Private AI: ChatGPT-style AI without your data leaving the EU
Want to use AI for client files, contracts or patient information? How your own language model on a European server works, what it costs and when it's worth it.
An accountant who wants to summarise annual accounts. A lawyer who wants to search contracts. A care provider who wants to write reports faster. They all want to use AI, but they run into the same question: is this information allowed in ChatGPT?
Often the honest answer is: better not. This article explains the alternative.
Where does your data go?
When you paste text into a well-known AI service, it is processed on servers run by the company behind that service. For the big providers, those are US companies. With a business plan they usually don't use your data to train their models, and many providers now offer storage in Europe. But you remain dependent on their terms, and for sensitive data that's a step too far for many organisations.
On top of that, for personal data the GDPR requires you to be able to explain where it is processed and by whom. The fewer parties, the simpler that is.
What is Private AI?
With Private AI a language model runs on a server dedicated to your organisation, in a European data centre. Your team works in a chat environment similar to ChatGPT: asking questions, drafting text, summarising documents. The difference is that the data never leaves that environment.
This is possible thanks to open-source language models. Companies such as Mistral (France), Meta and Alibaba publish models that anyone may run on their own server. Their quality has risen sharply in recent years.
What can you do with it?
- Summarise and write. Summarise long documents, draft emails and letters, write up reports.
- Ask questions about your own documents. You connect manuals, files or a knowledge base. The assistant looks up the answer and cites the source. This is also known as RAG (retrieval-augmented generation).
- Translate and rewrite. In your organisation's tone and terminology.
- Connect it to your processes. The same AI can later become part of automations, for example for processing incoming mail.
Honestly: what are the limitations?
Open-source models are excellent at summarising, writing and answering questions about documents. For the very hardest reasoning tasks, such as complex analyses across dozens of pages, the largest commercial models are sometimes still better.
A dedicated server also costs money, even when it isn't being used. For a team that uses AI occasionally for non-sensitive text, a business plan with a well-known service is often cheaper.
When is Private AI worth it?
Private AI is especially interesting if:
- you regularly work with personal data or confidential information, such as files, contracts or medical data;
- your clients or regulators ask where their data is kept;
- you want to use AI on a large collection of your own documents;
- you don't want to depend on the prices and terms of a single US provider.
What does it cost?
With us a Private AI environment starts from € 2,950 for setup and from € 495 per month for hosting, management and updates. The monthly price depends mainly on the size of the model and the number of users. See the Private AI page for all details.
Getting started yourself?
Not sure whether you need Private AI? Start with clear agreements on which information may and may not go into AI tools. Our free AI prompt guide includes a template for such usage rules. And if you want to know where AI saves your business the most time, take the free AI scan.