Almost every conversation about AI in accounting is about productivity: close faster, type less, get answers instantly. All true. But there is a question that comes first, and it gets skipped surprisingly often:
Where exactly does our accounting data sit the moment an AI reads it — and whose law applies there?
For a Swiss trust company that manages its clients' entire financial books, that is not an academic question. It's the question everything else hangs on.
The blind spot in the AI rush
Most AI assistants on the market are attractive front-ends over the APIs of a handful of large US providers. Every request — and the data inside it — travels to their cloud. For a newsletter draft, that's fine. For a client's receivables ledger, payroll or VAT return, it is not.
Because when the data moves, so does the legal order it lives under. Data on US infrastructure is subject to US law — including the CLOUD Act, which can compel access regardless of where the company itself is based. A Swiss fiduciary working that way has effectively handed over control of its clients' data. That sits badly with the profession's duty of confidentiality — and, increasingly, with data-protection law.
Sovereignty isn't a feature. It's the foundation.
So we didn't bolt the problem on at the end — we put it at the start. The guiding principle is Jurisdictional Integrity: the entire processing chain lives in one legal order, Switzerland. Not just the database. Every layer.
The stack — Swiss end to end:
- Data stays in the client's source system. The AI reads only (read-only) and never trains on it.
- Compute runs on our own GPU infrastructure in a Swiss data centre — no hyperscaler.
- The model is a self-hosted open-weight language model. It sends nothing outward.
- The application — client portal and admin backend — runs on the same sovereign foundation.
The effect is simple to state and hard to engineer: not a single byte of accounting data leaves Swiss legal territory. There is no "US hop" at which a foreign law could reach in.
Open-weight models — without the quality trade-off
The most common objection to sovereign, locally hosted AI is: "Sovereign, sure, but worse than the big cloud models." We didn't take that on faith — we measured it.
We ran several open-weight models on our own hardware against a leading cloud model, on exactly the tasks the accountant handles day to day: pick the right tool, set the correct parameters, invent no numbers. The result: the local model reached cloud parity. Full data sovereignty, with no perceptible loss in quality.
Sovereignty is therefore no longer a compromise you have to explain to a client. It's an advantage you can sell.
Trust is built in, not asserted
An accounting AI is only worth what you can trust its numbers to be. So traceability is at the centre:
- Sourced answers. Every figure comes with its source from the books — expandable, not asserted.
- Honest over guessing. If the data is missing, the AI says so plainly — rather than inventing a plausible but wrong number.
- Full audit trail. Every request is logged — fit for revision and compliance.
- Receipt upload. Drop in a PDF, get a structured booking proposal back.
The result
What came out is a 24/7 AI accountant: the trust company's own end-clients ask their numbers themselves — "how's my liquidity?", "who are my biggest suppliers?" — and get sourced answers, around the clock. The trust team is freed from routine queries and keeps full control through an admin backend. All of it on a foundation you can explain to a client in one sentence: your data stays in Switzerland, under Swiss law, on our own infrastructure.
Sovereign AI isn't a sacrifice. It's a competitive advantage.