Copilot Cowork is GA, but the headline for engineers is simple: you need a Microsoft 365 Copilot license first, and then Cowork bills separately at $0.01 per Copilot Credit on PayGo. That means this is not just “Copilot got a new feature.” It is a second meter, with a real FinOps and governance footprint, sitting on top of an already premium license.
That does not make it bad. It does make it something you should price and govern before you let users loose on it.
What is Copilot Cowork, really?
Cowork is Microsoft’s agent runtime for long-running, multi-step work across Microsoft 365. The practical distinction versus Copilot Chat is that Cowork is meant to do the task, not just suggest the next step. Microsoft’s own docs position it as work that can run for minutes to hours, across files, meetings, Teams, calendars, browser actions, plugins, and connected systems.
That is the good part. For teams already building Microsoft Copilot and AI agents, this is closer to an actual execution layer than the usual “AI assistant” wrapper.
At GA, Microsoft says Cowork supports usage-based billing for model responses, tool and skill calls, image generation, and browser tasks. It also adds model choice, plugins, Dynamics 365 integration, and local browser use through Edge.
How much does Copilot Cowork cost?
The official pricing facts we have today are these:
| Item | What Microsoft says |
|---|---|
| Base requirement | Microsoft 365 Copilot User Subscription License required |
| Usage billing | Separate from the license |
| PayGo rate | $0.01 per Copilot Credit |
| Prepurchase option | P3 annual commitment, discounted versus PayGo |
Microsoft also documents that P3 is layered on top of PayGo, not a separate billing system. When prepaid credits run out, usage continues on pay-as-you-go unless you stop it with policy.
That last part matters more than the marketing. The bill does not stop because your pilot spreadsheet was wrong.
Microsoft claims Cowork tested 30-40% cheaper than Claude Cowork with the Microsoft 365 connector in an internal 125-run comparison using Opus 4.8. Fair enough, but I would not anchor a budget on a vendor-run benchmark. What matters in practice is your task mix, approval flow, data retrieval pattern, and whether users discover that “just one more run” is someone else’s Azure-backed spend.
If you are serious about this, treat it like any other consumption service and budget it alongside AI workflow automation, not as a vague add-on to seat licensing.
What changed at GA that admins should care about?
The best GA decision is that Cowork is off by default. Keep it that way until you have policies.
Microsoft says admins can set spending limits at tenant, group, and user level, configure alerts, review credit requests, and monitor usage in the Microsoft 365 admin center. Cowork usage can also be tied to Azure subscription-backed billing for scale.
Security-wise, the story is mixed but better than many agent launches:
- audit log, eDiscovery, Insider Risk Management, DSPM, Communication Compliance, and Data Lifecycle Management are in scope
- prompts, responses, and generated artifacts flow through existing Microsoft 365 controls
- sensitivity labels are inherited end-to-end
- DLP is still listed as coming soon for Cowork GA scenarios
That last point is the one I would watch most closely. An agent that can act across files, messages, browser sessions, and plugins should not get a blank check to your tenant. If you built your control assumptions around DLP being there on day one, verify them. For some orgs there may be little impact; for regulated environments, that gap is not minor. An AI and automation audit is cheaper than finding this out during an incident review.
The part I would be careful with
Browser use via local Edge is powerful, and Microsoft deserves credit for making it inherit existing Edge, Entra, Conditional Access, and browsing restrictions. The docs also say the feature is disabled by default.
Still, this is where second-order risk shows up first. Once users see an agent can click through web workflows for them, expectations jump fast. Then plugin sprawl starts. Then least-privilege discussions arrive late. Then the finance team asks why “a few helpful automations” now have a monthly burn rate nobody forecast.
In practice, Cowork looks genuinely useful for high-friction knowledge work. But it also looks like a product that needs the same design discipline you would apply to custom MCP server development: explicit scopes, clear approvals, usage caps, and a default assumption that metered AI features expand unless someone actively contains them.
My take: good runtime, sensible admin controls, but the commercial model is the real product decision. If you already pay for Microsoft 365 Copilot and want an agent that can actually execute work inside your trust boundary, Cowork is worth testing. Just do not roll it out as “included AI.” It is not included, and the meter is the feature you need to understand first.
