Data boundaries: where Copilot and Foundry process.
Microsoft 365 processes within the EU Data Boundary, Foundry in Switzerland North, in the EU data zone or globally by model, each with date and source.
Every cell says whether a data class may be processed on a platform, needs a documented decision, or has no business being there.
Data class
Microsoft 365 and CopilotProcessing: EU Data Boundary; Claude models excluded and not enabled by default
Foundry, regional deploymentProcessing: Switzerland North, for the models available there and for hosted agents
Foundry, EU data zoneProcessing: Azure regions in the EU, without Switzerland
Foundry, Global StandardProcessing: Any Azure region with free capacity
Claude and Grok in FoundryProcessing: Globally or in the US data zone, no regional deployment
Public contentWebsite, brochures, published reports
applies
applies
applies
applies
applies
Internal content without personal referenceDirectives, templates, minutes without names
applies
applies
applies
applies
appliesDocuments without personal reference
Personal dataEmployees, customers, members
appliesWith permission checks, labels and Purview logging; Claude models only after a management decision
appliesRegion and log recorded before the start
documented decisionAfter a documented decision with the data protection officer
noOnly data without personal reference
documented decisionAnonymised cases or an explicit agreement with the people concerned
Sensitive personal data and professional secretsHealth, mandates, patient records
documented decisionDecision of the data protection officer, a dedicated label and a DLP policy for the Copilot location
documented decisionRegional deployment, private network, decision of the data protection officer
no
no
no
Legend:appliesdocumented decisionno
As of 3 October 2026, checked on Microsoft's pages. Our frame for the start, not legal advice; the decision per application is taken with your data protection officer.
01
Every layer has its own place.
A prompt in Copilot stays in Microsoft 365 and therefore within the EU Data Boundary; Claude models are excluded from it and enabled for no users by default in the EU, the EFTA and the UK. In Foundry, model and deployment type decide: regionally in Switzerland North, in the EU data zone, which does not include Switzerland, or globally. Copilot and data agents in Fabric use Azure OpenAI; if the capacity sits in Switzerland, that is outside the EU Data Boundary and the US, administrators must allow processing outside the region in the tenant settings. For every application we record before the start which of these rows applies.
02
The deployment type decides where a model computes.
In Foundry you create a model as a deployment in your subscription, and the type of deployment sets the place: Standard computes in the chosen Azure geography, Data Zone Standard within the data zone, Global Standard in any Azure region with free capacity. As of September 2026, GPT-4.1, GPT-4o and the embedding models run as a regional deployment in Switzerland North, newer GPT models process in the EU data zone or globally, Claude and Grok globally or in the US data zone. Hosted agents, that is your own agent code in a container, run in Switzerland North and Switzerland West. The model with the best test result and the model with the right place are two questions, and the proposal answers both.
03
What «no training» covers and what Purview keeps.
Enterprise Data Protection applies to Copilot Chat and the licensed experiences: prompts and answers do not serve to train the foundation models, and who stores them for how long is governed by your retention policies. Copilot writes prompts, answers and the sources used into the Purview audit log and into a hidden folder in the person's mailbox, where retention and eDiscovery reach them. Since 1 May 2026, prompts and responses from Foundry applications also flow into audit, the data security posture view, classification and loss prevention when the application calls with the person's Entra identity. Labels travel along, and encrypted documents need the extract right, otherwise Copilot may not summarise them.
04
The duties under the data protection act stay with you.
Whichever layer processes, you remain responsible. The register of processing activities names the AI application with purpose, data categories and recipients; with Microsoft, the Product Terms and the Data Protection Addendum serve as the processor agreement. Professional secrets, as in law firms, practices and hospitals, need their own review before a document reaches a layer, and for processing with a high risk the act requires a data protection impact assessment. We record the outcome per application in writing, with your data protection officer: layer, model, deployment type, region, retention, data classes, approvals.
05
Purview shows what the directive allows.
The data security posture view for AI shows which AI services are in use and where personal data appears in prompts; the Purview browser extension on managed devices also captures AI websites without a business contract. Endpoint DLP warns or blocks when someone pastes sensitive data into such a website, and a policy for the Copilot location keeps files with selected labels out of the answers; files someone uploads directly it does not check. A short directive says which confidentiality level may go into which tool, and the notice of a stopped prompt leads to the matching passage in it.
What you decide.
Which data class may go to which layer?
Options
1Public and internal content without personal reference
2Personal data of employees, customers and members
3Sensitive personal data and professional secrets
Our recommendation: Content without personal reference may go to any approved layer, including global deployments. Personal data we process in Microsoft 365 and in regional deployments in Foundry, with labels and log; the EU data zone joins after a documented decision. Sensitive personal data and professional secrets stay in regional deployments or in Microsoft 365 with a dedicated label and DLP policy, with the decision of your data protection officer.
What we advise against: Personal data in global deployments just because the newest model runs there. Development and test data without personal reference are enough to check a model.
Regional model or newest model?
