What is shadow AI and how do we stop it?

By LAN Service Group · Reviewed October 4, 2026

Why shadow AI is a data problem

Most shadow AI is well-intentioned. Someone wants a faster summary, cleaner code or a better first draft, and the free tool is one browser tab away. The problem is where the data goes and under whose terms.

OpenAI's help center, for example, says it may use content from its services for individuals to train its models unless the user opts out, while by default it does not train on ChatGPT Business, Enterprise, Edu or API data. A personal account also means no admin console, no audit trail and no way to recover or remove company data when the employee leaves.

Find it before you try to stop it

You cannot govern what you cannot see. Microsoft Purview groups generative AI apps into three categories, including "other AI apps" detected through browser activity and categorized as generative AI in the Microsoft Defender for Cloud Apps catalog, such as ChatGPT, Google Gemini, consumer Microsoft Copilot and DeepSeek.

Outside Microsoft tooling, look at expense reports and card statements for AI subscriptions, single sign-on and OAuth consent logs for AI apps connected to company accounts, browser extensions on managed devices, and meeting note-taker bots joining calls. Then simply ask teams what they use; an amnesty usually surfaces more than any scan.

Control it without a blanket ban

• Offer a sanctioned tool first: a company-licensed AI service with single sign-on, so the easy path is also the governed one.
• Use data loss prevention at the endpoint: Microsoft says onboarded Windows devices can warn or block users sharing sensitive information with third-party generative AI sites in a browser, such as pasting credit card numbers into ChatGPT.
• Watch for risky behavior: Microsoft Purview Insider Risk Management includes a risky AI usage policy template.
• Block clearly high-risk AI sites and review which AI apps have been granted access to company mail, files and calendars.
• Give people a fast way to request a new tool, so asking is easier than working around you.

Policy and people

Technical controls only cover managed devices and accounts. Personal phones and home laptops are outside that view, which is why a total ban tends to push AI use somewhere you see even less. A one-page acceptable-use policy that names the approved tools, the data that never goes into any AI tool and the request process does more than a blocklist.

Assign one accountable owner and review the approved-tool list each quarter. NIST's AI Risk Management Framework, which is voluntary, puts this under its Govern function and gives you a recognized structure to report against.

Common follow-up questions

Sources

• Microsoft Learn: Microsoft Purview data security and compliance protections for generative AI apps
• OpenAI Help Center: How your data is used to improve model performance
• NIST AI Risk Management Framework