Field guide · three modes

Chat, Agent or Cowork?

Choose by the unit of work: a conversation, a repeatable role, or a delegated outcome.

These capabilities overlap. The useful distinction is not what each can technically do, but how you work with it.

The operating model

  • CHAT

    You steer the conversation.

  • AGENT

    You reuse a defined specialist.

  • COWORK

    You delegate an outcome and supervise checkpoints.

The three modes

Copilot Chat

A general conversational AI for questions, exploration, drafting, summarizing and fast feedback.

Best when
The work is one-off, still ambiguous, or benefits from back-and-forth.
You provide
A question, prompt and relevant context.
You get
An answer, draft, synthesis or next move, usually within one session.
Control model
You drive every turn and decide what happens next.
Nuance
With Work IQ on, licensed users can ground Chat in accessible work content such as email, chats and files; with it off, Chat is grounded in the web.
Not this
A durable workflow just because you wrote a good prompt once.

In this workshop

“Help me search broadly across my weekly frustrations, then narrow them to one sharp Pain Card.”

Copilot Agent

A reusable specialist configured for a defined purpose, with instructions, selected knowledge and, where enabled, actions.

Best when
The same bounded job, method or standard should be applied repeatedly.
You provide
A role, instructions, knowledge, boundaries and examples.
You get
Consistent handling of the same class of problem; the exact response or action depends on the agent's configured capabilities.
Control model
An owner designs and governs it; a user invokes it and reviews the result.
Nuance
An agent can be simple or powerful. The distinguishing idea here is reusable specialization, not “a bot that works by itself.”
Not this
Automatic proof that a process is safe, correct or ready to scale.

In this workshop

“Run the Painkiller Solution Architect every time a new Pain Card arrives, and return one Build Brief.”

Copilot Cowork

A delegation surface for multi-step work across Microsoft 365. You describe the outcome; Cowork plans, executes and produces work using available skills and apps.

Best when
The outcome spans several steps, sources or applications, or should continue while you focus elsewhere.
You provide
The outcome, context, boundaries, quality bar and checkpoints.
You get
Completed artifacts and, when authorized, actions across Microsoft 365; work may take minutes to hours.
Control model
You can follow, steer, pause and review. Cowork pauses for approval before important actions such as sending or scheduling.
Nuance
Availability depends on license, tenant enablement, permissions and usage-based billing.
Not this
Invisible autonomy or guaranteed correctness. Finished work still needs human judgment.

In this workshop

“Take the approved Build Brief, create the smallest useful version, run it once, show the evidence and improve one thing.”

Which one now?

Pick the shape of the task. Nothing is saved — this is just for you.

Select a task shape to see the recommended mode.

Start with the smallest mode that fits. When the conversation becomes a repeatable method, make an Agent. When the job becomes a multi-step outcome, delegate it to Cowork.

The Pain Killer, translated

  • Chat finds and sharpens the pain.
  • The Agent applies a reusable design method and returns the Build Brief.
  • The Build Map is the human decision: solution pattern and delivery bar.
  • Cowork builds, runs, shows evidence and iterates.

Chat explores. Agent standardizes. Cowork executes. Leadership still chooses the outcome, boundary and evidence.

Reality check

Use what is actually switched on. Licensing, admin enablement, tenant configuration, permissions and governance determine what each participant can access.

Official sources

Checked 1 Sep 2026.