Shadow AI

Shadow AI is model, agent and AI-tool usage that exists inside an organisation without having been approved, reviewed or inventoried. A team wires a hosted model into a support workflow on a personal key; an engineer installs an MCP server from a public registry into an editor that holds repository credentials; a department uploads a customer spreadsheet into a chat product to summarise it.

It is the successor to shadow IT, and it spreads faster for a simple reason: adopting an AI tool needs no procurement, no infrastructure and often no payment. The barrier that made shadow IT visible — someone had to provision something — is gone.

The exposure is invisible by construction. Nothing appears in an asset register, no vulnerability scanner reports it, and the first accurate picture usually arrives during an incident or a customer security questionnaire.

Blocking is also the response that fails. Prohibition drives usage onto personal accounts and personal devices, where there is no logging at all — the organisation trades a visible risk for an invisible one. The workable answer is discovery plus a sanctioned path that is easier than the unsanctioned one.

And the questions are now contractual. Enterprise security reviews and the EU AI Act both ask which AI systems process which data. "We are not sure" is an answer with commercial consequences.