Why companies want a private ChatGPT
Public chatbots are useful, and your staff know it. They also send every prompt to an outside provider, on terms your legal team did not negotiate. Blocking them only moves the copying to personal phones.
A private assistant gives your teams the same experience with a different address. The chat window is Open WebUI, LibreChat or Dify. The model runs on your servers or on dedicated GPUs in Zurich, or in another region you choose. The conversation history sits in your own database, and the contract is with Pilae SA, a Swiss company.
Models hosted in your perimeter
We run open-weight models on hardware you choose: your own GPU servers, Pilae Cloud in the region you choose, or both. vLLM serves them when many people use the assistant at once; Ollama suits a pilot or a small team. Small models handle drafting and summaries quickly. Larger models handle reasoning over long documents. Each is a line in your contract, not a setting a vendor can change.
For air-gapped sites, the whole stack runs on premises with no outbound connection. Model updates arrive as planned changes that you approve.
The open-weight models we run
These are the model families we deploy most often. Each comes in several sizes, and the size decides the hardware: we match it to your GPUs and to how many people use the assistant at once. Licences differ by family and by size, so we check each one against your use before it goes into the contract.
| Family | Published by | Typical use |
|---|---|---|
| Llama | Meta | A general assistant for drafting, summaries and questions over your documents. |
| Mistral | Mistral AI, Paris | French, German and Italian as well as English, for multilingual teams and correspondence. |
| Qwen | Alibaba Cloud | Long documents, structured extraction and code, in many languages. |
| Gemma | Smaller models that answer quickly on modest hardware, for a pilot or a single team. | |
| DeepSeek | DeepSeek | Step-by-step reasoning over a problem, where a slower, more careful answer is worth it. |
| gpt-oss | OpenAI | Reasoning and tool use in the assistants and workflows you build with Dify and n8n. |
The weights are downloaded once, checked and served from your own machines. Whoever published a model never sees a prompt sent to it: it runs on your hardware, not on theirs.
Answers grounded in your documents
Retrieval-augmented generation connects the assistant to the places your knowledge already lives: file shares, Nextcloud, wikis and databases. The index sits in Qdrant or in Postgres, inside your perimeter. Each answer cites its sources, so a reader can check the passage behind it. Permissions come from your directory, so an intern does not get answers from the board papers.
Built for Swiss rules on outsourcing, secrecy and data
A private assistant does not make an organisation compliant, and no provider can do that for you. What the architecture can do is keep the facts on your side: where the data sits, who can reach it, and what the record shows. It is built to support the rules a Swiss compliance team asks about first.
- FINMA outsourcing. For banks and insurers, FINMA Circular 2018/3 asks for an inventory, due diligence, audit rights for you, your audit firm and FINMA, and a way out. Your contract names where the models and data run and grants those rights, and the whole stack can be handed back or moved. FINMA’s guidance 08/2024 on artificial intelligence adds governance, data and model risks; the model list per environment and the approval of every change support the governance it describes. Our FINMA outsourcing page covers the contract in detail.
- Professional secrecy. Article 321 of the Swiss Criminal Code punishes lawyers, doctors, notaries and their auxiliaries who reveal a secret entrusted to them. With the models in your perimeter, a client file pasted into the assistant goes to no model provider. On premises or air-gapped, the conversations never leave your machines, and every access by one of our engineers is one you approved and the audit log records.
- The Swiss data protection act. The Federal Data Protection and Information Commissioner states that the data protection act applies directly to AI. The act asks for transparency, security and care with disclosure abroad. The data stays in Switzerland unless you choose a region elsewhere, our data processing agreement covers the processing, and prompts, answers and sources are logged in your own database for access requests. See our nLPD page.
- The EU AI Act. Regulation (EU) 2024/1689 reaches Swiss organisations whose AI output is used in the EU, and sets its obligations by risk. An internal assistant for drafting and search is rarely high-risk; screening candidates or scoring credit is. For those uses, the logs, the listed model versions and the approved changes support the record-keeping and human oversight the act asks of a deployer. Which of your uses fall where is your decision, and we supply the facts it needs.
Operated, not just installed
The Pilae Agent watches the chat, model and retrieval stack, plans each update and waits for your approval. Monitoring probes every service every 60 seconds and alerts our engineers. Your data residency and security controls are written into the contract. To start, book a 30-minute private AI briefing with an engineer, or tell us about your project.