Can you trust Azure OpenAI with confidential data in Switzerland?
Short answer: it can be used with confidential data if it is configured and governed properly. Microsoft's published commitments cover how the service handles your data. They do not replace your own legal, security and access-rights work. Last checked against Microsoft Learn in October 2026; terms and features change, so check the current Microsoft documentation before deciding.
What Microsoft commits to (data, privacy and security page for Azure OpenAI in Foundry):
- Your prompts, completions, embeddings and training data are not available to other customers, not available to OpenAI, and not used to train or improve foundation models without your permission or instruction.
- The models are hosted in Microsoft's own Azure environment and do not interact with OpenAI services such as ChatGPT or the OpenAI API.
- Data stored by optional features (uploaded files, stored conversations) stays in the resource's Azure geography, is encrypted at rest, and you can delete it.
- Abuse monitoring exists: flagged prompts and completions may be sampled and reviewed, by default automatically and in some cases by authorised Microsoft staff (located in the EEA for deployments in the EEA). Customers who qualify can apply to modify abuse monitoring so that this storage and human review do not take place. Retention periods and eligibility are set by Microsoft: check the current documentation.
What remains your responsibility:
- Choosing a deployment type that matches your residency needs. Standard deployments process data within the chosen geography; Global deployments may process it in any region; EU Data Zone deployments process it within the EU data zone, so processing in Switzerland is not guaranteed. Data at rest stays in the designated geography. Availability depends on the model and region.
- Under the Swiss nFADP you remain the controller: keep a record of processing (smaller organisations with under 250 employees are partly exempt, unless they process sensitive data on a large scale or carry out high-risk profiling), inform the people concerned, sign a proper processing agreement, assess the risk of any disclosure abroad and put safeguards in place, and run a data protection impact assessment where the risk is high. Professional or banking secrecy duties are not lifted by a cloud contract.
- Who can see what in Microsoft 365, which data is sent to the model, and what you log. This is where most real leaks happen, and the rest of this guide covers it.
Why is data security crucial with Azure OpenAI and Microsoft 365?
AI models, whether generating documents, assisting with email analysis, or facilitating collaborative work in Microsoft 365, inevitably process internal data flows, sometimes sensitive ones. The integration of models such as GPT or add-ins that enhance Outlook and Word raises new risks: involuntary exposure, unauthorized access, or transfer outside Swiss borders. The entry into force of the new Federal Act on Data Protection (nFADP) in September 2023 imposes strict requirements on transparency, traceability and control of the processing of personal information. It does not require data to be hosted in Switzerland: it regulates the disclosure of data abroad (adequate country or appropriate safeguards, information duty).
Distinguishing common usage and critical deployments in private AI
The use of AI in Microsoft 365 is no longer limited to simple task assistants. Advanced semantic analysis, automatic generation of reports, extraction of structured information: each new use case adds complexity to the processing chain.
In many Swiss companies, the issue of data storage and transit is pressing. The adoption of local LLM (Large Language Model) deployments, coupled with Swiss hosting, helps DPOs and CISOs keep information flows within the national or company perimeter. Houle thus offers fully managed solutions on Swiss infrastructure, with fine governance for each compartment (Outlook, Word, SharePoint, Teams, etc.).
Swiss (FADP) and European (GDPR) requirements for responsible AI
Compliance with nFADP means each company must clearly identify its AI processing activities. The entire data journey must be tracked: from collection (email, document, form) to usage (analysis, generation, archiving). A precise inventory of tools integrated into Microsoft 365, AI uses, and external flows (to Azure OpenAI or other clouds) is necessary.
In parallel, the GDPR (where it applies) sets out data protection by design and data minimisation: every solution should limit collection, anonymise or pseudonymise information where possible, and let people exercise their rights (access, erasure, portability). Regular checks and, where consent is the legal basis, documented consent are good practice when adding add-ins and AI connectors in Microsoft 365.
Azure OpenAI in Switzerland: how to protect the confidentiality of processing?
Azure OpenAI is a managed Azure service, and Microsoft documents that prompts and completions are not used to train its models. Confidentiality then depends mostly on your configuration:
- Pick the deployment type deliberately (standard, data zone or global) and check that your model is available in the region you need; geography-based deployment types often arrive after global ones.
