Invoices that post themselves
You drop a pile of PDFs and the AI reads them, extracts everything, reconciles them with the accounting and, if a total doesn't match, deletes the invoice itself. You just review.
Entering invoices by hand, one by one
Each supplier invoice meant opening the PDF, copying the supplier, the tax ID, the lines and the taxes, and entering it into the accounting. Slow and error-prone: sometimes the supplier's tax ID got confused with the company's, the wrong email was picked, or an income-tax withholding went undetected and the total didn't match. Multiply that by hundreds of invoices a month.
A system that reads, reconciles and self-corrects
Now the accountant drags a batch of PDFs onto a screen and that's it. The AI reads each invoice, pulls all the data, finds the right supplier even if the name comes differently, splits the taxes line by line and creates the invoice in the accounting. Then it checks that the total the accounting calculates matches the PDF to the cent; if it doesn't, it deletes the invoice automatically so a bad figure never gets in. The human only reviews what the AI has already done.
How it works, step by step
- 1
You drop the PDFs
You drag all the invoices at once onto a screen.
- 2
The AI reads them
It extracts supplier, tax ID, lines, taxes and ledger account from each one.
- 3
It reconciles and creates
It matches the supplier, splits the taxes line by line and creates the invoice in the accounting.
- 4
It self-corrects
It verifies the total to the cent; if it doesn't match, it deletes the invoice itself. You just review.
Under the hood
For the curious: the real engineering behind it.
- OCR (Mistral) + AI extraction (Gemini) in strict JSON, with taxes per line (multiple VAT rates, income-tax withholding, foreign currency)
- Cascading supplier matching: tax ID → email → phone → fuzzy name match (85%)
- Creation in the accounting (Holded) via API, total verification and automatic deletion if it doesn't match
- Deduplication by hash and business key so the same invoice never gets entered twice
- An 8-stage pipeline with visual progress; tested with dozens of varied real invoices
- So solid it was reused as the core of another client's invoice pipeline (the data-platform one)
The result
| What we measured | Before | Now |
|---|---|---|
| Time per invoice | 5-10 min | seconds (just review) |
| Errors that get into the system | common | 0 (deleted if it doesn't match) |
| Taxes per line | by hand | automatic (multiple VAT + withholding) |
| The accountant's work | typing | reviewing and approving |
The wow
The system doesn't even trust itself: after creating the invoice, it checks that the total matches the PDF to the cent and, if it doesn't, deletes it automatically. It's impossible for a wrong number to get in. The accountant went from typing invoices to reviewing what the AI had already done.
Other cases
Want something like this in your business?
If someone on your team enters invoices by hand, this can be built over your accounting.
Tell us your case and we'll tell you what can be built. The first analysis is free, with no commitment.