CompanyScope
by Janus Compliance

Embedded productivity AI

Microsoft 365 Copilot compliance: GDPR, AI Act, DPA, training, transfers

Independent compliance research from Janus Compliance. Reviewed by Michael K. Onyekwere, CIPP/E. Last reviewed 2026-10-14. Not legal advice.

Share this Microsoft 365 Copilot profile:Share on XBluesky

TL;DR. Microsoft acts as processor under its DPA, and its Product Terms commit that Microsoft's generative AI services will not use Customer Data to train foundation models. For tenants provisioned in the EU or EFTA, Copilot is an EU Data Boundary service with three main Copilot-specific exceptions: Bing web search queries (outside the DPA and the boundary), flex routing (inferencing outside the boundary at peak demand) and Anthropic models (excluded from the boundary). Previews are also excluded from the boundary. Web search is on by default in every commercial tenant, whatever its creation date. Flex routing (eligible EU and EFTA tenants, not Multi-Geo) and Anthropic models in Word, Excel and PowerPoint (EU, EFTA and UK tenants) are on by default for tenants created after 25 March 2026. For older tenants (those that existed on or before 25 March 2026 for flex routing, and before 25 March 2026 for the Word, Excel and PowerPoint setting), Microsoft asks admins to check the setting rather than stating a default. UK tenants are outside the EU Data Boundary and flex routing. The wider Anthropic subprocessor setting is off by default throughout the EU, EFTA and UK. Copilot Chat, used with a work or school account, is under the same DPA and is listed in scope for the HIPAA BAA.

DPO action: in the Microsoft 365 admin centre (Copilot > Settings > View all), check "Flex routing during peak load periods" (EU and EFTA tenants) and "AI providers operating as Microsoft subprocessors", which holds the Anthropic and OpenAI settings and the separate Anthropic setting for Word, Excel and PowerPoint. Decide whether to allow web search, which is on by default in commercial tenants and sends generated queries to Bing outside the DPA. Record each decision in the DPIA. For PHI, the BAA applies by default through the DPA, but minimum-necessary access across the Graph is the main HIPAA control.

What the tool does

Microsoft 365 Copilot is an AI assistant embedded in Word, Excel, PowerPoint, Outlook, Teams and the wider Microsoft 365 stack. It draws on the user's tenant data (emails, files, chats) through Microsoft Graph and combines that with large language models to generate text, summaries, action items and draft documents. Microsoft has renamed it Microsoft Copilot, and Microsoft 365 Copilot Chat is now Microsoft Copilot Chat, on Microsoft Learn (Microsoft states there are no changes to security, compliance and privacy), although the Product Terms and some licences still use the old names.

Copilot Chat is used with a work or school account and is grounded in the web by default rather than in organisational content, but Microsoft applies the same DPA, Product Terms and enterprise data protection to it. Buyers should still record it separately in a DPIA because its grounding differs. Consumer Copilot (Microsoft 365 Personal and Family) is a different product, and Copilot Studio (the agent platform) is a separate compliance evaluation again.

Data processed

The hard part of this profile is that Copilot processes a lot:

Special-category likelihood: Very high. Copilot reads anything the signed-in user can read. In a typical SME, that includes HR records, performance reviews, financial data, and customer correspondence. DPIA is not optional for any deployment.

Default geographic processing: Copilot is an EU Data Boundary service for tenants provisioned in the EU or EFTA, subject to the exceptions set out under EU / UK transfer position below (flex routing, Anthropic models and Bing web search queries). UK tenants are outside the EU Data Boundary.

DPA availability

Microsoft 365 Copilot is governed by the Microsoft Product Terms and Data Protection Addendum (DPA), both publicly available and incorporated by reference into all M365 commercial agreements.

The DPA makes the customer the controller and Microsoft the processor of Personal Data, which covers Copilot prompts, responses and the tenant content Copilot reads. Microsoft acts as an independent controller when it processes Personal Data for business operations incident to providing the Products and Services to the customer, which the DPA defines as the processing operations the customer authorises in that section. Several Copilot routes fall wholly or partly outside the DPA:

Breach notification: the DPA commits Microsoft to notify the customer of a Security Incident "promptly and without undue delay", and Appendix A adds that notification will be made "without undue delay and, in any event, within 72 hours."

