into approved, limited and not allowed
- Publish a one-page acceptable-use policy
- Provide at least one approved tool for common tasks
- Schedule the next review
Want help with the first review?
We help organizations run discovery, set practical policies and roll out approved tools as part of our security and continuity work. If you would like a first review, contact us.