Should we allow Anthropic Claude models inside Copilot?

You’re right to pause. “Another model in the picker” sounds like a UX tweak; for UK and EU tenants it’s a residency and contract decision dressed as a feature flag.

Here’s what’s going on. Anthropic is onboarded as a Microsoft subprocessor for many Copilot experiences (Copilot, Researcher, Copilot Studio, Power Platform, Copilot in Microsoft 365 apps). Under that path, Microsoft Product Terms and the DPA generally apply — unless the model is labelled Anthropic models with Data Retention, which is a separate opt-in where Anthropic acts as an independent processor under Anthropic’s own terms and can retain inputs/outputs for a period. Critically, Anthropic models in these offerings are currently excluded from the EU Data Boundary (and related in-country processing commitments). Customers in the EU, EFTA, and UK have Anthropic disabled by default (No users). Commercial cloud elsewhere often has it on by default. AI Administrators (or Global Admins) manage this under Copilot → Settings → AI providers operating as Microsoft subprocessors, and can scope to security groups. Copilot Studio / Power Platform may need an extra allow-external-LLM step in the Power Platform admin centre after the M365 setting is on.

Jack Schofield’s First Law of Computing (historically attributed) asks whether you can get data out on terms you understand. Here the twin questions are: where does processing happen, and whose DPA covers the bits Anthropic sees?

Practical options:

1. Leave Anthropic off for the tenant (UK/EU default). Safest Steady default while Legal maps EUDB expectations to the Learn caveat. Users keep the Microsoft-hosted models already in scope for your boundary story.
2. Opt in for a named pilot security group only. Enable Anthropic as subprocessor, assign a small Entra group (Researcher power users or Studio makers), document the EUDB exclusion in the AI register, and review after 30 days. Best balance for most regulated UK orgs that still want a controlled trial.
3. Enable for all users because “it’s under the Microsoft DPA.” Contractually tidier than a direct Anthropic deal — still doesn’t erase the Learn statement that these models sit outside the EU Data Boundary today. Only do this with an explicit Legal acceptance note.
4. Never enable “Anthropic models with Data Retention” without a separate paper trail. Different terms, Anthropic retention of prompts/responses, off by default everywhere. Treat as a second decision, not a side-effect of turning Claude on.

The caution: UI labels change (Fable-class / Mythos-class / “with Data Retention”). Features can disappear when you disable the subprocessor. GCC High / DoD and some sovereign clouds don’t get Anthropic at all; non-federal GCC has its own dated opt-in with FedRAMP boundary warnings. And “Microsoft subprocessor” is not the same claim as “processed inside the EU Data Boundary.”

I’d suggest recording Claude as a Wave 0 GRC item: default off (or pilot group only), EUDB exclusion acknowledged, Data Retention models explicitly forbidden unless Legal signs a separate risk acceptance, and Power Platform external-LLM locked to the same scope. Steady before scale means model choice follows the boundary, not the demo.

You’re asking the right question. In the end it comes down to whether “optional Claude” is worth a residency exception you can defend.