Claude inside Microsoft 365 Copilot: the UK tenant setting most admins have not checked

Microsoft has made Anthropic's Claude models available inside Microsoft Copilot (formerly Microsoft 365 Copilot), and has been explicit that when they are used, processing happens outside the Microsoft EU Data Boundary. For UK organisations this is an administrator decision rather than a background change, and it is one that touches your data protection impact assessment and your record of processing activities.

There is a complication that most coverage has missed. There are two different settings, they behave in opposite ways, and they interact. A UK administrator who reads the headline coverage and concludes their tenant is safe by default may be wrong, depending on which setting is in play and when the tenant was created.

This article is written for the person who has to make and defend that decision: the IT director, the data protection lead, or the managing partner who signs off supplier changes. It is not a walkthrough of where the toggle lives, which Microsoft documents perfectly well. It is about what you are consenting to when you enable it.

The two settings, and why they conflict

The first is the global subprocessor setting, covering AI providers operating as Microsoft subprocessors. Microsoft's position on this one is unambiguous for the UK: Anthropic models in Microsoft 365 Copilot, Researcher, Copilot Studio, Power Platform and Copilot in Microsoft 365 apps "are currently excluded from the EU Data Boundary, and when applicable, in-country processing commitments. Customers within the EU Data Boundary and customers in the UK have Anthropic models disabled by default." For these regions the setting appears but defaults to no users.

The second is a narrower, apps-level setting covering Copilot in Microsoft 365 apps with Anthropic models. This one behaves the opposite way: Microsoft states it is on by default for tenants in the EU, EFTA and the UK created after 25 March 2026. Tenants that existed before that date are directed to the Message Centre.

So the correct answer to "are we opted in by default?" is: it depends which setting you mean and when your tenant was created. Stated without that qualification, either version is wrong.

The two also interact in a way that can trap an administrator mid-configuration. Microsoft notes that when the global subprocessor setting is scoped to all users, or to specific users and groups, the apps-level Anthropic setting becomes unavailable and cannot be changed. Configure in the wrong order and a control you expected to have disappears.

There is one more item for organisations that believe they have already handled this. UK, EU and EFTA organisations that previously opted in under Anthropic's own separate commercial terms must opt in again, because the newer toggle defaults to off. The Anthropic independent processor setting was decommissioned on 1 May 2026, and without the subprocessor setting enabled, access to Anthropic models is no longer available. A prior decision does not carry forward.

The UK is not inside the EU Data Boundary

This is the point that most UK commentary gets wrong, usually by treating "UK and EU" as a single bucket.

Microsoft’s 3 April 2026 update says local inferencing for Microsoft 365 Copilot interactions in the UK is expected by the end of 2026. That is an announced timetable, not evidence that a particular UK tenant already has it. Microsoft’s Anthropic subprocessor documentation checked on 14 October 2026 excludes Anthropic models from in-country commitments where they apply. Check the product surface, model and tenant’s current routing in writing.

That commitment is precisely what enabling Anthropic models sets aside. Microsoft states that when Anthropic models are used in Copilot experiences in Word, Excel or PowerPoint, data processing occurs outside the EU Data Boundary. Its main Copilot privacy page repeats the position plainly: models provided by Anthropic as a subprocessor are currently excluded from the EU Data Boundary.

Enabling an Anthropic model therefore requires a fresh processing and transfer assessment. Do not describe it as withdrawing a UK inference entitlement that has not been verified for the tenant.

Two different contractual positions, and one that catches people out

Where Anthropic operates as a Microsoft subprocessor, your contractual position stays inside the Microsoft estate. The Microsoft Product Terms, the Microsoft Data Protection Addendum, Enterprise Data Protection and the Microsoft Customer Copyright Commitment all apply. Your counterparty remains Microsoft.

Microsoft names Claude Fable 5 and Claude Mythos 5 as models with data retention. For these models, Anthropic acts as an independent processor under its own commercial terms and Data Processing Addendum, and stores data outside the commitments in your Microsoft Customer Agreement, Product Terms and DPA. Microsoft's documentation, read on 14 October 2026, separately says eligible organisations can use Fable 5.1 with Anthropic as a subprocessor and no Anthropic retention of customer content. Confirm which route applies to your model and tenant before enabling it.

Models labelled with data retention require a separate opt-in and acceptance of Anthropic’s terms. They are off by default, including in tenants where other Anthropic models are enabled. Assess the need and contractual route before enabling them.

