AI Models & Data Protection
SKH strictly distinguishes two model categories:
- SKH Standard – Swiss-hosted, data-protected, no API keys to manage.
- BYOK (Bring Your Own Key) – Frontier models from external providers, with your own API key. Data leaves Switzerland/EU.
What your members actually see depends on your organisation's Model Policy.
SKH Standard – Swiss data residency
All SKH Standard models run in Switzerland. Requests, responses and indexed content never leave Switzerland or the EU. FADP- and GDPR-compliant by default.
Available models
| Model | Description | Chat attachments |
|---|---|---|
| DeepSeek V3 | Strong reasoning model, ideal for code, analysis, multi-step thinking. | Text |
| Kimi K2.5 | Multimodal generalist. Reads images, PDFs, audio and video. Default for file uploads. | Images, video, PDF, audio |
| Mistral Medium 2505 | Fast, multilingual model with great price/performance. | Text |
Every SKH plan includes all Standard models; there's no extra token pricing. You pay for the plan, not the tokens.
Which model when?
- Summarising contract clauses, explaining technical concepts, code review → DeepSeek V3.
- PDFs with images, screenshots, video summaries → Kimi K2.5.
- Quick short answers, multilingual content → Mistral Medium 2505.
BYOK – Bring Your Own Key
If your organisation needs frontier models from other providers, the owner can enable BYOK via Model Policy.
Available providers
| Provider | Notes |
|---|---|
| OpenAI | GPT family |
| Anthropic | Claude family |
| Google Gemini | Gemini models |
| Mistral AI | Direct from Mistral |
| Azure OpenAI | Azure deployment, GDPR-fit, but not necessarily Swiss region |
| Azure DeepSeek | DeepSeek on Azure |
| Custom OpenAI-compatible | Your own endpoints (e.g. self-hosted models) |
A member can create any number of model profiles and pick them individually in chat.
Data-protection implications
As soon as a BYOK model is active:
- Requests, including the cited internal documents, are sent to the provider.
- Logging, caching and data residency follow that provider's terms.
- The chat displays an amber strip at the top and a flag next to the model name.
- On first switch per provider SKH shows a warning. You must explicitly confirm.
- The switch is recorded in the Audit tab of the Model Policy – with member, provider, profile and chat reference.
Creating a model profile
- Settings → Model Profiles.
- Add profile.
- Pick a provider, enter a model ID (e.g.
gpt-4o,claude-3-opus-20240229,gemini-1.5-pro). - Paste your API key. The key is stored encrypted.
- For Azure providers also: endpoint, API version, deployment name.
- For Custom OpenAI-compatible: enter a base URL.
- Optionally mark the profile as default.
Model Policy at organisation level
Owners control under Settings → Model Policy which BYOK providers are allowed at all. Providers not on the list don't even appear in the profile editor or model selector.
The Audit tab on the Model Policy page lists every BYOK switch in the last 30 days – including member, time, provider and chat reference.
LLM Scope vs Model Policy
These are two orthogonal concepts:
- LLM Scope controls which modes the chat offers (internal search, web, MCP, hybrid, deep research). See Search & Research Modes.
- Model Policy controls which models are selectable.
So an organisation can be:
- Maximally restrictive: Internal only + no BYOK = only SKH Standard models, only internal search.
- Maximally permissive: Full access + all BYOK providers allowed = full flexibility but maximum outbound data flow.
FAQ
Are my data used to train models? No. For SKH Standard models, training opt-out is contractual. For BYOK models the provider's terms apply; most enterprise tiers offer opt-out.
Can I host my own model? Yes, via Custom OpenAI-compatible. Provide a base URL that follows the OpenAI Chat Completions API (e.g. self-hosted vLLM or Ollama with the OpenAI bridge).
Is the model picked per chat or per message? Per message. The selector value applies to the next question. Mid-chat switches are allowed; SKH documents them as a visible divider in the conversation.
Can I enable BYOK per member? BYOK is currently enabled per provider at the organisation level. Profiles themselves are per member – everyone brings their own key.