Subprocessor list

Microsoft publishes the Microsoft Online Services Subprocessors List on the Service Trust Portal. The AI model providers that matter for Copilot are:

For subprocessors that support AI functionality, the May 2026 DPA gives "30 days' notice with the ability to disable use of that subprocessor until at least 6 months after notice", rather than the 6 months' advance notice that applies to other new subprocessors with access to Customer Data. A new AI subprocessor can therefore receive Customer Data 30 days after notice, which matches the OpenAI timeline (added 23 June 2026, on by default from 24 July 2026), and the admin's protection is the ability to disable it.

The carve-out worth flagging: when Copilot uses Anthropic models, the processing falls outside the Microsoft EU Data Boundary. Microsoft's page on Anthropic models in Microsoft Online Services (last updated 18 September 2026) states:

"Anthropic models deployed in Microsoft offerings such as Microsoft Copilot, Researcher, Copilot Studio, Power Platform, and Copilot in Microsoft 365 apps are currently excluded from the EU Data Boundary, and when applicable, in-country processing commitments. Customers within the EU Data Boundary and customers in the UK have Anthropic models disabled by default."

Microsoft turns Anthropic models on by default for most commercial-cloud customers, but not in the EU, EFTA or UK. In those regions the "AI providers operating as Microsoft subprocessors" setting for Anthropic appears with its default set to No users, as it has since the setting was introduced in December 2025. A separate setting introduced on 3 April 2026, described by Microsoft as "Copilot in Microsoft 365 apps with Anthropic models", makes Anthropic the default model for Copilot in Word, Excel and PowerPoint. Microsoft states that this setting is on by default for EU, EFTA and UK tenants created after 25 March 2026, and tells tenants that existed before that date to check the Message Center for their default. When Anthropic models are used in Word, Excel or PowerPoint, the processing takes place outside the EU Data Boundary. Both settings are reached through Copilot > Settings > View all > AI providers operating as Microsoft subprocessors in the Microsoft 365 admin centre. When the Anthropic subprocessor setting is enabled for all users or for specific users and groups, the Word, Excel and PowerPoint setting is unavailable and cannot be changed. Microsoft decommissioned the earlier Anthropic independent processor (IP) setting on 1 May 2026.

Training-on-customer-data position

Copilot does not use Customer Data, including prompts, responses and data accessed through Microsoft Graph, to train foundation models. The contractual commitment is in the Product Terms (Universal License Terms for Online Services, "Use of Content for Training"): "Microsoft Generative AI Services will not use Customer Data to train any generative AI foundation model, except pursuant to Customer's documented instructions." Microsoft's Copilot privacy documentation describes the position this way: "Prompts, responses, and data accessed through Microsoft Graph aren't used to train foundation LLMs, including those used by Microsoft Copilot."

Optional user feedback is treated differently. Microsoft may use it to improve Copilot, states that it does not use that feedback to train the foundation LLMs, and gives admins controls to manage it.

Sources: Microsoft Product Terms, Universal License Terms for Online Services; Microsoft Learn, "Data, Privacy, and Security for Microsoft Copilot".

Retention and deletion

Copilot stores each prompt and response, with citations to the content used to ground the response, as the user's Copilot activity history. Microsoft's Purview documentation states that this data is held in a hidden folder in the mailbox of the user who runs the AI app, where compliance administrators can search it with eDiscovery. Admins set retention and deletion with Microsoft Purview retention policies, and users can delete their own activity history through the My Account portal. Microsoft states that Copilot services have opted out of the abuse monitoring available in Azure OpenAI, which includes human review of content.

After a subscription ends, the DPA keeps Customer Data in a limited-function account for 90 days so the customer can extract it, and Microsoft then deletes it within a further 90 days unless the DPA authorises retention. The opt-in Anthropic models with Data Retention are the exception to this picture: Anthropic, not Microsoft, keeps most inputs and outputs for up to 30 days. Where Anthropic's trust and safety classifiers identify potential violations of its Usage Policy, Microsoft's page states that Anthropic "may retain content (inputs and outputs) for up to two years and trust and safety classification scores for up to seven years."

EU / UK transfer position

The position depends on where the tenant is provisioned, when it was created and how its admin settings are configured.

EU Data Boundary scope. Under the Product Terms, a Microsoft 365 tenant provisioned in the EU or EFTA is in scope for the EU Data Boundary unless the customer has also bought the Multi-Geo Capabilities add-on. Microsoft 365 Copilot and Microsoft 365 Copilot Chat are listed as EU Data Boundary Services, for which Microsoft commits to store and process Customer Data and Personal Data within the boundary. Copilot was added to Microsoft's data residency commitments on 1 March 2024. The Product Terms list limited transfers that apply to every EU Data Boundary service, for example remote access by Microsoft personnel outside the boundary to operate, troubleshoot, support and secure the service.

