Tax documents filed as soon as they arrive
The starting point
Two tax advisers facing a very busy season
Frontalien is a Geneva firm specialising in cross-border commuters' taxes: people who live in France, work in Switzerland, and therefore deal with two administrations. The firm has two people.
The mandates don't spread over the year: they all arrive within a few weeks. The difficulty isn't the annual volume, it's how concentrated it is.
What it cost
Repeated emails to gather the documents, then retyping everything
A simple return needs five to ten documents, a full one fifteen to fifty. The client often provides only some of them: each certificate, each supporting document is obtained over many email exchanges.
Once the documents were gathered, each one was read line by line, and every amount re-entered by hand into the final return.
The solution
The client uploads, the AI reads, the tax adviser decides
Each client has their own space, with a list of documents that adapts to their situation (children, property, insurance). Every document they upload is filed automatically, and the system immediately flags what's missing.
The AI then picks out the amounts specific to each type of document, converted to francs if they're in euros. The tax adviser sees the document on one side, what the AI read from it on the other, and validates, corrects or discards: nothing is saved before this step.
Before / after
Avant
- Documents requested email after email
- Omissions discovered along the way
- Every amount re-entered by hand
Après
- A list of documents specific to each situation
- What's missing, flagged right away
- Every amount read by the AI, checked by the tax adviser
Key decision
Why do the documents never leave Switzerland?
A salary certificate, a pension statement, a divorce ruling: these aren't documents you send to a public AI service. The model that reads them runs in Switzerland, and the server can only reach a few authorised addresses.
The firm and we reached this conclusion at the same time: its clients want their data to stay in Switzerland.
Key decision
When the AI doesn't recognise a document, it says so
The first versions always classified, even when wrong, with no way to correct it. Today, an unrecognised document is marked as such, and the tax adviser picks the right type in one click.
Their corrections don't train the model, and that's deliberate: with tax data, a system that's always wrong in the same way, and that a human catches, is better than a system that changes without warning.
Key learning
Not showing the client an error they can't correct
At first, clients saw in their space the type the AI had given each document. When it was wrong, they could do nothing about it, and they worried: was their document lost, ignored?
Now they only see their file names. The type only appears once the tax adviser has confirmed it.
What comes next
The last manual entry, on the tax website, is the next to go
After validation, the tax adviser still transfers the data by hand to the Geneva administration's website. From early 2027, the tool will fill in this form itself from what the tax adviser has validated.
The result
- [The time a file took, and the time it takes now: to be provided by the firm]
Do your teams spend their time extracting data from documents? Let's talk.
We start with a conversation, not a quote.
Discuss a similar project