What it does

Cowork is the agentic direction of Copilot: agents that take on a multi-step task and carry it through with limited hand-holding, working across your M365 content rather than answering a single prompt. It's the shift from assistant to something that executes on your behalf.

Key facts

• The model is delegation — you hand off a task, the agent plans and works it, and you review the result.
• Agents operate within the user's permissions and the tenant's data boundary.
• It builds on the broader Copilot agent platform (Researcher, Analyst and custom agents from Copilot Studio).
• This is a fast-moving, partly early-access area; capability and naming are still shifting.

When to use / skip

Use it for well-bounded, repeatable multi-step work where you can clearly describe done. Skip it for anything high-stakes or ambiguous where an unsupervised agent going slightly wrong is expensive to unwind.

Configuration decisions

• What an agent is allowed to touch and do unattended.
• The human-in-the-loop checkpoints for anything consequential.
• Which teams get to trial it during early availability.

Gotchas

• Autonomy raises the stakes on the oversharing problem — an agent ranging across your Graph exposes weak permissions faster than a person would.
• "Agentic" is doing a lot of marketing work right now; verify what actually runs unattended versus what still needs a click.
• Auditability of agent actions is something to check, not assume.

Consultant notes

• Treat this as the strategic horizon, not a today-feature for most tenants. Set that expectation with stakeholders.
• The governance conversation is bigger than for chat — permissions, audit and accountability all sharpen when agents act on their own.
• Nail your data hygiene before you let agents loose. Everything wrong with your permissions gets amplified.

---

*Worth revisiting after the nex