Microsoft AB-100: DLP Policies for Copilot Studio
Data loss prevention in Copilot Studio is not limited to blocking connectors. Power Platform data policies can govern authentication modes, knowledge sources, connector tools, HTTP requests, skills, publication channels, and event triggers. The goal is to prevent an agent from becoming an easy path between organizational data and services that should not be combined.
On this page
To design a Copilot Studio data policy, trace a real business interaction from the signed-in user through its knowledge sources, connector calls, API endpoints and publication channel. Record which identities can read each source, where sensitive information could cross a trust boundary, and what must fail if a connector is blocked. Use that model to configure tenant- or environment-scoped policies before an agent is released. A policy that blocks a required capability can affect makers and people already using the agent, so test both permitted and denied paths before changes reach production. Because the earlier agent-exemption route is retired, legitimate exceptions require deliberately scoped policy decisions or an agent redesign rather than silent bypasses.
Group connectors by data boundary
Power Platform data policies classify connectors into Business, Non-business, and Blocked groups. Connectors in different groups cannot exchange data through the same governed experience. Data classification should inform those groups so DLP reflects actual business sensitivity rather than a generic list of “safe” and “unsafe” products.
DLP design should begin with data flows rather than connector names. Identify which business data can move to which systems, which combinations are prohibited, and where exceptions are justified. The Business, Non-business, and Blocked groups then become an implementation of that policy rather than an arbitrary taxonomy.
Default connector grouping deserves review because new connectors can inherit the policy’s default group. A conservative default can prevent accidental exposure, but it can also break new production capabilities when administrators are unaware that a connector landed in Non-business or Blocked. Governance teams should monitor new connector adoption and classify intentionally.
DLP rules should be reviewed with solution architects and process owners, not only security administrators. A connector that looks risky in isolation may be essential to a controlled business workflow, while a seemingly harmless endpoint may create an unexpected data path when combined with another tool. Policy quality improves when it reflects how agents actually work.
Document why important connectors are in each group. This becomes valuable when an administrator inherits the policy later or a vendor introduces a new connector with a similar name. A short rationale prevents future changes from weakening the boundary through misunderstanding.
Data policies should be documented alongside architecture diagrams. The diagram should show which connectors, channels, knowledge sources, and HTTP endpoints the agent requires, while the policy record shows how those paths are classified. This makes security review concrete and helps explain why a seemingly harmless change cannot be published.
Require authentication when needed
Administrators can use data policy to block the Copilot Studio capability that allows chat without Microsoft Entra authentication. This prevents makers from publishing an agent that accepts unauthenticated access when organizational policy requires signed-in users. Agent authentication and DLP should be reviewed together because the policy can enforce which authentication modes are available to makers.
Govern knowledge sources
Copilot Studio data policies can block local document uploads, SharePoint or OneDrive knowledge, public websites, and other supported knowledge patterns. Endpoint filtering can allow or deny specific SharePoint sites or public web endpoints without blocking an entire knowledge category.
Purview governance complements connector restrictions by classifying sensitive material and controlling its retention, auditing and permitted use. A DLP rule can prevent an unapproved connector path, but it does not repair an overexposed SharePoint library or ensure that information labels and lifecycle rules match business obligations.
Control connector tools
Administrators can block Power Platform connectors from being used as agent tools or place them in appropriate data groups. This matters when a connector can move organizational data into an external service or perform a write operation. Copilot actions should be designed within the approved connector policy rather than assuming every available connector is permitted in production.
Use endpoint filtering for HTTP
Blanket blocking of HTTP can be safe but may be unnecessarily restrictive for mature teams. Copilot Studio data policy supports endpoint filtering so administrators can allow or deny specific HTTP destinations. This creates a more precise boundary around custom APIs while preventing makers from calling arbitrary endpoints.
Endpoint filtering is particularly useful when a connector category is too broad. An organization may allow HTTP only to approved internal APIs or permit SharePoint grounding only from designated sites. This creates a more precise control surface than allowing every endpoint simply because the connector itself is trusted.
