Find the AI nobody asked permission for. With the evidence attached.

Most AI tools are adopted by clicking "Sign in with Google" or "Sign in with Microsoft". That click leaves a record in your identity provider, and the record says more than that the tool exists: it says what the tool was allowed to reach.

Connect Google Workspace with a service account you create, limited to one read-only scope for the admin audit log. Connect Microsoft 365 with an app you register in your own tenant, with read-only directory permission. The first sync runs when you connect, and each AI tool found becomes a finding with its evidence: how many people granted it access, since when, and which permissions it holds. After that, every connected source is synced automatically once a night, and Sync now on the Integrations page runs one whenever you want.

Permissions drive the suggested risk. A writing assistant that can only see a name and an email address is one thing. A meeting assistant that was granted read access to calendars and files is another, and its finding says so in plain words.

What is read, and what is not. Application names and ids, permission names, the identity of the person who granted access, counts and timestamps. No mail, files, calendar entries or prompt content are requested from either provider. People are stored according to your identity mode: emails in named mode, the same salted hashes the sensors use in pseudonymous mode, and counts only in anonymous mode.

Limits. Google keeps these audit events for 180 days, so a grant made earlier is seen only if the app was authorised again or was active recently. Microsoft sign-in counts need Entra ID P1 or P2 and an additional permission; without them you still get applications, grants and assignments. Neither source sees a tool that someone signed up for with a personal email address; that is what the sensors are for.

A catalog is always a little behind. New AI products appear every week, and an identity provider reports every app in your tenant, most of which have nothing to do with AI.

So the apps that Google Workspace, Microsoft 365 or Okta report and the catalog does not recognise are screened by fixed rules: AI terms in the name, a domain that reads like an AI product, or broad access to mail, files or calendars. Those that look AI-related are listed as unclassified apps on the Discovery page, each with the reasons it was listed.

An admin or compliance manager goes through the list. Confirm turns the app into a finding in the review queue and adds it to your organization's own catalog, so the next sync recognises it. Dismiss removes it from the list. Nothing is added on its own.

Only the app's name, public domain and permission labels are kept for this list, not who uses it. The organization catalog is used when identity providers are synced; the browser extension and desktop agent keep to the built-in catalog.