Flex routing (EU and EFTA). Flex routing lets LLM inferencing for Copilot and Copilot Chat take place outside the EU Data Boundary during periods of peak demand. Microsoft's EU Data Boundary transfers page states that Copilot prompts, responses and grounding data may then be processed outside the boundary and that pseudonymous user IDs may be stored there for security and operational purposes. Other data at rest stays inside the boundary.

Microsoft's own pages differ on this point. Microsoft's main Copilot privacy page, as read on 14 October 2026, states "For EU customers, Microsoft Copilot is an EU Data Boundary service" and "EU traffic stays within the EU Data Boundary while worldwide traffic can be sent to the EU and other countries or regions for LLM processing." It does not mention flex routing. The flex routing page (last updated 29 September 2026) and the EU Data Boundary page on ongoing partial transfers describe the peak-demand exception. Those two pages also word the destinations differently: the flex routing page says inferencing "may occur in the United States, Canada, or Australia", while the transfers page says processing may take place outside the boundary "including in the United States, Canada, and Australia". A DPIA should record both statements alongside the tenant's actual flex routing setting.

Anthropic, OpenAI and web search. Anthropic models are excluded from the EU Data Boundary wherever they are used (see Subprocessor list above). OpenAI-operated models are included in the EU Data Boundary "except as otherwise noted in the EU Data Boundary documentation", which allows an AI subprocessor to process and store a pseudonymised user ID or session GUID outside the boundary, including in the United States. Generated web search queries sent to Bing are outside the EU Data Boundary, and web search is on by default in commercial tenants (see DPA availability above).

Taken together. For an EU or EFTA tenant created after 25 March 2026, Microsoft's defaults allow three routes outside the EU Data Boundary: generated web search queries sent to Bing, inferencing at peak demand through flex routing (eligible tenants), and Anthropic models in Word, Excel and PowerPoint. The wider Anthropic subprocessor setting stays off unless an admin opts in. For an older tenant, web search is still on by default, and the flex routing and Word, Excel and PowerPoint settings should be checked in the admin centre because Microsoft's public pages state no default for them.

UK tenants. A tenant provisioned in the UK is outside the EU Data Boundary and outside flex routing. Under the Product Terms, Microsoft stores certain Customer Data at rest in the UK for a UK-provisioned tenant, including stored content of interactions with Microsoft 365 Copilot and Microsoft 365 Copilot Chat, but LLM processing is not limited to the UK: Microsoft's privacy page states that "Customers outside the EU may have their queries processed in the US, EU, or other regions." The Anthropic subprocessor setting is off by default for UK tenants, while the Word, Excel and PowerPoint Anthropic setting is on by default for UK tenants created after 25 March 2026.

Transfer mechanism. The DPA makes all transfers of Customer Data and Personal Data out of the EU, EEA, UK and Switzerland subject to "the terms of the 2021 Standard Contractual Clauses implemented by Microsoft". It defines those clauses as the processor-to-processor module between Microsoft Ireland Operations Limited and Microsoft Corporation (Module Three in the Commission's numbering, which is our gloss rather than the DPA's wording), so the customer is not a party to them. Transfers from the UK are also subject to the IDTA, the name Microsoft uses for the UK Information Commissioner's international data transfer addendum to the SCCs.

EU-US Data Privacy Framework: Microsoft Corporation appears on the participant registry with all three certifications, EU-US, Swiss-US and the UK Extension, each running to 31 August 2027. When the register was checked on 9 and 14 October 2026 the status of all three was "Active". On 8 September and 4 October 2026 it had read "Active - Re-certification under Review", the register's wording for an annual re-certification in progress. The DPA also states that Microsoft is certified to the EU-US and Swiss-US frameworks and the UK Extension.

For EU and EFTA tenants, Copilot remains an EU Data Boundary service under the Product Terms, but Bing web search queries, flex routing and Anthropic models can each take processing outside the boundary. Web search is on by default in every commercial tenant, and the other two are on by default for tenants created after 25 March 2026 (flex routing for eligible tenants only). A DPIA should record the tenant's actual settings and the decision taken on each. - My read

Security documentation

Microsoft's certifications relevant to Copilot, as stated on its public pages:

Microsoft's standards coverage is broad, but its public pages name Copilot in fewer certification scopes than the wider portfolio suggests. For Copilot, the questions that most often need a decision concern default settings and configuration.

