How an AI specialist learns a country’s accounting rules

Our accounting and tax specialist is designed to read invoices and receipts, propose a booking and its VAT code, and answer accounting questions, quoting the law in force. Accounting isn’t the same in every country, so it is trained country by country. This is how it learns one, with Switzerland as the example.

Why a general chatbot isn’t enough

Every country’s bookkeeping looks simple until you hit the details. In Switzerland, for instance:

• VAT rates depend on the date. Since 1 January 2024 the rates are 8.1% standard, 2.6% reduced and 3.8% for accommodation; before that they were 7.7%, 2.5% and 3.7%. What matters is the date of the supply, not the date of the invoice or the booking, so a delivery made in December 2023 but invoiced or booked in January 2024 still takes the old rate.
• The accounts are specific. Input tax goes to 1170 or 1171 on the Swiss SME chart of accounts, VAT owed to 2200, and each amount ends up in a box on the tax administration’s VAT return.
• There are several languages. The same firm can receive invoices in German, French and Italian, and the law exists in all three.

Every other country has its own version of each of these. A general model trained on a snapshot of the internet can mix rates from different years and countries, and rarely says where an answer comes from. An accountant needs the rule in force in their country on the transaction date, and the source.

The law, from the source

For each country, the specialist reads the law from official sources: the accounting and VAT law and the tax authority’s publications. Each text is stored with the date it applies from, so it can quote the version in force on the day of the transaction.

Tools for the parts that must be exact

We don’t ask the model to do arithmetic. The specialist calls tools:

• a VAT engine that picks the rate by the date of supply, computes and rounds the amounts the way the tax authority does, and converts foreign-currency invoices at the official rate;
• a search over the legal texts, and over the company’s own manuals, contracts, minutes and notes, scanned pages included, in one private index per company;
• a posting writer: the model proposes the debit and credit accounts and the VAT code, and code writes the actual posting lines.

The tools are checked against worked examples published by the tax authority itself.

Books to keep

To train an accountant you need books to keep. For each country we build practice companies with the invoices, receipts and other paperwork a real business there produces, across industries and VAT situations, and for each document we write the booking a good local accountant would make, with the VAT code and the source behind it. Every case is checked before the model learns from it, and no client data is ever used.

A bar set in advance

Before it’s released in a country, the specialist is measured on cases it has never seen, in each of that country’s languages, against a bar we fixed before training began: how many routine proposals it must get right, how much it must handle on its own, that it never cites a source that doesn’t exist or doesn’t apply, and that no language falls behind the others.

Every update after release has to clear the same bar again before it goes live.

Two rules we don’t bend

• The accountant decides. The specialist proposes; a person confirms every booking and enters it in the firm’s own accounting software.
• ”Needs review” beats a guess. When a case is unclear, the specialist flags it for the accountant with the reason.

Working with it

It runs where your rules require: on your own servers or offline on one device, so client books never leave the firm. Why that matters is in AI for confidential work. If you run an accounting firm or keep a company’s books, write to us and tell us which country you work in.