Anthropic and OpenAI are competitors in the foundation-model market. Microsoft has now incorporated both into the same Microsoft 365 Copilot tenant. As of March 2026, that arrangement is in general availability. For most enterprise IT functions outside Europe, this is a straightforward improvement. For European tenants, the position is materially more complex.
What Is Actually in the Bundle
Copilot Cowork is Microsoft's name for an autonomous agent that handles long-running, multi-step tasks inside an M365 tenant, built on "the technology behind Claude Cowork". Three components shipped in Wave 3:
- Copilot Cowork in research preview, Anthropic's agentic engine running inside the M365 tenant context, initially $30 per user per month.
- Claude Sonnet selectable in Copilot Chat, alongside OpenAI's GPT family.
- Microsoft 365 E7 Frontier Worker Suite at $99 per user per month, generally available from 1 May 2026. UK and EU list pricing is not yet confirmed; at spring 2026 rates E7 lands around £78 (€90) per user per month.
The Surprise the Headlines Under-Played
Microsoft put $13 billion into OpenAI, yet in nine months it has signed a $30 billion Azure compute deal with Anthropic, integrated Claude Opus 4.6 into Microsoft Foundry, embedded Anthropic technology into Copilot Cowork, and made Anthropic an official subprocessor under its Product Terms. This is not Microsoft hedging with a second supplier; it is a committed multi-model position. At the level of platform strategy, the logic is sound for a company shipping AI to roughly 400 million seats. From a European procurement perspective, the picture is different: an EU customer is no longer engaging one US AI provider through a US cloud vendor, but two.
The Potential, Taken Seriously
There is genuine upside. Different models have different strengths: Claude performs well on long-context analysis and clean prose, while OpenAI's GPT family remains stronger on speed and ecosystem breadth. A tenant that can route work to whichever model suits the workload is, in principle, more capable than any single-model deployment. Multi-model bundling also removes the friction of separate DPAs, subprocessor registers, and audit trails. That is a reasonable case for adopting the bundle, and the compliance points below should be read alongside it.
The Data Boundary Gap
The Claude integration is disabled by default across the EU, EFTA, and the UK; an administrator must explicitly opt in. The setting lives at Admin Center → Copilot → Settings → Data Access → "AI providers operating as Microsoft subprocessors". It is the decision point at which personal data becomes eligible to be routed to a second US AI provider.
Crucially, the Anthropic integration is explicitly excluded from the EU Data Boundary. Where Claude is enabled, requests leave the boundary and inference runs on Anthropic's US infrastructure regardless of the Azure region selected. By comparison, AWS Bedrock (Frankfurt) and Google Vertex AI (Belgium) deliver EU-resident Claude inference today; the Microsoft route does not.
If your DPIA, your TIA, or your vendor risk register treats the Microsoft 365 environment as a single US-provider exposure, enabling Anthropic models invalidates that assessment.
Two US Subprocessors, One Extraterritorial Exposure
Both Microsoft and Anthropic are US corporations under the CLOUD Act, and both fall within Section 702 of FISA and Executive Order 14086, the instrument the 2023 adequacy decision rests on. The practical change for an EU controller is not the existence of CLOUD Act exposure but that the exposure surface has widened from one provider to two. Documentation needs to reflect two subprocessors, two transfers, and two extraterritorial-reach claims. An audit that finds otherwise is a finding against the controller, not the processor.
What Administrators Should Actually Do
Before changing the default configuration: map the use cases and scope enablement to Entra groups rather than the whole tenant; refresh the DPIA against the realistic worst case of what users actually paste into prompts; document a properly worked Legitimate Interests Assessment; update the subprocessor register to add Anthropic; and decide explicitly whether the transparent fallback to OpenAI when Claude errors is acceptable, communicated, and logged.
Where This Leaves European Procurement
Microsoft is shipping arguably the most polished multi-model productivity environment any vendor has produced. The same arrangement routes more workloads through two US providers rather than one. The procurement question is not whether to accept the bundle, but what plan exists for the workloads that should not run through it: sensitive HR matters, privileged legal correspondence, regulated caseloads, cross-border investigative work. A two-track architecture, productivity workloads on Microsoft and sensitive workloads on a European stack, is no longer aspirational. It is increasingly the practical answer for any organisation taking its GDPR exposure seriously.
