Purview can now see Claude Enterprise. Here is what you get, and what you do not

In May 2026 Microsoft added an Anthropic Claude data connector to Purview, in preview. Once configured, Claude Enterprise shows up as an Enterprise AI app alongside Microsoft 365 Copilot and ChatGPT Enterprise: who used Claude, when, and what content was involved, visible in DSPM reports, activity explorer's AI activities tab, and the unified audit log.

The connector works through Anthropic's Compliance API rather than network interception, so it sees Claude Enterprise interactions wherever they happen, not just on managed devices or in managed browsers.

Four prerequisites, and the first one rules out most tenants today:

An Anthropic Enterprise plan. Not Free, Pro, Team, or Max. The Compliance API only exists on Enterprise, and the primary owner has to enable it under Organization Settings, then generate a Compliance Access Key for the connector.

Pay-as-you-go billing consent in Purview. Managing Claude interactions runs on the consumption billing model, so finance needs to be in the conversation before you switch it on.

The Data Connector Admin role for whoever creates the connector under Settings in the Purview portal.

Patience. Data starts appearing in DSPM about a day after setup.

During setup you choose what to ingest: the activity feed (sign-ins, chat events, file uploads, admin actions across Claude.ai, the Console, and the API) or full chat conversations including prompts and responses. The second is what makes the visibility meaningful, and it is also the one to clear with your privacy team first.

Microsoft publishes a capability matrix for each AI app, and Claude's is mostly crosses. Supported: DSPM and auditing. Not supported: data classification, sensitivity labels, DLP, Insider Risk Management, Communication Compliance, eDiscovery, and Data Lifecycle Management.

In practice that means you can see a user pasted something into Claude, but Purview will not classify what it was, will not block it, will not flag the user in IRM, and cannot put the conversation on legal hold.

Compare ChatGPT Enterprise, where the connector supports classification and Insider Risk Management. Claude's integration is newer and shallower. Claude agents are not covered at all yet, so they will not appear in AI observability either.

The connector is a witness, not a control. Plan accordingly.

Visibility on the Enterprise side pairs with enforcement on the browser and network side:

Inline web DLP can audit or block sensitive text and files heading to Claude.ai in the browser, on the unmanaged-apps path through Edge or via the Network path for other browsers. That covers consumer Claude too, which the connector never sees.

The connector then tells you what got through on the Enterprise plan, with the audit trail to investigate.

This split matters for policy design: blocking happens before the prompt is sent, the connector records what was sent. Neither replaces the other.

Watch the capability table as the preview matures. ChatGPT Enterprise's deeper support suggests where Claude's integration is likely heading, and classification support would be the signal that real controls are coming.