Claude in Microsoft Copilot Goes On by Default in the EU on 9 December

On 9 December 2026, Microsoft will switch on Anthropic's Claude models in Copilot for every EU and EFTA tenant that has not made a choice. That is the core of Message Center notice MC1546438, published on 10 November. If your Microsoft 365 admin has never opened the Anthropic setting, the answer your company gives on 9 December will be "yes, for all users", whether or not anyone meant it.

The change itself is reasonable. Until now Claude in Copilot sat outside Microsoft's EU Data Boundary, which is why it was off by default in Europe. Microsoft announced on 5 November that processing for these models moves inside the boundary by the end of 2026, and the notice now puts that in "early December". Once the data question is answered, Microsoft treats Claude like any other model in Copilot and turns it on.

The part that deserves attention is the default. A setting nobody has touched becomes a decision about a new AI provider, made by a calendar.

What exactly happens on 9 December

The notice is precise, so it is worth reading the four cases as Microsoft states them:

• No selection made. Claude models are enabled for all users on 9 December.
• Claude already enabled for all users or for specific groups. Nothing changes in who has access. Processing moves into the EU Data Boundary automatically once it is available.
• Claude allowed "only in Copilot in Microsoft 365 apps". This option is retired. On 9 December the tenant moves to All users.
• Claude disabled ("No users"). The selection is respected.

The setting lives in the Microsoft 365 admin center under Copilot → Settings → View all → AI providers operating as Microsoft subprocessors → Anthropic. Changing it requires the AI Administrator or Global Administrator role, according to the same notice. After 9 December admins can still switch Claude off at any time, so nothing here is irreversible. But "we can turn it off later" is a weaker position than "we decided before it went on".

The notice lists Microsoft Copilot, Microsoft 365 Copilot Chat and Copilot in Power Platform among the affected services, and says users will get Claude "across Microsoft Copilot experiences and in Microsoft Copilot Studio". My reading is that this is not only a question for companies that pay for Microsoft 365 Copilot licences. If your staff use the included Copilot Chat, the setting applies to you too.

The detail most teams will miss

The retired option is the trap. A cautious IT team that chose "only in Copilot in Microsoft 365 apps" did the sensible thing: allow Claude in Word or Excel, keep it out of agent building. On 9 December that careful choice becomes the widest possible one.

A hypothetical example. A distribution company with 150 employees tested Claude in October and limited it to Microsoft 365 apps. The sales team is building a Copilot Studio agent over the price list and customer files. Nobody touches the admin center in December. On 9 December Claude becomes available to everyone, including the people building that agent, and the first time anyone hears about it is when a model picker shows a new option. No rule was broken. Access simply widened without a decision.

Data Boundary does not mean every Claude model

Microsoft's admin documentation on Anthropic as a subprocessor separates two groups. Standard Claude models in Copilot fall under the Microsoft Product Terms and the Data Protection Addendum, and Microsoft's announcement adds that customer data is not used to train generative AI foundation models. A second group, labelled "Anthropic models with Data Retention", is different: data is stored by Anthropic and is not subject to your Microsoft customer agreement, including the Product Terms and DPA. These models have their own control and stay off by default everywhere.

The notice is explicit that the Data Retention models are "governed by a separate setting and aren't part of this change". DataNorth's analysis makes the practical point: do not assume they move into the boundary with the rest. If someone in your company asks for the most advanced Claude model by name, check which group it is in before you say yes.

There is also a short gap in timing. Processing moves inside the boundary in "early December", without a precise day in the notice. If you enable Claude today, your prompts are processed outside the EU Data Boundary until that move happens. For most internal drafting that may be acceptable. For HR files, health data or client documents under NDA, waiting a few weeks costs very little.

What a deliberate answer looks like

There are three honest answers, and any of them is fine if someone chose it.

Yes, for everyone. Reasonable if you already treat Copilot as a general tool, your data protection officer is comfortable with Anthropic as a Microsoft subprocessor and people have been asking for Claude. Then do nothing in the setting, but tell your helpdesk, because users will see new options and ask.

No, for now. Reasonable if you are mid-way through a DPIA, have no use case yet or are in a regulated sector that needs the documentation updated first. Select "No users" before 9 December and put a date in the calendar to look again.

A named group. In my view this is the right answer for most companies with 30 to a few thousand employees. Create a Microsoft Entra ID group for the teams that have a reason to use Claude, such as analysts working with long documents, and allow only them. You learn from real use without widening access to everyone at once.

Whichever answer you pick, write down three things: who decided, why, and when you will review it. DataNorth also recommends adding Anthropic as a subprocessor to your DPIA and processing register. That is a few lines of text, and it is exactly what an auditor or a client's procurement team will ask for later. DataNorth adds that Copilot Studio and Power Platform have an additional setting in the Power Platform admin center, so check both places rather than assuming one switch covers everything.

Whose decision this is

It is tempting to forward MC1546438 to the Microsoft 365 admin and move on. The admin can click the setting. The admin should not have to decide alone whether a new AI provider processes company data, which teams get it and what goes into the register.

This is the small, recurring kind of decision that shows whether a company actually has an owner for AI. Vendors will keep shipping defaults like this one: a new model, a new region, a new agent feature, each with a date. If nobody is accountable, the defaults make the policy. If you would like a structured way to assign that ownership and keep a list of AI systems and providers current, our AI governance work starts exactly there.

For 9 December the task is smaller. Open the setting, check which of the four cases you are in, and make sure the answer on that day is one somebody actually chose.