AI, in your work.

What the reference describes

The Microsoft architecture proposes a Copilot Studio orchestrator that routes requests to agents covering personal data, time off, manager actions, learning, and HR policies. The intended entry points are Teams and Microsoft 365 Copilot. It is a reference for designing a solution adapted to each organization.

Separating tasks also adds work

Microsoft warns that orchestration across agents can increase latency and the effort required for testing, management, and governance. The split should be justified through evaluations of the assistant’s behavior.

Nexo interpretation: start with responsibilities. Separating a policy question from a data update can help give each task defined sources, permissions, and owners. Adding agents without clarifying those boundaries creates more components to maintain.

In the documented standard experience, connecting another Copilot Studio agent requires it to be published, in the same environment, and enabled to accept connections. The maker must own it or have it shared with them; depending on authentication, it may also need to be shared with end users.

Identity and permission to act require separate checks

User authentication and tool authentication have their own configurations. Microsoft allows tools to use author credentials or a user connection, recommending the latter when access must be restricted to particular people or groups.

Recognizing an employee in the conversation therefore does not, by itself, establish permission to change a record in the HR system. Each operation needs a check of the identity it uses and the permissions enforced by the destination service.

Nexo’s proposal for the design workshop: record the authoritative system, operating identity, required permission, and response to denied access for each task. HR should also define which subjects require human involvement.

Two journeys to discuss during design

The following is an original fictional example from Nexo. It describes proposed behavior, not a completed test or capabilities automatically available when an assistant is created.

Evaluate answers and operations

The Copilot Agent Kit documentation describes batch tests with observed responses, latencies, and pass or fail results. It is a tool for gathering evidence; acceptance criteria must reflect the service being designed.

Nexo proposes repeating a stable set of cases before each release and after relevant changes:

• An unambiguous question and an ambiguous one: check agent selection and clarification questions.
• A user without permission: check that the operation is denied without exposing restricted data.
• An external system that does not respond: check that the failure is communicated without an invented outcome.
• A request requiring human support: check that escalation provides a usable next step.

Availability and limits before deployment

The licensing documentation distinguishes the Copilot Studio for Teams plan from the standalone subscription: the former does not include generative orchestration. The trial license allows agent creation and testing, but not publishing. Choosing Teams as a channel does not, by itself, establish the required plan.

The agent guidance identifies connections to Microsoft Foundry, Fabric, and Microsoft 365 Agents SDK agents as public previews. It also warns that citations may not be retained when answers are returned between agents.

A deployment decision requires checking the specific features selected, connectors, administrative policies, and licensing terms for the environment. Nexo’s editorial recommendation is to give the first pilot a limited scope and explicit acceptance criteria: a convincing conversation does not establish that permissions, operations, and escalation work correctly.