How we use AI
SoloDesk is bookkeeping and admin software for small Swiss businesses and their accountants. One part of it uses AI: when you upload a document, a model reads it, so you do not have to type it in.
The +AI label is granted by Swiss Made Software to Swiss companies that already meet the swiss made criteria and use artificial intelligence in their products, and that commit to describing how that AI is built, sourced, and operated.
In SoloDesk, AI supports one feature: reading uploaded documents. Everything else in SoloDesk works without a model call. The model itself is a third-party service from Google; the software around it is our own.
This page describes that one step: where it runs, what it does, what leaves our infrastructure and what does not.
Where SoloDesk runs
SoloDesk is a browser-based service. There is nothing to install and no self-hosted version: the application, the database and every document you upload are hosted on Google Cloud in Zurich.
Reading documents happens in a separate service of ours, in the same place, and it only ever runs on the server. Your browser and the SoloDesk mobile app never contact a model.
The model itself is not ours and does not run on our infrastructure. For AI document reading, our service calls Google Cloud Vertex AI through its global endpoint, which may involve processing outside Switzerland.
What the AI does, and where it stops
When you upload a document, the model reads it: one call decides what kind of document it is, and, where there are details to read out, a second call reads them — a supplier invoice, a receipt, the transactions in a bank statement PDF. No document is read more than twice.
That is where the model's part ends. Everything after it is ordinary software with no model in it. Matching a document to a contact is a database lookup. Assigning an expense to an account uses a table we maintain. Reading a bank file in the standard Swiss format is a plain parser. Those steps are deterministic: the same input gives the same result, and we can name the rule that produced it.
When the model is not confident about what it read, SoloDesk does not apply the result: it is surfaced for your review, and you decide. Results the model is confident about are applied, and the original document stays attached to them, so you can check the two against each other.
Reading documents is part of the service rather than something you switch off, and it runs only on documents you upload.
The technology we use
One provider, one model: a Google Gemini model, used through Google Cloud Vertex AI. There is no second provider, no model running on our own hardware and nothing running on your device.
Each reading task has its own instructions and a fixed schema the answer has to fit, so what comes back is labelled fields rather than free text we then have to interpret. We write, version and maintain those instructions and schemas ourselves.
Where the model comes from
The model is a third-party foundation model, built, trained and operated by Google. We did not build it, we do not host it, and we have no model of our own.
What we do build is everything around it: the instructions, the schemas, the handling of uncertain results and the bookkeeping logic that turns an answer into an entry you can check. That part is our own in-house development. No open-source or self-hosted model is involved.
Google Cloud Vertex AI is listed as a sub-processor in our privacy policy.
What we send, and what we keep
Each time a document is read, two things go to Google:
- the file as you uploaded it — a PDF or a photo, in full, not an extract;
- your company name, so the model knows whose books it is reading.
That is what the model needs in order to read the document, and it is the only thing that leaves our infrastructure for this step. Our privacy policy names the providers we work with and explains how we handle those based outside Switzerland.
What Google does with it: under our Vertex AI configuration (zero data retention), Google does not retain your content after returning the result, and, per its enterprise terms, does not use your content to train its AI models. Your content is processed solely to deliver the result back to you.
What we keep is on the document record: the details read out of the document — counterparty name and address, amounts, dates, the category assigned — and the text of the document itself.
What we do not keep: our record of each call holds only technical facts — which task ran, which model, how long it took, whether it succeeded. When a call fails, the error text is kept for debugging. Beyond that, neither the instructions we sent nor the model's answer are stored there.
Matching to contacts, spotting duplicates and assigning expense accounts all happen inside our own infrastructure. Those steps are never sent to Google.
Fine-tuning and adaptation
We do not fine-tune the model, and nothing you upload is used to train one on our side. The model does not adapt to your business over time: each document is read on its own, with no memory of the documents before it.
What does change is the software around the model — the instructions and schemas we maintain, and which model version we call. Those are versioned changes we make deliberately and release to everyone, not something the system teaches itself.
The instructions and schemas are written for Swiss business documents. There is nothing for you to configure and nothing for you to train.