Story
Chat with your accounting app on WhatsApp
How FIAR360 turned WhatsApp into a full interface to its fiscal app. An AI agent on a dedicated Mac mini that runs local models for simple tasks, calls the app's functions directly and sends reports back in the chat.
Updated 2026-09-15
FIAR360 is a Romanian app that automates tax and bookkeeping for freelancers, sole traders, short-term rental hosts and rideshare drivers. It issues e-invoices, keeps the mandatory registers and prepares tax returns.
Its users are not accountants. They drive for Uber, rent out an apartment or run a small practice. They don't want to learn another app. They already have WhatsApp open all day.
So FIAR360 put its whole app behind a WhatsApp chat, with an AI agent running on a dedicated Mac mini.
From sending receipts to running the app
FIAR360 already let users send a photo of a receipt on WhatsApp. The app read the supplier, the amount, the date and the VAT, and filed it as an expense waiting for the user's confirmation.
Useful, but one-way: you could send documents, not ask questions or get anything done.
The new agent turns the same chat into the app itself. Users write the way they would to an accountant:
- How much tax do I owe so far this year?
- Invoice the client from last week again, same amount.
- How many guest registrations did I have in July, and what was the city tax?
- Send me a report of my expenses for the quarter.
The agent understands the request, decides what to do, calls the right functions inside FIAR360 and replies in the chat, with the report attached when one is needed.
How it works
The agent sits between WhatsApp and the FIAR360 application.
One message
A question, a photo of a receipt, or a document.
On the Mac: a small local model sorts it
What kind of request this is, and what it needs. No cost per message.
Then one of two
Routine, most messages
Answered on the Mac itself.
Needs judgement
A larger cloud model, for reports and unusual cases.
The agent calls FIAR360's own functions
Reads the fiscal situation, creates records, issues documents. No screens scraped.
The reply goes back to WhatsApp
A short answer, a figure, or a finished report.
- 1A message arrives.Text, a photo of a receipt, or a document.
- 2A local model sorts it.Small language models running on the Mac handle the simple work: what kind of request this is, which data it needs, whether it is a document or a question.
- 3A larger model reasons when needed.Requests that need judgement, such as putting together a report or deciding how to categorise an unusual expense, go to a larger cloud model.
- 4The agent acts inside the app.It calls FIAR360's own functions to read the fiscal situation, create records, issue documents or build reports. It does not scrape screens or copy data by hand.
- 5The reply goes back to WhatsApp.A short answer, a figure, or a finished report.
Why local models for the simple tasks
Most messages are routine: a receipt, a balance check, a status question. Sending every one of them to a large cloud model would be slow and expensive.
Small models on Apple silicon answer these in the time it takes to read the message, at no cost per request. The larger model is kept for the minority of requests that need real reasoning.
Why a Mac mini
- It runs local models well. Apple silicon shares memory between the CPU and GPU, so a Mac mini can hold and run language models that would need a dedicated graphics card elsewhere.
- It is always on. The agent answers at any hour, without a laptop that sleeps or a desktop someone restarts.
- It is dedicated. The machine does one job. Nobody else's workloads, no shared host.
- It stays in the EU. The Mac runs in MacDuty's racks in Romania.
What this means if you are building an agent for your app
- Meet users where they already are. For many small businesses that is WhatsApp, not a dashboard.
- Let the agent call your app, not imitate a user. Direct functions are faster and far more reliable than clicking through screens.
- Split the work by difficulty. Local models for routing and routine requests, a larger model for reasoning. You pay for intelligence only where it matters.
- Keep the user in control of what counts. Anything that changes the books should be easy to review and confirm.
The FIAR360 agent runs on a dedicated Mac mini at MacDuty.
Next