Local AI for confidential work: run it where your data is allowed to be

Most AI assistants work the same way: you send your document to a provider’s servers, a large general model reads it there, and the answer comes back. For a lot of work that is fine. For a fiduciary holding a client’s books, a law firm with case files, an HR team with payroll or a clinic with patient records, it is the part that stops the project.

We build the other kind: specialist AI that runs where the data is allowed to be.

The problem isn’t the AI, it’s where the data goes

Confidentiality is part of the job in many professions. In Switzerland, professional secrecy for lawyers, notaries, auditors, doctors and other health professions is set out in Article 321 of the Swiss Criminal Code, and the revised Federal Act on Data Protection has applied to personal data since 1 September 2023. In the EU, the GDPR does the same for personal data. Fiduciaries and accountants are bound by their contracts and their professional rules on top.

None of this bans AI in the cloud. But every document sent to an outside provider means more for the organisation to settle: a data processing agreement, a look at where the servers are and who can access them, sometimes a transfer abroad, and often a conversation with the client. For a firm with no legal or IT department, that is usually where it ends.

Where a specialist can run

So we deploy where the organisation’s rules say the data may be:

• On your own servers. One server in the office or your data centre serves the whole team over the local network.
• On one device, offline, for the most confidential work and for places with no network. Our note taker, Kimiko, works this way: it transcribes on the Mac or Windows PC it’s installed on.

Wherever it runs, documents are read where they are stored, and they are never used to train models for anyone else. Why this has only recently become possible is in What is sovereign AI?

A specialist, not a generalist

A general chatbot knows a little about everything, including a version of tax law from whenever it was trained. A professional needs the rule that applies on the date in question, and needs to see where it came from.

So we build specialists, and we build each one the same way:

• Train it for one profession, on cases written and checked for that profession in that country.
• Ground it in the sources. It looks things up before it answers, in the official texts and in the organisation’s own documents, and cites them.
• Give it tools for the exact parts. Arithmetic, rates and posting lines are computed by code, not guessed by the model.
• Test it before it ships, on cases it has never seen, and again before every update goes live.
• Let it say “needs review”. When it isn’t sure, it hands the case to a person and says why.

What a specialist doesn’t replace

Keeping the data where it belongs doesn’t replace professional judgement, and no specialist makes every answer right. Ours propose, a person confirms, and the organisation’s own backups and access rules still apply. We think that’s the honest way to bring AI into work that carries a signature.

Where we apply it

Our specialists serve offices and industry, public administration, and defence and security forces; What is sovereign AI? lists who each line is for. How an AI specialist learns a country’s accounting rules shows the method on one profession. And Kimiko takes AI meeting notes, transcribed on your own computer, with no bot joining the call.

If your organisation has data that can’t go just anywhere, write to us.