An employee asks Copilot to summarise a file. Before assessing the quality of the answer, their company wants to know where the information will be processed. It was in this area—the location of data processing—that Microsoft made an announcement on 5 November: Anthropic’s Claude models used in Copilot are to join its European data processing boundary by the end of 2026.
As of 6 November, this is a deployment commitment. Enabling it in a console does not mean that this new data location is already in effect. For organisations that have been waiting for this change, the distinction determines the timing of any potential trial.
What is being announced, and for whom
According to Microsoft, organisations in the European Union and the European Free Trade Association will be able to use Claude in Copilot experiences, including Cowork, with processing within the EU Data Boundary. This boundary covers the EU and EFTA. It does not mean there is a data centre in each customer’s country.
For eligible Microsoft 365 customers, this regional commitment is part of the contractual terms, with no additional purchase specific to this arrangement. That does not mean that all Copilot products or subscriptions become free.
The announcement concerns Claude through Copilot. It does not allow us to conclude that the same terms apply to a Claude subscription purchased directly from Anthropic. The model’s name is therefore not enough to identify the terms of the service: you also need to consider which product you are using it through.
Three decisions not to confuse
The first decision is whether to authorise users to access the model. Microsoft says this setting is off by default in the EU and EFTA; the administrator can enable it for everyone or for specific users and groups.
The second concerns where processing actually takes place. For organisations already using the service, the change announced will happen automatically once it is available. Microsoft has promised further details as availability approaches.
The third is an internal company decision: which documents can be entrusted to the tool? This question does not go away because a provider announces a regional change. A file may contain data that its user is not meant to share, including with a service authorised for other tasks.
Take a hypothetical example: a team wants to prepare a summary of sales meetings. It can start with minutes that contain no customer data to assess the quality of the summary. Moving on to contracts or personally identifiable information is a different decision, involving a different level of sensitivity. The first trial does not automatically validate the second.
Location, training and accuracy: three different questions
Microsoft says that customer data in this context is not used to train generative AI foundation models. This commitment answers a question about data use; on its own, it says nothing about the accuracy of a summary.
In its FAQ of 18 July 2024, the CNIL points out that generative systems can produce plausible but inaccurate information. It recommends, in particular, defining permitted uses, training users and checking the results. It also urges organisations to examine their own responsibilities and those of their providers with regard to personal data.
European processing will not correct an invented date, a misquoted amount or a caveat omitted from a summary. Conversely, an accurate result in a trial is not enough to answer questions about privacy. Assessing these two dimensions separately makes the decision clearer.
What would a genuinely useful trial look like?
Rather than asking whether Claude “works well”, a team can define a task and an expected result. For a summary, this might mean identifying the decisions made, who is responsible for them and their deadlines, while flagging anything that remains undetermined. This framework is a proposed method, not a test we have carried out.
The answer can then be compared with the original document. Has the system added a decision that was never made? Has it turned a hypothesis into a commitment? Has it preserved disagreements? An elegant answer that smooths over these nuances may be less useful than a more restrained and faithful text.
The conditions of the trial should also be noted: the product used, the settings available, the type of document and the date. A validation without this context is difficult to reproduce. The choice of model is therefore best guided by the task at hand, rather than by a general preference for a brand.
The next signal to look out for
For an organisation that makes this use conditional on processing within the EU Data Boundary, the decisive signal will be confirmation that the feature is available for its service. The 5 November announcement paves the way for this step; it does not replace it.
In the meantime, it is possible to clarify use cases, which documents are permitted and the criteria for success. This avoids treating the regional launch as a blanket authorisation. The value of the new feature lies in offering an additional choice within a specific framework, not in exempting users from exercising judgement over the data they send and the answers they receive.
Archive photo: a path on the Microsoft campus in Redmond, in July 2006. It does not show a European data centre.
