Skip to content

Product · · 3 min read

How AI reads invoices: PDFs, scans and phone photos

What happens to an invoice after you upload it to reko.ai: text extraction, amount, BIN, VAT and date recognition, confidence scoring and human review.

Key takeaways

  • reko.ai accepts invoices as PDFs, scans and photos — no per-supplier templates needed.
  • It extracts the invoice number, supplier, BIN, amount, VAT and date; the original is stored next to the payment.
  • The model's answer goes through strict validation: a field that isn't in the document stays empty instead of being made up.
  • Confident payment–invoice pairs are approved automatically; doubtful ones go to a person for review.
  • Recognition quality depends heavily on the source: an original PDF reads better than a scan, and a straight photo better than a cropped screenshot.

From file to data

Invoices arrive in every possible form: a PDF from the supplier's accounting system, a stamped scan, a phone photo, an email attachment. To an accountant they're the same document; to software they're several different problems. That's why recognition in reko.ai is built as a pipeline.

  1. Intake. Invoices are uploaded to the workspace — one at a time or in bulk. PDF, PNG, JPG and WEBP are supported.
  2. Text extraction. The text content is pulled out of the document for the model to work with.
  3. Field extraction. The AlemLLM language model finds the invoice number, supplier, BIN, amount, VAT and date in the text and returns them in a strict structured format.
  4. Validation. The model's answer is checked: the amount must be a number, the date a real date, the BIN twelve digits. If a field isn't found in the document, it stays empty.
  5. Linking. The recognized invoice becomes a candidate for matching against payments from the bank statement.

Why no templates

Classic recognition systems rely on coordinates: “the total is always in the bottom-right corner”. That works while you have five suppliers with identical forms. A language model finds fields by meaning, so a new counterparty with its own invoice layout needs no setup.

The flip side of that flexibility is the risk that the model “fills in” a value. That's why in reko.ai the model is explicitly forbidden to invent data, and everything it returns is validated before it's saved.

What a confidence score is

Every proposed payment–invoice pair gets a confidence score — a number from 0 to 1. If the counterparty's BIN, the amount and the period match, confidence is high. If the amount matches but the counterparty is uncertain, confidence drops.

The score decides where the pair goes next. On the Business plan, pairs with a confidence of 0.95 or higher are approved automatically, while doubtful ones land on a review screen that shows the amount, BIN and original document side by side. People spend time only where it's really needed.

How to help recognition

  • Upload the original PDF if you have one: it almost always beats a scan or a screenshot.
  • Photograph the whole document, in even light and without a strong tilt.
  • One file, one invoice. Several documents in one file make parsing harder.
  • Don't crop the party details or the total: those are exactly what reconciliation needs.

Where documents are stored

Originals are kept in encrypted S3-compatible storage in a data center in Kazakhstan. Files are served through short-lived links, so a document can't be opened from an accidentally forwarded URL a week later.

Frequently asked questions

Which invoice formats does reko.ai support?

PDF and images in PNG, JPG and WEBP — so scans and phone photos work too.

Do I need to set up templates for each supplier?

No. Fields are extracted by meaning rather than by position on the page, so a new supplier needs no setup.

What happens if the AI gets it wrong?

Doubtful pairs get a low confidence score and aren't approved automatically — a person decides, looking at the original document.

Where are uploaded invoices stored?

In encrypted S3-compatible storage in a data center in Kazakhstan. Access is granted only through short-lived links.

Close the month while competitors are still hunting for lost invoices.

14-day trial. No card required.