Use least-privilege endpoint patterns. Allowing one trusted API hostname is safer than allowing an entire wildcard domain when the agent only needs one service. Narrow rules also make incident containment easier because an administrator can revoke one path without disabling unrelated integrations.
Control channels and triggers
Policies can restrict publication to channels such as Direct Line, Teams and Microsoft 365, SharePoint, WhatsApp, and other supported endpoints. They can also block event triggers, which matter because an autonomous agent can run without an interactive user present.
Process-shaped DLP evaluates not only a connector name but the movement of business data. A customer-service agent may read case status through an approved tool while being prohibited from exporting full case notes to an external endpoint. That distinction should be implemented in scoped connector and endpoint policy, with downstream authorization still enforced by the target service.
Plan for enforcement in production
A data policy violation can prevent publishing or cause a runtime error for users. Administrators should review policy changes against production agents before broad rollout, especially when default connector groups or endpoint rules change. DLP is a security control, but poorly planned changes can create production outages.
Policy changes need impact analysis. Before moving a connector between groups or blocking a channel, identify which published agents depend on it. Since Copilot Studio enforces policy in real time and exemptions are no longer supported, a governance change can become an immediate production incident.
Exception governance is still necessary even without technical exemptions. A business may have a legitimate reason to permit a connector or endpoint that is normally restricted. Document the risk, scope the policy narrowly, assign an owner, and review the exception periodically instead of weakening the tenant-wide baseline indefinitely.
Use test agents or nonproduction environments to validate new policies before tenant-wide enforcement. Check both positive and negative cases: approved connectors should still work, blocked paths should fail, and the maker error should be understandable enough to guide remediation.
Give makers useful error context
Copilot Studio surfaces policy violations and allows administrators to configure help links and contact details for policy error messages. This matters because makers need a sanctioned path to request a policy change or redesign the agent. Governance works better when the error explains the boundary than when teams respond by searching for an unofficial workaround.
Production support should know how DLP violations appear. Makers can see publish-time policy details, while end users may receive a DataLossPreventionViolation error. Runbooks should tell administrators where to inspect the violation, which policy applies, and whether the correct response is policy change or agent redesign.
Use zones to vary governance depth
Different environments can have different data policies and governance expectations based on their purpose and risk. Environment strategy can separate personal experimentation, team development, and governed enterprise deployment. For current Microsoft business AI architecture, DLP should protect data without making every environment identical. Strong programs use clear zones, deliberate connector groups, endpoint filtering, runtime enforcement, and a review path that lets useful agents move safely toward production.
Enforce DLP on autonomous workflows
DLP should also cover autonomous use cases. Event triggers can cause agents to act without an interactive user, which increases the importance of controlling connectors, destinations, and data movement. Blocking or limiting trigger-related capabilities can be appropriate in lower-trust environments.
Finally, DLP should be tested as part of ALM. A release that adds a new connector or knowledge source should verify policy compliance in the destination environment before broad publishing. Security and deployment become much easier when the agent’s intended data paths are known and approved before the Publish button is reached.
Update DLP as agents evolve
Keep DLP aligned with lifecycle. When an agent is retired, its special endpoint rules or exception-oriented environment policy may no longer be necessary. Removing obsolete allowances reduces the long-term governance surface and keeps the policy understandable.
Test policy inheritance when environments or tenant scope change. A policy that applies to all environments can affect new environments automatically, while an environment-specific policy can create behavior different from the rest of the tenant. Makers should know which policy set governs their target environment before they design around a connector.
DLP review should include user experience. If a policy requires Microsoft authentication, blocks a channel, or removes a knowledge source, the product team should understand what users will see. Good governance avoids surprise by aligning policy, publishing, support documentation, and adoption communication.
For high-risk agents, combine DLP with lifecycle and monitoring. DLP constrains data movement, but owners still need to review unusual usage, connector changes, and new business requirements. Governance is strongest when policy enforcement and operational oversight reinforce each other.