How do you build a shadow AI discovery program from scratch?

Build it in four steps: define what counts as an AI security tool, pull signals from five sources rather than surveying employees, classify findings by the sensitivity of data flowing in, then sanction or block with a real replacement path. The step that determines whether the program works is the second one, because self-reported inventories consistently miss most usage.

Definition first, because scope creep kills the effort. Decide whether AI features embedded in already-approved software count, whether coding assistants are in scope, and whether browser extensions count. Most programs should answer yes to all three, but the answer needs to be written down.

The five signal sources, in order of yield: OAuth and third-party app grants in Google Workspace or Microsoft Entra, which reveal tools users authorized against corporate identity; browser extension inventory through enterprise browser management; egress and secure web gateway logs filtered for model provider endpoints; expense and card data for personal subscriptions being reimbursed; and code repositories for API keys against model providers.

Classification should rank by data class and autonomy. A tool receiving customer records outranks one receiving marketing copy. A tool that takes actions outranks one that returns text.

Then close the loop. Every blocked tool needs a sanctioned alternative announced at the same time, or the usage moves to personal devices and your visibility drops to zero.

Akto Atlas runs the discovery layer continuously rather than as a point-in-time audit, surfacing new AI tools, extensions, and connected agents as they appear along with the data classes they receive.

Discovery precedes control. Start counting before you start writing policy.