Guide to using Claude with Microsoft 365 while maintaining compliance

On 18 September 2026 Microsoft revised its guidance on Anthropic models in Microsoft Online Services, eight days after its previous revision. The page now sets out three positions for those models: under Microsoft's contract, under Anthropic's with data retention, and, for the newest models, one or the other depending on an organisation's eligibility. That page governs the first of two arrangements by which Claude, Anthropic's AI assistant, reaches Microsoft 365. The second is Anthropic's own product, connected through a connector and a set of Office add-ins. Each has its own contract, control point and default state, and a supervised institution in Geneva or Dubai needs to know which one it is running.

The two look alike from a desk. In both, a member of staff types a question and Claude answers over the organisation's documents. Underneath, the party processing the data, the agreement that covers it and the administrator able to stop it are all different, and so is the outsourcing register entry each requires.

Claude models inside Microsoft Copilot

In the first arrangement Microsoft licenses Anthropic's models and serves them inside its own products: Microsoft Copilot, the Researcher agent, Power Platform, Copilot in the Microsoft 365 apps and Copilot Studio, Microsoft's tool for building agents. The user stays in Microsoft's interface, signs in with Microsoft Entra ID, the identity service behind every Microsoft 365 account, and, where the feature offers it, picks Claude or sees an indicator that a Claude model is in use.

Since 7 January 2026 Anthropic has been a Microsoft subprocessor, a supplier that handles data on Microsoft's behalf, working under Microsoft's Product Terms and Data Protection Addendum. The Customer Copyright Commitment, Microsoft's undertaking to defend clients against copyright claims over Copilot output, extends to Anthropic models in Microsoft Copilot and Copilot Studio. Microsoft lists its subprocessors in the Service Trust Portal, the reference an outsourcing entry should cite. Retrieval runs on Microsoft Graph, the layer that connects Copilot to mail, files and chats, so Copilot reaches only what the signed-in user may already open.

The two look alike from a desk. Underneath, the party processing the data, the agreement that covers it and the administrator able to stop it are all different.

One category sits outside that contract. Models Microsoft labels "Anthropic models with Data Retention" run under Anthropic's own commercial terms and data processing addendum, are off by default everywhere and need a separate opt-in. Under that arrangement Anthropic stores most inputs and outputs for up to 30 days, content flagged by its trust and safety classifiers for up to two years and classification scores for up to seven years, and does not use retained data for training without express permission. For its newest models the page adds an eligibility test: some organisations receive them as subprocessor models under Microsoft's terms, with no retention and no further opt-in, and others receive them only as data-retention models. Which side an organisation is on is a question for the Microsoft account team.

The default state depends on region.

• Most commercial cloud tenants have Anthropic models on by default; a tenant is an organisation's own instance of Microsoft 365.
• Tenants in the European Union, the European Free Trade Association, EFTA, which includes Switzerland, and the United Kingdom have them off by default, because the models are excluded from the EU Data Boundary, Microsoft's commitment to store and process EU and EFTA clients' data inside those regions, and from in-country processing commitments where those apply. A second setting, introduced on 3 April 2026, lets those tenants make Anthropic the default model for Copilot in the Microsoft 365 apps.
• Non-federal organisations in the United States Government Community Cloud, GCC, gained a setting on 22 July 2026; federal organisations in that cloud, and everyone in GCC High, Department of Defense and other sovereign clouds, have no option at all.

For a Geneva institution the switch is off until someone turns it on. For a DIFC institution on commercial cloud it is on until someone turns it off. The control sits in the Microsoft 365 admin centre under Copilot, Settings, View all, then "AI providers operating as Microsoft subprocessors", where an administrator assigns access to named users or security groups.

Questions people ask