Options
1Regional deployment in Switzerland North with the models available there
2EU data zone with newer models, after a documented decision
3Global Standard with the newest model, for data without personal reference
Our recommendation: For personal data, test the regional deployment first; for classification, extraction and summarisation the model there is often enough. If the test set shows that only a newer model solves the task, we review the EU data zone with your data protection officer or anonymise the data before it reaches the model.
What we advise against: Choosing the model first and justifying the region afterwards.
Enable Claude models in Copilot?
Options
1Do not enable; Copilot works with the router and the OpenAI models
2Enable for a named pilot group, after a management decision
3Enable for everyone
Our recommendation: For Swiss tenants the approval stays off until a case demands it; then for a named group with data without personal reference and with the written decision that records that these models are excluded from the EU Data Boundary. Models with data retention at Anthropic need a separate approval.
What we advise against: Enabling for everyone because one team wants a particular model. Without a data class and a decision, the exception from the data boundary runs along unnoticed.
Who signs off the data route of an application?
Options
1Your IT alone
2Business team and IT
3Business team, IT and the data protection officer
Our recommendation: All three, on one page per application: which layer processes with which model and deployment type, in which region, how long prompts and logs remain, which data classes are allowed, who approves exceptions and when the statements are checked again. This sheet belongs in the register of processing activities.
What we advise against: A decision that covers the whole tenant. The answers differ per layer, model and deployment type, so per application.
Status and sources.
As of: October 2026
As of 18 September 2026
Claude in Microsoft Copilot: in the EFTA only after approval.
Anthropic is a Microsoft subprocessor, and Microsoft's Product Terms and Data Protection Addendum apply. In the EU, the EFTA and the UK the models are enabled for no users by default and are excluded from the EU Data Boundary. Models with data retention at Anthropic need a separate approval.
All Claude models run in Microsoft Foundry as a Global Standard deployment, hosted on Azure or by Anthropic; a data zone exists for some models in the US only. Authentication uses Entra ID or an API key. Subscriptions from the Cloud Solution Provider programme are not supported for Claude.
Grok models run in Foundry as a Global Standard deployment with global processing. A data zone for Grok exists in the US only, and there is no regional deployment.
Hosted agents run in Switzerland North and Switzerland West.
Your own agent code, built with Agent Framework, LangGraph or no framework at all, runs as a container in Foundry Agent Service: with its own Entra ID, its own endpoint, an isolated sandbox per session and traces in Application Insights. Both Swiss regions are on the list of supported regions.
Prompts and responses from Foundry applications flow into the audit log, the data security posture view for AI, classification, loss prevention, retention and eDiscovery in Microsoft Purview. Policies apply when the application calls with the Entra identity of the user.
Data agents on Swiss Fabric capacities need an approval.
Copilot and data agents in Fabric use Azure OpenAI. If the capacity sits outside the EU Data Boundary and the US, as in Switzerland, administrators must allow in the tenant settings that data is processed and stored outside the region. If a data agent is called from other services, their rules apply in addition.
This is how we put it into practice: data protection under the FADP, enforced in Microsoft 365, with register, labels and logs. Compliance & data protection
Is the EU Data Boundary enough for Swiss personal data?
For most personal data, yes: Switzerland lists the EU states as countries with adequate data protection, and Microsoft processes and stores Microsoft 365 customer data within the EU Data Boundary. The exceptions are in the contract and the documentation, such as certain support cases or the models from Anthropic; for professional secrets and sensitive personal data we review them one by one with your data protection officer.
What does the EU data zone mean for us in Switzerland?
A model with the Data Zone Standard deployment type computes in Azure regions within the EU; Switzerland is not among them. For data without personal reference that is irrelevant. For personal data we record the decision in writing, as for any processing outside Switzerland, and first check whether the regional deployment in Switzerland North is enough.
May we process professional secrets with Copilot?
Your data protection officer decides that, and we prepare the groundwork: which documents carry a professional secret, which label they get and which policy keeps them out of Copilot's answers. Many organisations start with content without secrets, such as directives, templates and internal procedures, and open further areas only after the decision.
Do Microsoft staff see our prompts?
The Product Terms and the Data Protection Addendum govern access; prompts and answers sit in your tenant, where the Purview audit log, retention and eDiscovery reach them. Which logs arise at the service itself, for instance in Azure OpenAI, we record per application with your data protection officer, together with the data class allowed there.
What happens to files someone uploads directly into Copilot?
They are processed with the person's rights and appear in the audit log, but the DLP policy for the Copilot location does not check them; it acts on labelled files and emails in SharePoint, OneDrive and Exchange. That is why the directive says what may be uploaded, and endpoint DLP acts on the device before a file reaches an AI website.
Do we need a data protection impact assessment?
If the processing can bring a high risk to the personality of the people concerned, the data protection act requires one. With AI applications that is often the case when sensitive personal data is processed at scale or decisions about people are prepared. We supply the technical details for it, data route, region, retention, access and logs; the assessment itself is carried out by your data protection officer.
Next chapter · Chapter 7
Which levers move the AI bill at Microsoft.
How does the meter run at Microsoft, and what moves the bill?