Back to blog

Automated expense tracking in a Telegram bot with AI and n8n

AI Assist case: automated expense tracking in a Telegram bot with AI

AI Assist case: automated expense tracking in a Telegram bot. A screenshot, voice or text from three users — and data lands in Excel by itself.

We launched an automated expense-tracking system for a client, built on a Telegram bot. Three employees log expenses in whatever way is easiest — and the data lands in a single Excel sheet with no manual entry.

What we did

A user sends an expense to the bot in one of three ways: a screenshot of the receipt, a voice message, or plain text. The bot recognises the amount, category, date and comment and adds a row to the sheet.

Under the hood is a combination of n8n for orchestrating the scenarios, OpenAI for recognising text, voice and images, and a Chroma vector database that stores context and helps classify expenses correctly. The result is written neatly into Excel.

What it costs

The economics turned out remarkably modest: 15 requests to the model cost the client about $0.07 via the API. The server is ours, so the client pays only a $4 monthly subscription. For a small business that is effectively the price of a coffee for fully automated bookkeeping.

Time to launch

We built a working bot in three days, and the client spent one more day testing it on real expenses. By day five the system was already running in production.

Why input data decides everything

The main takeaway of this project: the quality of input data determines 70–80% of the success of any AI system. If a user sends a blurry screenshot or dictates the amount unclearly, even the best LLM will get it wrong. That is why we worked separately with the client's team on simple rules: how to photograph a receipt, what to say in a voice message, which fields matter. That — not "model magic" — is what delivered a stable, accurate result.

A simple example: a clear photo of a receipt with a visible total is logged flawlessly, while a blurry "on the go" snapshot already needs a follow-up. So we advise treating input as part of the process, not a trifle — and we bake these rules into the team's internal guidelines. The approach also scales easily: add more users, new expense categories, or an export to another system.

Want to automate routine the same way? See our solutions or get in touch — we'll design a scenario for your business.

Frequently asked questions

How exactly does a user log an expense?

They send a receipt screenshot, a voice message, or text to Telegram. The bot recognises the amount, category, date, and comment and adds a row to Excel.

What does it cost per month?

About $4 for our server plus a tiny API fee — roughly $0.07 per 15 requests to the model. For a small business that is effectively the price of a coffee.

Why do you emphasise input data quality?

Because it drives 70–80% of any LLM's accuracy. A clear receipt is logged flawlessly; a blurry "on the go" snapshot causes errors. So we bake simple input rules into the team's guidelines.

Need an AI agent for your business?

Get a free audit for your niche and a solution demo.

Get a free audit