Microsoft publishes retention figures for that category which are directly usable in a DPIA: Anthropic stores most inputs and outputs for up to 30 days, may retain content flagged by trust and safety classifiers for up to two years and classification scores for up to seven years, and does not use retained data for training without express permission. Cite those to Microsoft. Anthropic's Covered Models article gives the 30-day baseline. Anthropic’s retention page states that flagged inputs and outputs may be retained for up to two years, and trust and safety classification scores for up to seven years.

Does enabling this create a restricted transfer?

This is where the two routes diverge, and where a UK organisation should take the question seriously rather than assuming the answer.

The ICO published updated international transfers guidance in January 2026 introducing a three-step test: does UK GDPR apply to the processing; are you initiating the transfer to an organisation outside the UK; and is the recipient a separate legal entity. Every restricted transfer needs an applicable UK adequacy route, appropriate safeguards or an exception, and where a safeguard is relied on, a completed transfer risk assessment is required.

The second step is the one that matters here, because it turns on who initiated the transfer. Under the subprocessor route, Microsoft appointed Anthropic; your contractual relationship remains with Microsoft, and the transfer analysis follows that structure. Under the models-with-data-retention route, your tenant is directly appointing an overseas independent processor under Anthropic's own DPA, which looks materially more like a controller-initiated restricted transfer requiring an IDTA or Addendum plus a transfer risk assessment.

Legal commentary on the ICO update suggests the practical rule of thumb is to follow the contractual relationships and identify the party that designed the transfer structure or first chose the recipient, and notes this analysis is particularly important in cloud, SaaS, managed service, processor and sub-processor arrangements. The same commentary notes that the Data (Use and Access) Act 2025 has introduced material UK and EU divergence, so a single transfer analysis may no longer satisfy both regimes.

We would not tell you the answer for your organisation in an article. We would tell you that the two routes are not equivalent and should not be assessed together, and that if your DPIA treats them as one thing it is probably wrong.

What this means for your DPIA and ROPA

Three practical consequences follow, and none of them is satisfied by flipping a switch.

Your record of processing activities needs to reflect Anthropic processing outside the EU Data Boundary, because a new processor in a new location is exactly what a ROPA exists to record. Your DPIA needs revisiting rather than reusing, since the tool it assessed is not the tool you are now running. And your scope is wider than Copilot chat: a single subprocessor decision affects Microsoft 365 Copilot, Researcher, Copilot Studio, Power Platform and Copilot in Microsoft 365 apps, with Copilot Studio and Power Platform requiring a second set of controls in the Power Platform admin centre.

The middle option is the one most firms should look at first. Access can be scoped to specific users or Microsoft Entra ID security groups rather than enabled tenant-wide, applied at provider level and enforced across Copilot and Copilot Studio. For a firm with mixed data sensitivity, where a marketing team's use case and a client-confidential matter are not the same risk, that is usually the proportionate answer.

If you have not yet produced a DPIA for your assistant rollout at all, our guide to whether you need a DPIA before rolling out Claude or Copilot sets out the screening test.

What to do, and what not to do

• Check both settings, not one. Confirm the global subprocessor setting and the apps-level setting separately, and note your tenant creation date, because the defaults differ.
• Re-check if you opted in previously. The independent processor setting was decommissioned on 1 May 2026 and prior opt-ins do not carry forward.
• Decide scope before you decide on or off. Specific users and groups is a real option and is usually the proportionate one.
• Leave preview models with data retention alone unless someone can explain why the change of contracting party is acceptable.
• Update the ROPA and revisit the DPIA as part of the change, not afterwards.
• Do not assume the UK is covered by EU Data Boundary commentary. It is not inside the boundary and its protection is a separate commitment.
• Do not treat the two routes as one for transfer purposes. Who initiated the transfer differs, and so does the analysis.

Where The AI Consultancy fits

Working out which settings are live in your tenant, what the change does to your existing documentation, and whether scoped access is the proportionate answer is the kind of review our Claude implementation engagements open with. Where the question is broader than one toggle, the AI readiness assessment covers the governance position as a whole. Our guide to whether Claude is GDPR compliant covers Anthropic's own retention and training position, and Claude data residency for UK organisations covers what can and cannot be kept in region on the direct routes.

Verified on 31 July 2026 against Microsoft Learn documentation for the AI subprocessor setting and the Anthropic apps setting, Microsoft's Copilot privacy and EU Data Boundary pages, Anthropic's data retention documentation and the ICO's international transfers guidance. This admin surface changed in April, May and July 2026 and is expected to keep moving; confirm the current position in your own tenant before acting. This article is general information, not legal advice.