- Manage roles and assignments in Microsoft Entra ID (formerly Azure Active Directory), with least privilege.
- Use the default encryption at rest, and a customer-managed key if your policy requires it.
- Decide whether to request modified abuse monitoring, and understand the trade-offs before you do.
- Treat preview features with caution: Microsoft notes they may have different privacy practices.
In our projects we start from the residency requirement, choose the deployment type and region accordingly, and document where prompts, logs and stored data end up. We do not promise more than the Microsoft terms and your configuration allow.
Risk prevention: audit, training and collaborative governance
The real challenge of securing AI is not only technological. Supporting business teams and IT in understanding limitations, the new risks (prompt injection, hallucinations, false positives), and access settings is essential.
Houle recommends a three-step approach:
- Compliance audit of Microsoft 365 and Azure OpenAI, to map all data entry and transit points for sensitive data.
- Definition of internal policies for AI use, via specific charters, consent models, and setting rules in Microsoft Purview and Azure Security Center.
- Personalized training and workshops, adaptable for non-technical employees, to support proper use of add-ins and raise awareness on data sharing challenges.
The entire document workflow – from automated contract generation in Word to trend analysis in Outlook – can then be secured by continuous supervision mechanisms (alerts, logging, reporting shared with the DPO).
Houle solutions: private AI, productive add-ins and local compliance
Houle offers tailor-made solutions, built around add-ins for Outlook, Word, and SharePoint, which integrate natively with Foundry to support data sovereignty. Our offerings include:
- Deployment of private AI on Swiss infrastructures (on-premises or Microsoft Switzerland cloud);
- Integration of local language models to limit exposure to public clouds;
- Addition of structured reporting modules to support traceability at the request of the DPO or legal department.
The aim is that the organisation knows, and can document, where each data flow goes and who can read the logs. Absolute guarantees do not exist, so Microsoft's current terms and your contracts should be checked regularly. Internal workflows (document management, data extraction, automation) gain strongly in productivity, while respecting a demanding regulatory framework under constant audit.
Govern data and access before connecting AI
A private AI does not fix a poorly governed Microsoft 365 environment: it inherits it. Before any deployment, three workstreams are worth tackling in this order.
Map and classify content
Start with targeted inventories of SharePoint, OneDrive and Exchange, then classify content with Microsoft Purview and sensitivity labels. A simple scale is enough: public, internal, confidential, HR or health data. Each level should trigger concrete rules for sharing, export and access to AI add-ins. Avoid copying files solely to feed a model, and document every extraction or automated processing step in line with your retention policies.
Keep access rights clean
Permissions accumulate in Microsoft 365: inheritance, one-off shares, documents copied into several Teams channels, access never removed after a change of role. AI can help spot these gaps (unusual access, inactive accounts, mass exports) on two conditions: proposals to grant or remove access stay under human validation, and the access logs are themselves protected. A monthly review of critical locations, easy to correct by hand, is a sensible starting rhythm.
Filter prompts and trace usage
An AI add-in should check what it sends to the model: detection of risky identifiers (names, IBANs, AHV numbers), automatic anonymisation or a validation request before sending. Log requests and responses so that you can answer an inspection or a data subject's request.
Prepare the teams
| Measure | Purpose |
|---|---|
| AI usage policy by data category | State clearly what is and is not allowed |
| Training for HR, IT and business teams | Reduce sharing and prompt mistakes |
| Incident drill (shadow AI, accidental leak) | Test the response and assess the impact under the nFADP |
| Periodic review of approved add-ins | Limit the exposure surface |
Conclusion
Securing private AI in Microsoft 365, coupled with compliance with nFADP and GDPR, requires a comprehensive approach that relies on technology, governance, and human support. Houle, as a trusted partner on the Swiss market, enables organizations to adopt AI in their daily tools (Outlook, Word, SharePoint) while reducing the risk of compromise or non-compliance.
Do you want to audit your AI uses, strengthen data protection, or deploy tailor-made solutions in a Swiss legal framework? Contact our experts for an initial concrete analysis, and discover the real levers of innovation, productivity, and peace of mind for 2025 and beyond.