HIPAA & BAA position

Microsoft's HIPAA Business Associate Agreement is part of the DPA by default for any customer that is a covered entity or business associate. Under the DPA, execution of the customer's agreement includes execution of the BAA, and the customer can opt out by written notice, so there is no separate BAA to sign. Microsoft lists both Microsoft Copilot and Microsoft Copilot Chat as in-scope services for Office 365 Commercial and GCC. Microsoft's HIPAA page still calls the document the Online Services DPA, which is now the Products and Services DPA.

The harder obligation is the minimum-necessary standard, and it is where Copilot differs from a standalone API. Copilot reads anything the signed-in user can read across the Microsoft Graph: every mailbox, file, Teams chat, and SharePoint site within that user's permissions. In a healthcare tenant that is a large PHI surface, and the HIPAA Privacy Rule's minimum-necessary standard (45 CFR 164.502(b)) requires reasonable efforts to limit PHI used or disclosed to the minimum necessary for the purpose. A BAA does nothing to limit what Copilot can surface. Permission hygiene and a DLP layer (Microsoft Purview or equivalent) are the controls that do, and they are the buyer's to configure.

The BAA is the contractual piece. HIPAA's remaining obligations (the Security Rule's risk analysis, security awareness training and audit controls, the Privacy Rule's minimum-necessary standard, and the Breach Notification Rule) still fall on the covered entity, and with Copilot the minimum-necessary and audit-control work is unusually heavy. See HIPAA for AI tools for the full walkthrough across vendors.

AI Act role + risk classification

Microsoft publishes AI Act readiness materials. Useful evidence for your audit file but does not relieve your own obligations.

DPIA prompts (for your use case)

  1. Have you checked your flex routing setting? It is on by default for eligible EU and EFTA tenants created after 25 March 2026. For older tenants, Microsoft asks admins to confirm the setting rather than stating a default.
  2. Have you decided whether to allow Anthropic models? Anthropic has been a Microsoft subprocessor since 7 January 2026, on by default in most commercial regions but off by default in the EU, EFTA and UK. Check the separate Word, Excel and PowerPoint setting as well, which is on by default for EU, EFTA and UK tenants created after 25 March 2026. Anthropic processing is outside the EU Data Boundary.
  3. Have you decided on OpenAI-operated models and web search? OpenAI-operated models have been enabled by default for eligible commercial customers since 24 July 2026. Web search is on by default in commercial tenants, and its generated queries go to Bing outside the DPA, HIPAA compliance and the EU Data Boundary.
  4. Have you mapped which roles can access Copilot? Copilot reads anything the user can read. Over-permissioned mailboxes and SharePoint sites become AI-exposure points.
  5. Have you implemented Microsoft Purview or equivalent DLP to control what Copilot can surface? Copilot surfaces whatever the signed-in user can access, and Purview sensitivity labels and DLP policies are the controls that narrow it.
  6. AI Act Annex III applicability: is Copilot being used for any purpose listed in Annex III, such as recruitment, promotion or termination decisions, or monitoring and evaluating staff performance? If so, the high-risk deployer obligations will apply from 2 December 2027 unless the Article 6(3) exception applies (it never covers a system that profiles people), a system put into service before that date is caught only if its design changes significantly afterwards (public authorities: 2 August 2030), and a use that turns Copilot into a high-risk system can make you its provider under Article 25(1)(c).
  7. Copilot Studio agents: if you've built or enabled agents, treat each as a separate AI system for AI Act purposes - risk-classify each independently.

Unresolved questions / red flags

Related profiles

Sources checked

Related reading


Share this Microsoft 365 Copilot profile:Share on XBluesky

Talk to Michael about Microsoft 365 Copilot - or your AI vendor governance more broadly

CompanyScope's public profiles cover the general picture. Michael runs Janus DPO-as-a-Service for businesses that need ongoing AI vendor governance, and writes one-off CIPP/E-reviewed Vendor Risk Notes for specific procurement decisions. Tell him what you're actually trying to clear.

A sentence or two is plenty.

Your context goes only to Michael. We don't share with the vendor or anyone else. Privacy notice.

Subscribe to the AI Agent Incident Register

Every new Register entry delivered with the legal analysis: the incident, the duty engaged, who is liable across the chain, and what governance would have prevented it. Written by Michael K. Onyekwere, CIPP/E. Free.

Subscribe - free

Delivered via Compliance Engineering on Substack, which handles your subscription and consent. Unsubscribe any time. Privacy notice.

For ongoing AI compliance support, work with Janus DPO-as-a-Service. For other vendors, browse the full index, or see real agent failures analysed legally in the AI Agent Incident Register.