Reading invoices with AI (OCR): how it works

Reading invoices is the first step in automating accounting. But "reading" an invoice means more than turning a PDF image into text: you have to understand which number is the total, which is VAT and which lines are the items. This article explains how invoice reading and recognition work with OCR and AI, which documents you can send, where recognition still gets things wrong, and what it costs.

OCR vs AI – the difference

OCR (optical character recognition) turns a scanned or PDF image into text. AI goes further – it understands the meaning of that text: which number is the total, which is VAT, which lines are products. Template-based tools break when the layout changes; AI adapts to different suppliers' formats.

In practice the difference shows within the first month. A template-based tool has to be prepared for every new supplier – someone shows it where the number sits on that invoice and where the total is. With a hundred suppliers and five new ones every month, that work never ends. AI-based recognition needs no templates: an invoice from a supplier it has never seen is read the same way as one from a regular.

Scanning, recognition, digitisation – are they the same thing?

The words are used interchangeably, but they are different steps. Scanning turns paper into a file. Recognition pulls the data out of that file – the supplier, the amounts, the lines. Digitisation is the whole route: from receiving the invoice to an entry in the accounting system and an archive where you can still find the document a year later. When you compare tools, ask which of these steps each one actually does – more than one "scanning" product stops at a file in a folder.

What fields are recognized

• •supplier name and details (company number, VAT number)
• •invoice number and dates
• •individual product or service line items
• •net amounts, VAT rates and the total
• •currency and payment terms

The lines matter most. Header amounts are quick to retype; the time goes into an invoice with forty lines, each needing an expense account and a VAT code. If a tool recognises only the header, the hardest part of the work stays with you.

Which documents you can send

• •a PDF that arrived by email – you forward the message and the invoice is read with no further action from you
• •a scanned paper document
• •a receipt or invoice photographed on a phone – handy for employees who buy things on the road
• •one file holding several invoices – it is split into separate documents

A digital PDF is recognised most accurately, because the text is already in it. With a scan or a photo the result depends on quality: a flat, well-lit page reads without trouble; a crumpled, half-faded fuel receipt does not always.

The hard cases where AI wins

It is exactly the complex invoices where manual entry tends to go wrong and AI shows its advantage:

• •mixed-VAT invoices with several rates
• •multi-currency invoices
• •large invoices with dozens of lines
• •poor-quality or scanned PDFs
• •several documents in one file

A typical example is the monthly fuel invoice hiding two hundred separate fills, or a telecom invoice with three different VAT treatments on one page. Entering one of those by hand takes half an hour, and it is very easy to slip.

Where recognition still gets it wrong

• •quantities or corrections added by hand on a printed invoice
• •faded till receipts, and photos with shadows or cropped edges
• •a stamp or signature covering an amount
• •an invoice where the supplier's own arithmetic does not add up – the lines do not sum to the total

No system gets these right every time, so the question that matters is not whether it makes mistakes but whether you will see them. Ezura checks the line totals against the header and the VAT against the rate, and shows any mismatch on the invoice as a warning. The decision is yours: correct it, or leave it if that is how it should be.

Reading is only the start

After reading, the real automation begins: checking amounts and VAT, assigning dimensions and exporting to accounting. The data is not just recognized but classified the way your accounting needs it.

Accuracy and review

Automation does not mean blind trust: you review the prepared record before approving, and amounts are reconciled so the buckets match the invoice. That pairs speed with control.

What invoice recognition costs

Pricing is usually per document, not per user or licence. Ezura's standard rate is €0.25 per invoice, dropping to €0.10 at higher volumes. That covers the checks on amounts and VAT as well as the recognition itself. The fair comparison is not with zero but with the time the same invoice takes to enter by hand: at three minutes each, five hundred invoices a month is twenty-five working hours.

What to have ready before you try it

• •last month's invoices – as they arrive, not hand-picked
• •the list of expense accounts and VAT codes from your accounting system
• •the dimensions you split costs by – projects, departments, vehicles

You see the recognition result straight away. The rules – which supplier's invoices go to which expense account – build up gradually as you handle real documents.