Shadow IT is an old problem: employees use tools IT does not know about. Shadow AI is the same phenomenon with higher stakes, because AI tools are typically fed exactly the content that is confidential: contracts, customer data, source code, internal strategy.
How shadow AI arises
Almost never out of malice. Employees want to work faster, the company offers no approved tool, and the free browser offering is one click away. Add AI features built into existing software, often via updates and without anyone reviewing a new processing activity, and meeting assistants that join calls through calendar permissions.
The risks
Confidential data leaves the company without a processing agreement, possibly being used to train models. Professional and trade secrets may be breached. AI uses that fall under the AI Act appear in no AI inventory. And outputs of AI tools flow unchecked into decisions and documents.
How to find it
Technically, indications come from several sources: OAuth grants through which employees give third-party apps access to email, calendars or files, proxy or DNS logs, installed browser extensions and card statements for subscriptions. Organisationally, a survey without threat of sanctions yields surprisingly honest answers.
How to govern it
Bans alone move usage to private devices. More effective is a short AI usage policy stating which tools may be used with which data, combined with an approved alternative good enough that nobody needs to work around it. New tools go through a simple approval process as part of AI governance. An annual comparison of approvals with technical findings shows whether the rules work.
Further reading
Technically, data loss prevention and an approved tool such as Microsoft 365 Copilot help. Where agents use systems, the rules for the Model Context Protocol apply.