Copilot Cowork is Anthropic's Claude Cowork technology built into Microsoft 365 - the same Claude models and the same skills standard, generally available since 16 June 2026. For UK organisations it arrives switched off: an administrator must enable Anthropic models before anyone can use it. Activating it well is a governance decision, not a checkbox.
On 9 March 2026, Microsoft announced Copilot Cowork - an agentic layer inside Microsoft 365 Copilot, built, in Microsoft's own words, by working closely with Anthropic to bring the technology that powers Claude Cowork into Microsoft 365. After a preview through Microsoft's Frontier programme, it became generally available worldwide on 16 June 2026.
Cowork is not another chat window. You delegate an outcome, and it plans and executes multi-step work across Word, Excel, PowerPoint, Outlook and Teams - drafting the document, building the spreadsheet, coordinating the follow-ups. Tasks run in Microsoft's cloud, so the work keeps going after the laptop closes, with visible progress and points where your people steer.
For MDs and CEOs the headline is simple: the agentic way of working we train teams in is no longer a separate tool decision. It is now built into the platform most UK firms already run their business on.
Under the surface, yes - it is the same engine behind two doors. Microsoft's door adds tenant governance and Microsoft Graph grounding; Anthropic's door is the native product. What matters commercially is what they share:
At general availability, Copilot Cowork runs on Anthropic models - Claude Opus 4.8 and Sonnet 4.6. The intelligence doing the work is the same intelligence your team would use in Claude directly.
Cowork skills use the open Agent Skills standard - the same SKILL.md format Claude uses. A skill that encodes your process, tone and structure is written once and works in both.
Both connect to your systems through MCP, the open Model Context Protocol that Anthropic created and donated. Cowork plugins mirror Claude's plugin structure.
The practical consequence: the Cowork skills your team builds with us work in both. Claude-native or inside Microsoft 365 - the capability is portable, and so is your investment in it. That is exactly the portability we argue for in why we specialise in Claude.
Copilot Cowork is priced in three layers, and the third one is the one to watch:
The base subscription your firm already pays for. Nothing new here - but Cowork does not come with it.
Cowork requires the Microsoft 365 Copilot licence on top - typically around £16-£25 per user per month for UK SMEs, depending on plan and route.
Cowork itself is then billed by consumption. Every task draws Copilot Credits based on model use, context retrieval, tool calls and runtime - pay-as-you-go at $0.01 per credit, or an advance commitment at a discount.
Compare that with Claude's flat per-seat subscription, where heavy use and light use land on the same predictable bill. Consumption billing is not wrong - but unmanaged credits are the new shadow spend, and the meter runs per task, per person, across the whole tenant. Microsoft provides spend caps, alerts and usage reporting; someone has to actually configure them. That is a governance job, and it is one of ours.
Anthropic models deployed in Microsoft's offerings are currently excluded from Microsoft's EU Data Boundary and its in-country processing commitments. Microsoft's response is blunt: for UK tenants (and those inside the EU Data Boundary), Anthropic models are disabled by default. Because Copilot Cowork at launch runs only on Anthropic models, that means Cowork simply does not work for a UK organisation until an administrator explicitly enables Anthropic models in the Microsoft 365 admin centre - accepting processing outside the UK and EU - with access scoped per user or group.
There is a second wrinkle: Data Loss Prevention support for Cowork was still rolling out at launch, while the rest of the compliance surface - audit logs, eDiscovery, sensitivity labels, retention - was in place. For firms in regulated sectors, that gap belongs on the decision record, not in the small print.
None of this is a reason to avoid Cowork. It is a reason to switch it on deliberately: a documented opt-in decision, scoped access, and controls configured before the first task runs. Firms that do that get the upside with their eyes open. Firms that treat it as a checkbox find out later what they agreed to.
This is our Setup & Deployment and AI Operations & Governance work, applied to the Microsoft door. Three moves, in order:
We put the opt-in decision in plain English for your leadership: what data goes where, what the exclusion from the EU Data Boundary means for your obligations, and who should get access first. You decide on the record, not by default.
Spend caps set at tenant, group and user level. Usage alerts and reporting configured so the credit meter is visible from day one. Access scoped to the people who are ready for it.
Your acceptable use policy updated for agentic work, accountability lines drawn, and your people trained to build skills that do their real job - skills that work in Copilot Cowork and Claude alike.
The result: Cowork arrives in your firm as a governed capability with a measurable return - not as an unmanaged meter and an unread consent screen.
Yes. Microsoft built Copilot Cowork in collaboration with Anthropic, and at general availability it runs on Anthropic's Claude models, including Opus 4.8 and Sonnet 4.6. In Microsoft's own words, the technology that powers Claude Cowork has been brought into Microsoft 365 - the same models that power Claude when your firm runs it directly.
Yes. Copilot Cowork requires the Microsoft 365 Copilot licence on top of your base Microsoft 365 plan - typically around £16-£25 per user per month for UK SMEs. Cowork tasks are then billed separately in Copilot Credits as your people use it, so the licence is the gate, not the whole bill.
Anthropic models are currently excluded from Microsoft's EU Data Boundary, so Microsoft disables them by default for UK tenants. Because Copilot Cowork runs on Anthropic models, it will not work until an administrator explicitly enables them, accepting processing outside the UK and EU. Access can be scoped per user or group.
Yes. Both products use the same open Agent Skills standard - the SKILL.md format - and both connect to business systems through the Model Context Protocol. A skill your team builds for Claude Cowork is structurally reusable in Copilot Cowork, which is why the capability you build with us is portable, not locked in.