Publication · Microsoft Copilot Studio · BYOP Preview

Give Copilot Studio tools
an independent policy decision

For eligible generative-orchestration agents, Copilot Studio can send every proposed tool invocation to an external provider before execution. AgenticDome turns that native Microsoft extension point into a tenant-governed allow-or-block decision—without rewriting each covered tool—using the user, agent, conversation, tool, reasoning, and proposed inputs Microsoft supplies.

AgenticDome Research·2026·Approx. 8 minute read

A safe prompt can still produce the wrong business action

Microsoft documents built-in UPIA and XPIA protections for Copilot Studio custom agents. Those controls matter, but prompt safety is not the whole authorization question. An agent can receive an ordinary request and still select a tool, amount, record, destination, or scope that the business should not approve.

Copilot Studio’s external threat-detection extension creates a decision point immediately before the tool runs. AgenticDome uses the proposed action and Microsoft-provided context to apply the customer’s policy, return allow or block, and create tenant-scoped evidence alongside the organisation’s other AgenticDome-protected agents.

What changes when AgenticDome is in the path

On the supported BYOP path, Copilot Studio calls the assigned AgenticDome endpoint automatically before a proposed tool invocation. Teams do not rewrite each covered tool to call AgenticDome; administrators establish Entra trust, register the endpoint for each environment, and choose the outage behavior. AgenticDome then evaluates the action through the same tenant policy plane used for supported SDK, MCP, and agent-runtime integrations.

Why AgenticDome is valuable in a Microsoft agent estate

AgenticDome complements Microsoft-native security with a business-action decision that can be shared across Microsoft and non-Microsoft agents. The integration uses Microsoft’s own external-provider contract rather than inserting an unsupported proxy into Copilot Studio.

Configuration-led adoption

AgenticDome provides the assigned HTTPS endpoint and guided federated-credential values. The covered pre-tool path is enabled through Entra and Power Platform administration rather than changes to each agent’s tool code.

Policy before business impact

Evaluate the proposed tool and inputs in the context of the user, tenant, agent, conversation, previous tool output, and stated reasoning before execution.

One policy plane

Bring Copilot Studio tool decisions into the same tenant governance and evidence model used across supported Python, TypeScript, MCP, and cloud-agent integrations.

Deliberate failure posture

Choose whether provider timeout or error should block the query, then test the response deadline and availability instead of accepting an accidental fail-open posture.

The setup is administrative, not an agent rewrite

The customer receives an assigned AgenticDome endpoint, establishes the Microsoft Entra application and Federated Identity Credential, enables Additional threat detection for each intended Power Platform environment, and validates allow, block, timeout, and error behavior. AgenticDome’s guided setup prepares the endpoint and exact trust values; Microsoft still requires per-environment administrator action.

How AgenticDome creates the control point

1. Environment setup

A Power Platform Administrator enables the external provider separately for each environment.

2. Entra trust

A Federated Identity Credential provides secretless authentication to the configured endpoint.

3. Proposed tool

Copilot Studio sends recent context, identities, prior outputs, reasoning, tool, and inputs.

4. Allow or block

AgenticDome evaluates the proposed invocation and returns the provider response contract.

5. Evidence and review

Teams validate tenant mapping, availability, data handling, and the configured outage posture.

What this changes for real teams

Power Platform admin

Add protection without rebuilding the agent

“We want a supported control before tool use, but we do not want to rewrite every tool implementation.”

AgenticDome contribution: Configure the assigned AgenticDome endpoint and federated trust through Microsoft’s native external-provider workflow. Copilot Studio then invokes the decision service automatically for eligible proposed tool calls.

Security team

Apply business policy to proposed actions

“Built-in prompt protections do not express all of our rules for refunds, exports, account changes, or privileged tools.”

AgenticDome contribution: AgenticDome evaluates the proposed tool, inputs, actor, agent, tenant, and available conversation context against tenant policy before Microsoft allows the tool to proceed.

AI governance team

Avoid a Microsoft-only control silo

“Our governance evidence should cover Copilot Studio alongside the Python, MCP, cloud, and custom agents used elsewhere.”

AgenticDome contribution: AgenticDome gives supported agent environments a shared policy vocabulary and tenant evidence model while retaining the native enforcement point appropriate to each platform.

The risks this helps contain

RiskWhy it mattersPublic AgenticDome control
Business-inappropriate tool call A prompt can be benign while the proposed action or arguments violate policy. Evaluate the tool, inputs, actor, agent, tenant, and available conversation context before execution.
Fail-open timeout Microsoft documents a one-second default wait and Allow the agent to respond as the default error behavior. Choose Block the query where appropriate, engineer the endpoint for the latency budget, and test timeout/error paths.
Coverage assumption External threat detection is skipped for classic agents and is configured per environment. Inventory generative-orchestration agents and environments; do not report estate-wide coverage from one successful setup.
External data sharing Context sent for a decision can include prompts, history, prior tool output, identity, reasoning, and inputs. Review provider terms, residency, minimization, retention, access control, and the customer’s regulatory obligations.
Output-policy gap The BYOP call documented by Microsoft occurs before tool invocation, not as a universal returned-content DLP hook. Protect output through separate supported application boundaries where available; do not claim BYOP alone sanitizes every tool response.

A practical path to production

Confirm eligibility

Verify preview availability and generative orchestration for every intended agent.

Configure FIC

Use the documented issuer, subject, App ID, and exact provider base URL.

Set each environment

Enable data sharing and choose error behavior separately in Power Platform admin center.

Test the deadline

Measure normal, blocked, timeout, and provider-error behavior within Microsoft’s response window.

Where this control ends

This integration covers the external threat-detection request Microsoft sends before a proposed tool invocation for eligible generative-orchestration agents. It does not cover classic agents, guarantee every tool or output path, replace Microsoft-native protections, provide a tenant-wide enablement switch, or remove the customer’s responsibility for Entra, data governance, environment inventory, availability, and provider review.

Current lifecycle note

Microsoft labels external threat detection and protection for Copilot Studio as preview, prerelease, subject to change, and not intended for production use. Recheck the official documentation before each rollout or public coverage statement.

Primary references and public implementation

Make Copilot Studio tool use answer to your policy before it acts.

Use AgenticDome’s assigned endpoint and guided trust values, enable Microsoft’s native external-provider path for each eligible environment, and prove allow, block, timeout, and evidence behavior before relying on it.