A global manufacturer is rolling out a Microsoft 365 Copilot agent for supplier onboarding that must answer policy questions and trigger approved…

A. An app-packaged declarative agent definition (e.g., a manifest file) enables consistent, repeatable deployments and central governance of the agent's capabilities across dev/test/prod environments, which is a core requirement.

C. Using connector-based grounding with Access Control List (ACL) trimming allows the agent to access supplier data in-place, respecting data residency and the "no new data movement" constraint, while enforcing user-specific security permissions.

D. An action allowlist with approval gates provides a robust safety and governance mechanism. It restricts the agent to only pre-approved actions and enforces separation of duties for high-risk operations, meeting a key safety requirement.

B. Per-user prompt override in Teams chat threads: This approach undermines central governance, making the agent's behavior inconsistent and difficult to audit, which is contrary to the requirements.

E. Separate agent per subsidiary tenant: This would lead to fragmented management and violate the "centrally governed" requirement, increasing operational complexity and cost.

F. Direct embedding of SOPs into prompts: This is not a scalable or maintainable solution for grounding. It is subject to token limits and is far less effective than using a dedicated retrieval mechanism like a connector.

1. Microsoft Copilot Studio Documentation, "Plugins for Microsoft Copilot". This explains how plugins (actions) can be defined and governed, including the use of allowlists and approval flows for safety. (See section: "Build and use plugins").

2. Microsoft Learn, "Implement a data connection for Microsoft 365 Copilot". This resource details how Microsoft Graph connectors index data in-place, respecting source permissions (ACLs) and avoiding data duplication. (See section: "Security and compliance").

3. Microsoft Power Automate Documentation, "Create and run approval flows". This document describes how to build approval gates for actions, which can be triggered by agents to enforce separation of duties. (See section: "Get started with approvals").