
An AI Company Brain makes approved organisational knowledge accessible through search and generated answers. It is not an all-knowing system: quality and access depend on sources, retrieval, permissions and freshness.
The technical statements were reviewed on 8 September 2026. For data residency, Microsoft's current pages on Copilot Studio, Microsoft 365 Copilot and Azure OpenAI are authoritative.
Five layers
1. Knowledge sources
These include SharePoint, OneDrive, approved business systems, databases, documents and APIs. Before technical integration, clarify ownership, freshness, retention and the permitted audience.
2. Retrieval and index
Documents are typically prepared, split into sections and made searchable. Vector search is one option, but not mandatory for every source: structured data can be queried through controlled functions or APIs. Chunk sizes or a particular embedding model are not a universal rule; they must be evaluated with representative questions.
3. Language model
The model formulates an answer from the context it is given. Azure OpenAI and models in Microsoft Foundry are available depending on region and deployment. A product name alone does not prove Swiss data residency.
4. Orchestration and security
Orchestration selects sources and tools, executes steps and logs the process. Access control must be enforced at every data source and during retrieval. In supported SharePoint scenarios, Copilot Studio can take the signed-in user's identity and permissions into account (SharePoint as a knowledge source). For a custom Azure AI Search solution, by contrast, permission filtering must be explicitly designed and tested; automatic inheritance of all SharePoint permissions is not guaranteed.
5. Interface and audience
| Interface | Typical use | Key question |
|---|---|---|
| Teams / Microsoft 365 | internal employees | identity, licence, permissions |
| SharePoint | internal intranet | supported publishing route and authentication |
| Custom web app | internal or external portal | own authentication, API and operations |
| Public channel | customers/B2C | data protection, abuse protection and consumption-based costs |
An internal Microsoft 365 canvas is not automatically an external portal. Copilot Studio lists the supported publishing channels separately. For anonymous or external audiences, identity, consent, data release and licensing must be planned separately.
Data sources and data protection
Email, HR documents, meeting minutes and CRM data can contain particularly sensitive or confidential information. “Present in the tenant” does not mean “permitted for the agent”. Define allowlists of sources, test permissions with different roles and log only what operations and audit require.
The revised Swiss Federal Act on Data Protection (FADP) has applied since September 2023. For disclosure abroad, Articles 16 et seq. FADP and adequacy or appropriate safeguards in particular must be assessed; data subjects must be informed where the law requires it. The FDPIC guidance explains this assessment.
Data residency: no blanket Swiss guarantee
- Copilot Studio: Microsoft lists Switzerland North and Switzerland West for the Swiss Azure geography and generally keeps customer data in the chosen geography. However, the documentation names exceptions for global, external and Microsoft 365-powered services.
- Microsoft 365 Copilot: commitments depend on the tenant, the Product Terms and the Microsoft 365 workloads involved. Not every data type necessarily resides in the same place.
- Azure OpenAI / Foundry: regional Standard deployments process within the chosen geography; Global and Data Zone deployments can route differently. Models and tools are not available in every Swiss region.
- Replit: according to the Replit documentation, published apps run on Replit infrastructure. The primary hosting region is the US; Enterprise customers can request EU hosting. That is not Swiss data residency.
An architecture must therefore document the specific services, deployment types, connectors, logs and subprocessors.
Operations
A production knowledge agent needs quality measurement, source owners, automated updates, role reviews, an incident process and a way to give feedback. Savings can only be quantified after a measured baseline and a comparable period of operation.
Conclusion
Start with data and access rights, not with the model. Only once audience, sources, identity, residency and quality criteria are settled can the right platform be determined.
Further reading: Platform comparison
