Govern AI Agents in Microsoft Purview: Guide
Implementation Guide

Governing AI Agents in Microsoft Purview

Microsoft Purview governs AI agents primarily through data-centric controls: sensitivity labels, DLP, Audit, Insider Risk Management, and Communication Compliance, extended to Microsoft 365 Copilot and Copilot Studio agents via Data Security Posture Management (DSPM) for AI. These controls manage data access, label inheritance, and interaction logging. They do not govern real-time agent decision-making, tool-call authorization, or agent-to-agent communication, which require separate runtime enforcement alongside Purview.

What Governing AI Agents in Purview Actually Means

Microsoft Purview approaches AI agent governance as an extension of its existing data governance stack rather than as a dedicated agent control plane. Instead of inspecting what an agent decides to do, Purview governs what data an agent can see, how classification labels travel with that data, and how prompts and responses are logged for later review.

This model is extended to Microsoft 365 Copilot and Copilot Studio agents through Data Security Posture Management (DSPM) for AI, which brings sensitivity labels, Data Loss Prevention (DLP), Audit, Insider Risk Management (IRM), and Communication Compliance into scope for AI-generated interactions. In practice, this means Purview answers questions about data access and retention well, but it is not designed to answer questions about what an agent chooses to do with that data once a session is underway.

Configuring Purview for AI Agent Governance

The core configuration surface for AI agent governance in Purview is made up of five controls. Each one addresses a different stage of the data lifecycle, from what an agent can retrieve, to what it can output, to what gets logged for investigation afterward.

Sensitivity Labels

Govern data access and label inheritance for agent-retrieved content.

DLP

Restrict sensitive content from appearing in agent prompts and outputs.

Audit (Premium)

Logs Copilot prompts and responses for investigation.

Insider Risk Management

Flags risky AI usage patterns across the organization.

Communication Compliance

Monitors prompts and responses when explicitly configured to do so.

Configured together, these controls give administrators visibility into what data agents touch and a record of what was said. They do not, on their own, constitute a decision-time policy engine for agent behavior.

Where Purview's Governance Model Reaches Its Limits

Purview's controls apply to data access and logging within Microsoft's ecosystem. They were not built to evaluate or intervene in an agent's real-time decision-making, authorize individual tool calls, or mediate communication between agents. That gap matters most in three recurring situations.

  • Custom and third-party agents: coverage depends on connector support. DSPM for AI primarily documents visibility into Microsoft 365 Copilot and Copilot Studio agents, so third-party or custom agent frameworks require additional connectors, and documented coverage for these remains limited.
  • Label inheritance: sensitivity labels do not automatically or uniformly apply to Copilot-generated content. Inheritance varies by application and configuration, so label behavior should be validated per scenario rather than assumed.
  • Agent-to-agent communication: interactions between agents, or calls out to external tools, fall outside Purview's documented scope and require separate runtime enforcement.

The practical takeaway

Purview is a strong foundation for data governance and audit trails around AI agents. Enterprises running agents that call tools, invoke external systems, or communicate with other agents still need a runtime enforcement layer alongside Purview to govern those interactions directly.

Frequently Asked Questions

Does Microsoft Purview cover custom or third-party AI agents?

Coverage depends on connector support. DSPM for AI primarily documents visibility into Microsoft 365 Copilot and Copilot Studio agents. Third-party or custom agent frameworks require additional connectors, and documented coverage for these remains limited.

Do sensitivity labels automatically apply to Copilot-generated content?

Label inheritance varies by application and configuration and is not guaranteed to be uniform across all Microsoft 365 workloads, so label behavior should be validated per scenario rather than assumed.