Don’t Turn On Anthropic in Microsoft Copilot. Here’s Why

Important update! Microsoft have now updated their approach, so this is outdated information.  You can read the update here:

A welcome update on Microsoft Copilot and Anthropic

I’ve left this blog post here in case anyone wants to understand the history of this and the original issue.

---

An unusually blunt blog title for me, but you’ll see why!

Before I go on, I want to stress that this is about something that your admin has to enable, so as a user, you don’t need to take any action at the moment, but you might be interested in reading the details and what it means.

Microsoft has just enabled the use of Anthropic’s Claude models inside Copilot 365. This is specifically in the Researcher agent and Copilot Studio, and has to be enabled by your system administrators. On the surface, this looks like a welcome expansion of choice. But enabling it comes with a major caveat: when you switch Anthropic on, your data leaves Microsoft’s managed environments. That means many of the protections and guarantees you’ve come to expect from Microsoft’s AI tools no longer apply.

Microsoft’s own documentation makes this clear:

Important
When your organization chooses to use an Anthropic model, your organization is choosing to share your data with Anthropic to power the features. This data is processed outside all Microsoft‑managed environments and audit controls, therefore Microsoft’s customer agreements, including the Product Terms and Data Processing Addendum do not apply. In addition, Microsoft’s data‑residency commitments, audit and compliance requirements, service level agreements, and Customer Copyright Commitment do not apply to your use of Anthropic services. Instead, use of Anthropic’s services is governed by Anthropic’s Commercial Terms of Service and Anthropic’s Data Processing Addendum.

Why Microsoft is doing this

This is part of Microsoft’s move to diversify beyond OpenAI. Anthropic’s Claude models are now available for Copilot’s Researcher tool and Copilot Studio, once enabled by your admin team.

Early reports suggest Claude performs well in tasks like Excel automations and PowerPoint generation, in some cases outpacing OpenAI. Microsoft is also making model choice available at no additional cost, presenting this as “more capability, same price.” And this isn’t about replacing OpenAI entirely; Microsoft is building a catalogue of models for Copilot.

All of this sounds reasonable until you look at the details.

What changes when you enable Anthropic

When Anthropic’s models are used in Copilot, you leave the security perimeter that Microsoft has worked hard to build for Copilot. Here’s what you lose:

• Data governance and control: data is no longer processed in Microsoft’s environment.
• Audit and compliance protections: requirements many organisations rely on simply don’t apply.
• Data residency guarantees: you cannot assume your data stays where you expect.
• Customer Copyright Commitment: Microsoft’s legal protection against copyright claims does not cover Anthropic use.
• Service levels: Microsoft’s SLAs don’t apply.

Instead, you’re bound by Anthropic’s own terms and guarantees. At the moment we’ve not fully analysed these, so this isn’t a comment on whether there are specific risks in them, but just losing the elements mentioned above is significant enough.

Why this undermines trust

So far, Microsoft has built confidence in Copilot by reassuring customers: your data is safe, processed within trusted boundaries, and covered by Microsoft’s commitments. This has been incredibly useful for us, helping with our messaging, building trust in AI, and almost all progress has been in the right direction.

By introducing an option that discards those guarantees, Microsoft has created confusion. You can no longer simply say, “Copilot is safe.” Instead, the message becomes conditional: it is safe, as long as you don’t turn on Anthropic. That nuance is easily lost, and once it’s lost, so is trust.

Conclusion

This is a deeply concerning move from Microsoft. Their efforts to build trust in Copilot and Microsoft’s AI products, among users, IT, and governance professionals, have been strong. But this decision erodes that.

Copilot 365 can still be safe and secure, as long as you do not enable the Anthropic option. But adding an optional feature that weakens guarantees muddies the message. Until Microsoft brings Anthropic into its trusted perimeter, the safest advice is simple:

Don’t turn it on.

We’ll watch this closely – I expect some changes to this arrangement as feedback on social media etc, is equally strong, and Microsoft has tended to listen and act.

In the meantime, we are reaching out to Microsoft, and I’ll share any more news on this as we get it.

---

Find out more by visiting our Artificial Intelligence page to explore publications and resources, learn more about our communities and sign up for our AI Literacy training.

For regular updates from the team sign up to our mailing list. 

Get in touch with the team directly at AI@jisc.ac.uk