The Microsoft 365 Copilot admin world doesn't get shaken up every week, but September 24, 2025 was one of those days for anyone following AI integration closely. Anthropic models are now part of Copilot. Since this week, Claude Sonnet 4 and Claude Opus 4.1 have been available to Frontier program customers. You can select them in the Researcher agent, and you can use them to build custom agents in Copilot Studio. The topic has already come up in tenant governance calls and works council meetings I've been part of, and people ask the same things every time. What is actually new? Where do users' prompts end up? And what has to be enabled (and documented) before anyone gets access to models that don't come from Microsoft?

Claude Sonnet 4 and Opus 4.1 send Copilot prompts outside Microsoft's cloud

This is the first time you can pick Anthropic's Claude models next to Microsoft's own GPT-based models, both for research in Copilot's Researcher agent and for your own agents in Copilot Studio. The important detail is where processing happens. With Anthropic models in Copilot, prompts and completions are handled outside Microsoft's cloud, meaning outside Microsoft-managed environments and outside the EU Data Boundary. Prompt traffic goes to Anthropic's service and falls under Anthropic's terms rather than Microsoft's. The Microsoft 365 Copilot privacy documentation leaves no doubt here: once you enable Anthropic, responsibility for data handling sits with you as the tenant admin. So before Claude goes live for your users, a proper data protection review and an update of your documentation are mandatory, not optional. That applies even more in regulated or privacy-sensitive environments.

Admins switch on Anthropic models with one toggle in the admin center

Technically, turning on Anthropic models in Copilot is simple, although Microsoft attaches a very visible warning. In the Microsoft 365 admin center, go to Copilot settings → AI providers. You'll find a new toggle called "Enable Anthropic Claude models". It is off by default, and the models only show up in Researcher and Copilot Studio after you give explicit organizational consent. That friction is intentional. Microsoft wants admins to decide consciously instead of letting new AI slip into a pilot without oversight.

What has worked for me is writing down exactly which use cases will benefit and how you plan to monitor data flowing to and from Claude. The smoothest rollouts I've seen were the ones where legal, IT and the works council saw the changes early. The Copilot admin center now includes a data flow diagram for external AI providers. Here is a simplified version that shows the model boundaries clearly:

flowchart TD
  User["User prompt"] --> Copilot["Microsoft 365 Copilot"]
  Copilot --> Claude["Anthropic Claude model"]
  Claude --> Copilot
  Copilot --> M365Data["Microsoft 365 data"]

Figure: Data flow with Anthropic models enabled in Copilot. Prompts are processed outside Microsoft's cloud.

The picture is hard to misread. With Anthropic enabled, your users' queries and their context leave the usual Microsoft 365 perimeter, and every data protection impact assessment should cover that.

Data protection and works council paperwork has to be done first

For admins, this means the (often thankless) job of updating data protection documentation and works council briefing material for Anthropic models in Copilot. Most Copilot features stay within Microsoft's own privacy standards (I covered that in my post on Copilot data residency). Once Anthropic is involved, though, Anthropic's terms and hosting locations apply. Your documentation should at least state which scenarios send data to Claude, which types of data are in scope, and which teams or roles will use the models. Works councils pay close attention to these distinctions in my experience, particularly when it comes to auditability and oversight. When I prepare a rollout, I always point out the admin center toggle, the external data flow and the available audit logs, so every stakeholder can see who is trying out Anthropic and when.

Pro tip: If you only want to pilot Claude models in selected Copilot Studio agents, restrict access to a small group and check your compliance documents before rollout. Copilot Studio's agent permissions let you control exposure. Keep access narrow until your privacy and governance teams have signed off.

If you want to dig into the legal and technical side, Microsoft's AI subprocessor list is required reading.

Copilot Studio agents can now run on Claude models

Once Anthropic is enabled for the organization, the models can be selected in Copilot Studio. Admins and makers get new options, and with them more compliance and documentation work. In real tenants I regularly see power users who want to try Claude Opus 4.1 right away while the compliance team holds them back. Copilot Studio lets you choose the model per agent, and usage shows up in the audit logs. If you're already introducing the new GPT-5 models in Copilot, it's worth comparing output, language handling and above all data residency. Where scenarios must stay on Microsoft-hosted models, say so in the agent name, the description and your policy documents. The technical separation works well in practice, and agent permissions plus careful rollout planning do most of the heavy lifting.

I see real innovation here, but only with eyes wide open

Anthropic models in Copilot, starting with Claude Sonnet 4 and Opus 4.1 in the Frontier program, give you genuinely more flexibility. They also bring risk. If you plan to enable them, do the data protection impact assessment, review your documentation and talk to your privacy and works council teams up front. The new models extend what Copilot can do in research and custom agents, but once your data leaves Microsoft's walls, you carry more of the responsibility.

My advice is to move quickly, but only after the checklist is done, and to bring compliance and privacy stakeholders in before the first pilot starts. Handled that way, Anthropic models in Copilot can become a real asset instead of a compliance headache.