---
title: "Claude in Copilot enters the EU Data Boundary: what it covers and what it leaves open"
url: https://zerofive.ai/en/blog/compliance/claude-in-copilot-eu-data-boundary
canonical: https://zerofive.ai/en/blog/compliance/claude-in-copilot-eu-data-boundary
language: en
published: 2026-11-10
updated: 2026-11-14
author: "ZeroFive.AI"
tags: EU Data Boundary, Microsoft Copilot, Claude, flex routing, GDPR, AI Rating
abstract: "Claude in Copilot enters the EU Data Boundary by end of 2026: what it secures, who it suits and what it leaves out, from the CLOUD Act to flex routing."
---

# Claude in Copilot enters the EU Data Boundary: what it covers and what it leaves open

**In short.** By the end of 2026 Microsoft will move the processing of Claude models inside the EU Data Boundary for organisations in the EU and EFTA, Copilot Cowork included. The objection about where processing happens, which for many DPOs was the reason for a negative opinion, falls away, and Claude in Copilot becomes an acceptable destination for internal data and ordinary personal data. What does not change is the US jurisdiction of the providers, flex routing at peak load, continuity of access to the model, the use of Claude outside Copilot and, above all, what an agent like Cowork can read and do inside the tenant.

On 4 November Microsoft updated the Microsoft Learn page on Anthropic models in its services, and the highlighted sentence was the same as in January: Claude, in Copilot, Researcher, Copilot Studio, Power Platform and the Microsoft 365 apps, is "currently excluded" from the EU Data Boundary and, where applicable, from in-country processing commitments. The next day Charles Lamanna announced on the Microsoft Copilot blog that by the end of 2026 that sentence will stop being true, and for the DPOs who have kept the switch off in the admin center since January it is the news they had been waiting for.

The announcement deserves to be taken seriously for what it actually moves, that is, the place where the model processes the data. Companies that were asking only for this can restart, those that had other questions will find them all still open, and a new question joins them, because with Copilot Cowork the model works inside mailboxes and document repositories with the permissions of whoever uses it.

## The 5 November announcement and the 4 November page

The commitments are few and precise. For organisations in the EU and EFTA, Claude models will be usable across all Copilot experiences, Cowork included, with processing inside the EU Data Boundary, where Microsoft 365 data is already stored. In these regions the setting is currently off by default and administrators can switch it on for all users or for specific groups from the Microsoft 365 admin center, while those who have already switched it on need to do nothing, because processing will move on its own once the service is ready. Claude in Copilot is covered by Microsoft's enterprise data protection, and customer data is not used to train foundation models.

The text has no precise date, no list of the experiences available on day one and no reference to the United Kingdom or to flex routing, and Microsoft says it will share more details as availability gets closer. Until then the technical and contractual documentation is the 4 November version, so a company that switches Claude back on today on the strength of the announcement sends its data outside the European perimeter for every week left before the release. Anyone with a negative opinion on file should revisit it when the Microsoft Learn page changes, not when the blog post came out.

The timeline arrived a few days later, in message MC1546438 in the Microsoft 365 admin center. The move into the EU Data Boundary is planned for early December, and from 9 December Claude is switched on for all users in EU and EFTA tenants where no administrator has made a selection in the Anthropic setting. Anyone who wants to decide for themselves has to do it before then, choosing no users or specific groups, otherwise the default decides.

## A perimeter for green and yellow data

In our piece on [protecting data in the use of AI](/en/blog/compliance/protecting-data-ai-choices-before-architecture) we described the classes we use in our projects, from the green of public data, or data meant to become public, to the yellow of confidential internal information and ordinary personal data, up to the red of what never leaves the perimeter the company has chosen. Each class has its permitted destinations, and Microsoft's announcement changes where Claude sits on that map: for a company that accepts processing in any European region, Claude in Copilot becomes a legitimate destination for green and yellow data, on the same terms as the GPT models Copilot already uses.

Until now anyone working on Microsoft 365 had to choose between the model best suited to a given task and the residency commitments on which their impact assessment was built, and many chose residency, leaving Claude to pilot groups or to public data. With the announcement the choice of model goes back to being a choice about quality, made case by case within a contract the legal team already knows.

It is enough for companies that already have Copilot, have classified their data and had issued a negative opinion only because of where processing took place. Think of a research office using Researcher on public reports and internal documents, or a sales team preparing offers with Cowork from its own material, without touching records of identified customers. For them the decision can be taken as soon as the documentation is updated.

## The CLOUD Act does not care about data centres

The EU Data Boundary is a commitment about location and sets where data is stored and processed. The companies processing it belong to US groups, Microsoft as well as Anthropic, so the 2018 CLOUD Act, which allows US authorities to request data held by American companies even when it is stored abroad, applies to Claude in Copilot inside the European perimeter exactly as it applied outside. For those who have justified in writing the choice of a national perimeter or a European provider for their most critical data, the announcement changes nothing, and it should not change their classification.

