Microsoftโs recent announcement about introducing Anthropicโs Claude models into Microsoft 365 Copilot raises an interesting question about AI governance that deserves attention.
From 9 December 2026, Microsoft intends to enable Claude by default for eligible EU and EFTA Microsoft 365 tenants where administrators have not explicitly configured their preferences. Microsoft has also announced that processing for the covered Claude models will move into its EU Data Boundary in early December.
On the face of it, this is a positive development. Organisations gain access to additional AI capabilities, Microsoft is addressing European data residency requirements, and administrators retain the ability to decide whether Claude should be available to all users, selected users, or nobody.
So whatโs the concern? For me, it isnโt necessarily about Claude, or whether it is any more or less secure than the AI models already available through Copilot. Itโs about who is making the decision, and whether the organisation has actually considered the implications. Yes itโs a change in the technology, but what about the approval?
When Copilot changes the flight plan
Consider an organisation that has already introduced Microsoft 365 Copilot. It may have undertaken a security assessment, reviewed the data protection implications, considered the relevant suppliers and sub-processors, updated its policies, and approved the use of Copilot within the business. That approval was presumably based on an understanding of what was being introduced, how information would be processed, and what controls were available.
Now Microsoft introduces another AI model provider and changes the default configuration. The organisation may not have requested that change. The people responsible for information security, data protection or AI governance may not even be aware of it.
And yet, unless an explicit configuration decision has already been made, the additional capability could become available to users. Iโm not suggesting this automatically creates a security vulnerability or that sensitive information is suddenly exposed. Existing permissions and security controls remain relevant.
However, the environment that was originally assessed and approved may no longer be quite the same environment. That, to me, is the issue.
Why this isnโt really a Microsoft problem
I donโt think we should single Microsoft out here. This is a wider challenge associated with the way cloud services, and increasingly AI services, are delivered. Weโve become accustomed to suppliers
introducing new features, changing functionality and updating services without organisations having to undertake a traditional upgrade. In many respects, thatโs one of the benefits of cloud computing.
But AI introduces some additional considerations. A change to the underlying model, the provider processing information, the data available to an AI system or the actions it can perform may have implications that go beyond a routine product enhancement. And those implications may not be immediately obvious to the people using the service.
For example, an organisation may have approved an AI solution based on particular data processing arrangements, acceptable uses or technical controls. If the supplier subsequently introduces new models or integrations, are those original assumptions still valid? Perhaps they are. Perhaps nothing material has changed from the organisationโs perspective.
But surely thatโs something the organisation should establish, rather than assume? AI governance doesnโt finish when the system goes live.
AI governance after go-live
One area Iโve been considering recently is how organisations approach AI governance beyond the initial assessment and approval. We spend a lot of time discussing AI policies, risk assessments, DPIAs, acceptable use, supplier due diligence, and the controls needed before a solution is deployed. All of those are important, of course. But what happens six months later? Or twelve months later? Who is monitoring changes introduced by the supplier? Who determines whether those changes affect the original risk assessment? And who decides whether additional capabilities should be enabled?
I donโt think the answer is to create another complicated governance process or require a full reassessment every time a supplier releases an update. That would quickly become unmanageable. What I do think is needed is a proportionate approach to identifying material changes and determining whether they require further consideration.
For example:
- Has an additional AI model or provider been introduced?
- Have the processing arrangements or subprocessors changed?
- Does the AI system now have access to additional information or functionality?
- Are new integrations or automated actions available?
- Do existing security controls, policies and approvals still reflect how the system operates?
These arenโt particularly complicated questions, but someone needs to be responsible for asking them. And where a change is material, there should be a clear decision, an identified owner and a record of what was agreed.
What to do about Claude?
Coming back to the Microsoft announcement, I think the immediate action is relatively straightforward. Organisations using Microsoft 365 Copilot should establish whether the change applies to their tenant and review their current Anthropic configuration. Microsoft provides administrators with options to enable Claude for all users, restrict it to selected users and groups, or
disable it. The important thing is to make an informed decision rather than simply accept the default.
From a data protection perspective, organisations should also review their existing Copilot DPIA and determine whether it requires updating. The relevant subprocessor arrangements should be considered, along with any changes needed to AI Acceptable Use Policies or employee training. Microsoftโs move to bring the covered Claude processing into the EU Data Boundary is welcome, but data residency is only one aspect of a wider data protection and governance assessment.
There is also a particular point for organisations that previously configured Claude for use only within Microsoft 365 applications, as Microsoft has indicated that this separate setting is being retired. None of this necessarily means that organisations should disable Claude. There may be perfectly good business reasons to enable it.
The important thing is that the decision is theirs.
A wider question for organisations and advisers
I think this example highlights something we need to consider more carefully as AI becomes embedded in everyday business applications. Traditionally, we might assess a system when it is introduced and revisit it periodically or following a significant change.
With AI, the underlying technology can change quite substantially while the application itself appears much the same to the user. An organisation might continue using the same Copilot interface, but the models, providers, capabilities or processing arrangements behind it may have evolved. That raises a question for those of us working in cybersecurity, data protection and technology governance.
How do we provide assurance that an AI environment remains appropriately controlled, rather than simply confirming that it was appropriately assessed when first deployed? I think this is where ongoing AI assurance will become increasingly important: not necessarily as another standalone compliance exercise, but as part of existing security governance, supplier management, change management and risk oversight.
It also creates an opportunity for organisations and their advisers to work together to understand supplier changes, validate their implications and ensure that the appropriate controls remain in place. At BH Consulting, this is one of the areas weโre exploring as we develop our approach to AI Security and Governance Assurance.
Final thoughts
I donโt see the Microsoft announcement as a reason for alarm. Nor do I think organisations should automatically reject new AI capabilities simply because they introduce additional considerations. AI technology will continue to evolve, and organisations should be able to take advantage of the opportunities that brings.
However, there is an important distinction between a supplier making a capability available and an organisation deciding that the capability is appropriate for its own environment. So as Microsoft changes the default availability of an additional AI model, organisations need to review that change and decide whether it is appropriate for them.
For me, the wider lesson is that AI governance cannot be a one-off exercise undertaken at the point of deployment; it needs to remain relevant as the technology changes. And perhaps most
importantly, organisations need to retain ownership of the decisions, even when the technology itself is being managed and updated by somebody else.
John McWade is head of technical services at BH Consulting