OCR

What it means

OCR is optical character recognition: it turns a scan or a phone photo of a document into text and reads the fields out of it, from the company number and the amount to the date and the payment reference. Paper becomes a record the accounting software can take in.

It is not an electronic invoice. On the Peppol network and in electronic invoicing the structured data arrives straight from the supplier, whereas OCR only infers it from an image, and every inference carries an error rate.

OCR on its own only knows which characters are in the image. Which number is the amount payable and which the tax base is decided by a further layer, today often artificial intelligence or templates for a particular supplier, and that is where most errors arise.

How it works in practice

In an accounting firm OCR is the first step of document digitisation: the document arrives by e-mail or from a phone, the software pre-fills the header and a person confirms it. Robotic process automation takes the routine on from there.

The Accounting Act allows a paper record to be converted into electronic form and kept only that way, provided it stays complete, legible and verifiable. For document retention it is the scan that counts, not a neatly recognised header.

Recognition quality depends on the input. We advise clients to photograph the whole document, straight and in good light, to scan thermal receipts from eKasa at once because they fade, and to send long invoices as a PDF rather than a string of phone photos.

What to watch out for

The usual loss: OCR fails to tell a prepayment invoice from the final one and the same purchase is posted twice. VAT deducted twice surfaces in the cross-check of the control statements and is assessed back with interest, though a machine made the error.

Documents often carry personal data: names, addresses, account numbers or salaries. If a cloud service runs the OCR, a data processing agreement with its operator is required and you need to know where the data is stored; otherwise it is a GDPR breach, not a technical detail.