The perimeter also has more jagged edges than its name suggests. Microsoft documents a set of partial transfers that continue even for covered services, tied to security and support, while the United Kingdom, where many Italian groups have subsidiaries or service centres, sits outside the EU Data Boundary and outside the announcement. A group with tenants in several countries has to check, country by country, which commitment covers each tenant.

## Flex routing and revocable access

Since 17 April 2026, EU and EFTA tenants using Copilot have had a setting called flex routing, which at peak load allows model inference to move outside the EU Data Boundary, to the United States, Canada or Australia, with a limited amount of pseudonymised data also stored outside the perimeter. According to the documentation updated in November it is on by default for tenants created after 25 March 2026, while administrators of older tenants are asked to check its status. It can be switched off in the admin center, in the Copilot settings, with the AI Administrator role, and anyone who has never opened it cannot be sure which perimeter they are actually using. The Claude announcement does not say whether flex routing will apply to its models, and until Microsoft clarifies it the safer assumption is that it will.

There is also a second switch, one no administrator controls. On 12 June Anthropic suspended access to Claude Fable 5 and Claude Mythos 5 to comply with export controls from the US Department of Commerce, and restored it on 1 July, after those controls had been lifted. For three weeks anyone who had built processes on those models went without them, regardless of where the data was processed. Residency describes where processing happens, while continuity depends on the decisions of the provider and its government, and a process resting on a single model, inside a single catalogue, needs a plan for the day that model does not answer.

## Inside the tenant, the agent

The impact assessments written for Copilot in 2024 and 2025 reasoned about an assistant that answers questions, and the main concern was where the text of the request ended up. Cowork works differently: it takes on a task, moves through mailboxes and repositories with the user's permissions, produces documents and can act on the user's behalf. Risk therefore shifts from where processing happens to what the agent can read and do, and on this the EU Data Boundary says nothing.

The most frequent case concerns permissions. Many companies have SharePoint sites and shared folders with wider permissions than needed, inherited from years of fast collaboration, and as long as people were the ones searching, that width stayed hidden behind the effort of finding things. An agent makes no effort, and if a file is readable by the user it enters the context of the task, from where it can end up in a shared document or in an email to someone outside. Before extending Claude in Cowork to a group, an audit of that group's permissions is worth more than any commitment about the data centre, and what the agent can do on its own should be decided with the [delegation levels](/en/blog/ai-governance/ai-agents-four-layers-delegation-perimeter) we use for any agent.

The announcement will also have a cultural effect, and inside companies it will arrive simplified as "Claude is fine now". The commitment, however, covers only Claude inside Copilot. Anyone using Claude directly, with an account on claude.ai, with Claude Code or with Anthropic's APIs, stays outside this perimeter, because the direct channel does not currently offer an EU-only processing region and those who want one have to go through the EU regions of the cloud platforms. Without a [policy](/en/blog/ai-governance/ai-policy-2-0-what-the-system-can-do) that explains which Claude is authorised and for which data, the message risks widening the very shadow AI that the switched-off toggle was trying to contain.

## When the data is red

For red data the announcement moves nothing, and it was not meant to. Special categories under Article 9 of the GDPR, judicial data, financial information on identified customers, trade secrets and, for a NIS2 entity, information on the security of its systems have a different destination if the company has decided they never leave a national perimeter or its own infrastructure, with or without the EU Data Boundary. The same goes for environments that must run disconnected from the network and for anyone who needs models outside the Microsoft catalogue, or with update cycles different from those decided in Redmond.

For these cases the routes are different and can be combined: a gateway that routes each request according to the data class, open weight models on infrastructure in Italy or directly on premises for red data, Claude through the European regions of the cloud platforms when that model is needed outside Microsoft 365. In a scheme like this, Copilot with Claude inside the European perimeter becomes one of the destinations, the natural one for everyday work on green and yellow data within the productivity suite.

![Three perimeters: the Microsoft 365 tenant, the EU Data Boundary and what remains outside it](/claude-copilot-eudb-perimetri-en.png)

The choices to make before switching the toggle back on can be checked with a few questions for your CAIO, or for whoever plays that role in the company:

- Will the opinion on the processing be updated on the basis of the Microsoft Learn documentation or of the announcement?
- Is flex routing on in the tenant, and did someone decide to leave it that way?
- Has anyone made a selection in the Anthropic setting before 9 December?
- Which groups will use Claude, and which data classes do they work with?
- Have the permissions on the repositories those groups can reach been checked before giving them an agent?
- Does the record of processing list Anthropic as a sub-processor of Microsoft, with its purpose?
- Which rule applies to Claude used outside Copilot?
- What happens to processes built on Claude if the model is unavailable for three weeks?

When the answers already exist, Microsoft's announcement is the green light many were waiting for, and Claude can enter everyday work as soon as the documentation is updated. When the answer to any of these questions is the name of a licence, the toggle can wait a few more weeks, the time needed to write the rules that the EU Data Boundary does not write on the company's behalf. These are the questions we start from in the Risk dimension of the AI Rating, and if you want to begin there you can [book a call](https://calendly.com/fabiolalli/zerofive) or write to us at hello@zerofive.ai.
