Data residency
Your files do not leave Europe.
| Hosting | European regions for every model we run |
| Standard | GDPR and the Swiss DSG, built in from the start |
| Data | Only what the job needs — retention by agreement |
| Logging | Traceable: which document was read and what was produced |
If a model does not meet these conditions we do not use it, however well it scores in tests.
What that means with American providers
What we run, row by row
Group A — Commercial models
| Model | Provider | Hosting / region | What we run it for |
|---|---|---|---|
| Claude | Anthropic, USA | European region | Long documents, structured extraction, analysis and reporting. Every answer keeps the pointer to the page or figure it came from. Our route when the result has to stand up to review. |
| ChatGPT | OpenAI, USA | European region | First drafts, summaries and day-to-day office work: e-mail, calendar, files, tenant replies, weekly reports. Whoever signs edits first. |
| Gemini | Google, USA | European region | Image and video analysis at volume: site photos, drone footage, defect reports, property images. We choose it when the existing stack points there. |
| Microsoft Copilot | Microsoft, USA | Inside the client's Microsoft 365 tenant | Work that already happens in Outlook, Teams, Word and SharePoint: research in existing files, workflows between e-mail and document. Existing permissions continue to apply. |
| Mistral | Mistral AI, France | France ⚑ | The direct route when data residency decides: general work at moderate cost, plus OCR that turns scanned and photographed documents into structured data. |
| Kimi | Open model from China | Self-hosted, on the client's infrastructure | For organisations that may not hand any data outside. Runs on your own infrastructure; nothing leaves the building. |
Group B — Open models for self-hosting
| Model | Provider | Hosting / region | What we run it for |
|---|---|---|---|
| GLM | Open model from China | Self-hosted, on the client's infrastructure | Our first open route. In our assessment one of the strongest open models in 2026. Open weights, no provider in the data path. |
| Qwen | Alibaba, China | Self-hosted, on the client's servers | The broad all-rounder for the everyday middle of the work: drafting, summarising, extracting, routine analysis. Open weights, Apache licence. |
| Llama | Meta, USA | Self-hosted, on the client's infrastructure | We use it mainly in marketing and PR. Meta's research on audience response gives an early indication of how a piece might land ⚑. |
Group C — Infrastructure
| Platform | Provider | Hosting / region | What we run it for |
|---|---|---|---|
| AWS Bedrock | Amazon Web Services, USA | AWS regions in Europe: Frankfurt, Zurich, Ireland ⚑ | Not a model but the place many models run. For clients whose document storage and specialist systems already sit in AWS and whose data should not change cloud. |
⚑ marks a statement taken from our existing material that is not backed by an external source.
Why origin rather than provider is given for Kimi and GLM, and what a provider's seat means for your data: US providers and data residency
Where a workflow runs
01 Microsoft Azure For clients already on Microsoft whose IT manages the environment itself.
02 AWS When document storage and specialist systems already sit in AWS and the data should not change cloud.
03 Google Cloud When the image analysis and the existing platform both point there.
04 What this is not Not a choice of model. Which model does the job is decided separately.
Bring your compliance position. We tell you what can run, where it runs, and what will not work.
Related entries: US providers and data residency · Using AI under Swiss data protection law