Accounting firms: one bot for receipts, one for client files

In an accounting firm, half of all client conversations come down to the same request: “could you send us your receipts?”. They arrive in a jumble, by email, as blurry photos, or in an envelope at the end of the quarter. And when a manager wants to know where their expenses go, someone at the firm opens the software, filters, exports, and answers two days later.

Bots change both moments, but not with the same tool. Some bots execute; others you ask questions.

The bot that executes: the client sends, the file fills up

The client takes a photo of the receipt and sends it on Telegram or WhatsApp, like any message. The bot reads the amount, VAT, date and supplier, picks the category, then creates the expense in the firm’s software with the photo attached. If the supplier already exists, it reuses it; otherwise it creates it. The client gets a confirmation within seconds.

This bot does not need to be a conversational AI. Once the receipt is read, it is a fixed sequence: read, check, write to the software, confirm. That is what makes it reliable and cheap. It can even write to two systems at once during a software migration without the client noticing: if the second one fails, the first stays up to date and the confirmation still goes out.

For the firm, the gain is simple: documents arrive as they happen, already sorted, instead of in one batch at closing.

Structured at the door: what makes everything else fast and cheap

The real job of the execution bot is not filing a photo. It turns a receipt into clean data: a recognised supplier, net amount and VAT kept apart, a date, a category from the chart of accounts, a client, and the photo attached to all of it. That work happens once, when the document arrives, while the client is still there to answer if something is missing.

Everything that comes after benefits:

• Questions cost almost nothing. “How much on fuel this month?” becomes a simple sum over rows already sorted. The AI does not reread a hundred photos to answer: it reads a few figures, which takes far less computing, so it costs far less.
• Answers are right. A figure computed from structured data is the same every time. An AI that had to guess a total from loose documents can get it wrong, and nobody would notice.
• A small model is enough. When data arrives clean, the model only has to understand the question. That is what lets everything run on a modest server, inside the firm.
• Anomalies stand out. A receipt with no vehicle, an unknown supplier, VAT that does not match the rate: the bot flags it on arrival, not three months later at closing.
• The file stays readable by every tool. Well-sorted data can be read by the accounting software, a spreadsheet, a dashboard or another bot, with no rework. Changing software no longer means re-entering everything.
• Duplicates do not get through. The same receipt sent twice has the same supplier, date and amount: it is caught at once.

The bot you question: the file answers in plain language

The second kind serves the firm itself. An accountant, a chartered accountant or a manager asks the question as they would ask a colleague: “how much did this client spend on fuel this month, compared with last year?”. The bot fetches the entries, does the maths, and flags what looks unusual.

Here you need an AI that understands the question. It connects to the accounting software through a connector (an MCP server, for instance) that lets it read, and only read, the data in the file. It does not replace the accountant’s judgement: it saves the half hour of exports and spreadsheets before judgement can start.

Confidentiality: a closed environment

An accounting file holds everything sensitive about a business: salaries, margins, suppliers, bank accounts. The firm’s professional secrecy does not stop at the door of an online AI service. That is why these bots are set up in a closed environment, depending on how strict the firm needs to be (the options are detailed in the four levels of private AI):

• on its own servers, in its offices or data centre;
• on a dedicated server in Switzerland or the European Union, rented in the firm’s name and used by no one else;
• in a private cloud where the AI model runs without the data leaving, or being used to train anything.

In all three cases, receipts, entries and the questions asked stay within a perimeter the firm controls. Access is set per person: a client sees only their file, an accountant only their mandates.

What about performance, without the big online services?

This is the objection we hear most: an AI hosted in-house would be worse. That was true two years ago. For this kind of work it no longer is, for four reasons.

1. Recent open models have caught up. To read a receipt, understand a question about figures or summarise a file, models you can install yourself now do the work only the big services could do before.
2. Computing power can be rented. A server with a dedicated graphics card rents by the month, in Switzerland or Europe. It serves only the firm, so it never waits in line.
3. The bot does not guess, it reads. Figures do not come from the model’s memory but from the accounting software, at the moment of the question (that is the idea behind RAG, explained in enterprise RAG: an AI that answers from your documents). The result depends on data quality far more than on model size.
4. Each task gets its own model. A small fast model reads receipts, a stronger one answers analysis questions. This routing keeps answers fast and the bill low, without sending anything outside.

What remains to check before launching is compatibility with the firm’s software: not every package opens its data in the same way. It is the first thing we look at, as in any project described in integrating AI in your business.

Where to start

The execution bot is the easiest to launch: it plugs into the existing software, changes nothing in clients’ habits, and its effect shows from the first month in documents received on time. The analysis bot comes next, on the same data, once it arrives clean and on time.

To price both, start from our AI at the right cost offer, or write to us with the name of your accounting software.