Claude Is Now Inside Microsoft 365 Copilot. For UK Tenants, It Is Off Until You Say So.

Anthropic’s Claude models now sit alongside OpenAI’s inside Microsoft 365 Copilot, Copilot Studio and Power Platform. For most commercial customers they are switched on by default. For tenants in the UK, EU and EFTA they are not: the setting starts at Off and stays there until an administrator turns it on.

That default is a sensible one, because enabling Claude changes where your prompts may be processed and whose contract governs them. This guide explains what Microsoft’s own documentation says, why UK data residency is the sticking point, and how to make a documented decision rather than letting the setting drift. If you are weighing Microsoft 365 AI automation or AI governance for SMEs, this is a ten-minute decision that deserves to be made on purpose.

What has actually changed

Microsoft’s documentation, Anthropic models in Microsoft 365 Copilot, sets out the position. Anthropic has been onboarded as a Microsoft subprocessor, which means that for standard models the Microsoft Product Terms, the Data Protection Addendum and Enterprise Data Protection apply, not a separate Anthropic contract. The models appear in Microsoft Copilot, Researcher, Copilot Studio, Power Platform and Copilot in Microsoft 365 apps. In Copilot Studio, builders pick the model when they create an agent, and users can choose Claude in features such as Researcher.

The regional split is the important part:

Two further details matter. First, if you opted in earlier under Anthropic’s own commercial terms, Microsoft says you must opt in again under the new arrangement. Second, since 3 April 2026 there has been a separate admin setting, Copilot in Microsoft 365 apps with Anthropic models, which can make Anthropic the default model in Copilot in the Microsoft 365 apps for UK, EU and EFTA tenants. Enabling the subprocessor and choosing the default are separate decisions.

The residency catch

This is where UK businesses need to slow down. Microsoft states that Anthropic models in Copilot, Researcher, Copilot Studio, Power Platform and Copilot in Microsoft 365 apps are currently excluded from the EU Data Boundary and, where applicable, from in-country processing commitments. An independent UK consultancy write-up reaches the same conclusion and recommends checking Microsoft’s subprocessor documentation on the day you decide, because boundary coverage may change.

Separately, Microsoft has been rolling out in-country processing for Copilot interactions. Its revised plan, last updated on 3 April 2026, lists the United Kingdom among the countries due to receive it by the end of 2026. Microsoft describes the offer as aimed mainly at government and highly regulated organisations. If your contracts, your regulator or your customers rely on that commitment, then enabling Claude for a user means that user’s Claude-routed prompts sit outside it.

In practice this is a scoping decision, not a yes-or-no one. Enable Claude for a named group doing work where the benefit is clear and the data is low risk, and leave it off for teams whose obligations depend on the boundary. The admin centre lets you assign access to specific users or Microsoft Entra ID groups.

The second trap: models with data retention

Microsoft draws a hard line between standard Anthropic models and those labelled Anthropic models with Data Retention. Its page lists some of the newest Claude models in this second category. They are off by default for every tenant, including regions where other Anthropic models are on, and nobody has access until an administrator explicitly opts in and chooses who can use them.

The terms differ materially. For these models Anthropic acts as an independent processor under its own Commercial Terms and Data Protection Addendum, not as a Microsoft subprocessor. Microsoft says Anthropic stores most inputs and outputs for up to 30 days, content flagged by its trust and safety classifiers may be kept for up to two years, and classification scores for up to seven years. Anthropic does not use retained data for training without express permission. Microsoft also says some models in the newest class may be offered under the standard subprocessor terms for certain organisations, so check which label a specific model carries before you enable anything.

Opting in to standard Anthropic models does not switch these on. Keep the two decisions separate in your records.

Where Claude helps, and where you may not need it

Model choice is rarely the reason a Copilot rollout succeeds or fails. The usual causes are permissions, data quality and adoption. Our oversharing guide covers the first, and it matters more than which model answers the prompt. Claude inside Copilot is worth testing where you already have a reason to prefer it, such as longer document analysis in Researcher, or where a Copilot Studio agent benefits from a different model. It is not a reason to rush.

It also does not replace a bespoke build. If you need an audit trail, your own approval steps or a fixed processing region, a purpose-built tool on the Claude API, with the region you choose, may fit better. See Copilot versus custom AI and our Claude API buyer’s guide for where that line sits.

A decision checklist for UK administrators

This is the order I would work in. For a small tenant it is an afternoon, not a project.

1. Check the current state. In the Microsoft 365 admin centre go to Copilot, then Settings, then View all, then AI providers operating as Microsoft subprocessors. Confirm Anthropic shows as Off, and check whether anyone opted in earlier under the old terms.
2. List your residency obligations. Look at customer contracts, regulator expectations and any internal commitment that depends on EU Data Boundary or UK in-country processing. If there are none, the decision is easier.
3. Pick a pilot group and a use case. Choose a small Entra ID security group and one task where you want to compare models. Avoid teams that handle special category or regulated data in the first round.
4. Assess the data protection impact. If enabling Claude materially changes how personal data is processed, complete a data protection impact assessment, update your record of processing activities and your privacy notice, and check how any transfers outside the UK are covered. Microsoft’s subprocessor documentation lists the current position.
5. Decide on the data-retention models separately. Unless you have a specific need and have read Anthropic’s terms, leave them off.
6. Enable, tell people and test. Select Anthropic under available subprocessors, assign the group, and save. Users see a Claude indicator when it is in use. Tell them which tasks are in scope and compare results against the current model.
7. Record the decision and review it. Write down who approved it, for which group and why, and set a review date. Microsoft’s scope and in-country timeline are still moving.

If you later change your mind, the same page has an option to disable Anthropic as a Microsoft subprocessor. Microsoft notes some features may become unavailable, and you can re-enable it afterwards.

Update your AI policy to match

An approved-tools list that says “Copilot” no longer tells staff enough. Your AI usage policy should now say which models are enabled, for whom, and which data types are out of bounds for them. Our AI usage policy how-to and the free policy template give you the structure, and the AI governance framework shows where a model-level decision like this one belongs in your register of approved tools.

FAQ

Sources: Microsoft Learn, Anthropic models in Microsoft Online Services; Microsoft Copilot blog, in-country data processing for Microsoft 365 Copilot (original November 2025, revised 3 April 2026); Talk Think Do, Claude in Microsoft 365 Copilot: UK checks (August 2026). Microsoft changes these settings often, so check the current documentation before you act.