Implementation Guide

Governing AI Agents in Microsoft Purview

Microsoft Purview governs the data AI agents access and the activity logs they generate, through DSPM for AI, sensitivity labels, DLP, Audit, Insider Risk Management, and Communication Compliance. It does not natively control runtime behavior such as tool-call permissions, least-privilege execution, or agent-to-agent interactions, which require a separate runtime governance layer.

Purview Capabilities That Apply to AI Agents

Four Purview capability areas are documented as applying directly to Microsoft 365 Copilot and Copilot Studio agents. Each addresses data visibility, access control, or logging rather than runtime execution.

DSPM for AI

Visibility into agent usage and sensitive data exposure.

Purview Audit

Unified log entries for Copilot prompts and responses.

DLP & Sensitivity Labels

Restricts and labels content Copilot can reference.

Insider Risk & Communication Compliance

Alerting on risky generative AI interactions.

Where Purview's Coverage Ends

Several architectural gaps are evident when mapping Purview's documented feature set against the operational needs of autonomous agents.

  • Tool-call permissioning and least-privilege execution

    Not addressed by any current Purview capability. The product scopes consistently to data classification, access control, and logging rather than execution control.

  • Agent-to-agent interactions

    Not explicitly addressed in current Purview documentation, representing an unconfirmed coverage area rather than a stated gap.

  • Third-party or custom-built agents

    Coverage of agents operating outside Microsoft 365 Copilot and Copilot Studio appears more limited than for first-party Microsoft agent surfaces.

  • Entra Agent ID

    Introduces agent identity governance as a separate capability track from Purview, requiring manual integration to achieve combined identity-plus-data oversight.

Plan around confirmed boundaries

Organizations should treat these as confirmed boundaries of native tooling rather than temporary limitations, and plan architecture accordingly.

Operationalizing Purview Controls for AI Agents

Together, DSPM for AI, Purview Audit, DLP and sensitivity labels, Insider Risk Management, and Communication Compliance form the data governance baseline for Microsoft 365 Copilot and Copilot Studio agents. Each control addresses what data an agent can access, how that access is logged, and how risky interactions are flagged after the fact. None of them, individually or combined, governs what an agent is permitted to execute at runtime. Teams rolling out agents should treat Purview as the system of record for data exposure, and pair it with a dedicated runtime governance layer for tool-call permissions, least-privilege execution, and agent identity oversight.

Data-Layer Governance Versus Runtime Governance

Purview's AI-related capabilities are built to govern data: what agents can see, what gets logged, and what gets flagged. Runtime behavior, including tool permissions and agent-to-agent calls, sits outside that scope. The questions below address the most common points of confusion.

Does Purview block AI agent actions in real time?

No. Purview's AI-related capabilities, including Audit, Insider Risk Management, and Communication Compliance, provide after-the-fact visibility and alerting rather than real-time interception of agent actions. DLP policies can prevent Copilot from referencing certain content, but this is a data-access restriction, not a runtime execution control.

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

Purview's documented feature scope is strongest for Microsoft 365 Copilot and Copilot Studio agents. Coverage of third-party or custom-built agents operating outside these surfaces appears more limited and should be confirmed against current Microsoft documentation.

How does Entra Agent ID relate to Purview governance?

Entra Agent ID is a separate identity-layer capability that assigns distinct identities to agents built in Copilot Studio and Azure AI Foundry. It complements Purview's data controls but does not replace them; the two must be integrated manually today.

What governs tool-call permissions and least-privilege execution for agents?

Based on current documented Purview capabilities, native tooling does not address tool-call permissioning or least-privilege execution enforcement. These runtime behaviors require a separate governance layer outside Purview's data-centric controls.

Close the Runtime Governance Gap

Purview governs the data layer around your AI agents. Runtime behavior, tool permissions, and agent-to-agent interactions require dedicated runtime governance.

Explore Runtime Governance