Microsoft 365 Copilot works with information available through its supported services and the user’s permissions. A governance review needs to inspect those permissions as well as the AI interaction.
Microsoft states that prompts, responses and Microsoft Graph data accessed by Microsoft 365 Copilot are not used to train foundation language models. That is distinct from the storage, retention, access and processing arrangements for user interactions. Read the current Microsoft privacy and security documentation for the service and configuration in use.
Review source access first
Identify the sites, files, mail and other sources included in the intended use. Check broad sharing, guest access, group membership and stale permissions.
Copilot respecting permissions does not mean those permissions are appropriate. An employee may already have access to material that the business did not intend them to find easily. Search and summarisation can make that oversharing more visible.
Use a representative set of accounts to test expected access and denied cases. Include people who change roles or leave a project. Keep permissions changes and source ownership in the operating process, rather than treating a pre-launch scan as permanent assurance.
Be precise about sensitivity labels
A label’s name alone does not enforce a prohibition. The effect depends on the configuration, encryption rights, source and supported application surface.
Review how the current Copilot experience handles labelled and encrypted content. Do not apply an old rule such as “removing one encryption permission blocks every Copilot route” without testing the actual content and experience. Microsoft documents important considerations for Microsoft 365 Copilot, including product-specific behaviour.
Test a protected file in the intended experience using an authorised account and a denied account. Record the result, policy version and licence assumptions. A screenshot of a configured label is not equivalent to an enforcement test.
Apply data policies to the correct surface
Microsoft 365 content controls, Copilot Studio connector policies and Power Automate policies are related governance concerns but not one interchangeable switch. Identify the product, environment, connector and action you are controlling.
Copilot Studio data policies can classify and constrain supported connectors. Review the current data-policy guidance and test both creation and execution paths relevant to the deployment. Power Automate has its own runtime data-policy enforcement details.
Feature availability, licensing and rollout can differ. Record those conditions rather than promising that a policy applies identically to every agent or flow.
Treat connected actions as a separate boundary
A connector can move data or change a record outside the immediate chat experience. Review its authentication model, permissions, allowed operations and recipients.
Use narrowly scoped access and validate parameters before consequential actions. Where review is required, show the proposed change and relevant evidence to the authorised person before execution.
Agent identities should be understood through the actual supported platform. An Entra agent identity is not simply another name for a managed identity, and not every automation automatically receives one. Consult the Entra agent identity documentation for that deployment.
Protect records and investigate actual events
Determine which interaction and activity records are retained, where they are searchable and who can access them. Apply the organisation’s retention, investigation and privacy requirements to those records.
Capture enough evidence to investigate source access and observable actions. Do not claim that a transcript or generated rationale reveals every internal reason for the model’s answer.
Define who can revoke access, disable a connector or pause an affected workflow. Check how external actions are reconciled if an interruption leaves their outcome uncertain.
Keep security review tied to changes
A new data source, connector, feature or permission model can change the exposure. Maintain an inventory and repeat relevant tests when those boundaries change.
This guide explains implementation questions, not a certification that Copilot makes the organisation compliant. Ryzolv’s AI Security and Governance work can help with the agreed assessment, testing and control implementation.