Invoice OCR is often sold as the end of data entry. It is not, and the gap between that promise and what the technology actually does is where most disappointing implementations live.
What OCR is genuinely good at
Optical character recognition, paired with a bit of layout intelligence, is reliable at reading recurring documents from known suppliers. When the same utility company sends the same invoice layout every month, the system learns where the invoice number, date, net amount and tax sit, and extracts them accurately.
For a payables team processing a few hundred such invoices a month, that is a real reduction in keystrokes.
What it is not good at
- First-time suppliers. There is no learned layout, so extraction is a guess.
- Judgement. OCR can read a line item. It cannot decide which cost centre it belongs to, or whether the charge was authorised.
- Poor scans. A photographed invoice at an angle, in bad light, will produce plausible-looking wrong numbers — which are more dangerous than obvious failures.
- Tax treatment. Reading a tax amount is not the same as knowing whether it is recoverable.
The word that matters: drafts
The right mental model is that OCR drafts a journal. It does not post one. The draft arrives with the fields it is confident about already filled, and a human confirms, corrects and codes it.
This distinction is not timidity about automation. It is what keeps the audit trail meaningful. A posting that no person ever looked at is a posting no person can explain.
Where approval fits
OCR and approval workflow solve different problems and should not be collapsed into one. Extraction reduces typing. Approval establishes authority. An invoice can be perfectly extracted and still be one nobody agreed to pay.
In a sound setup the sequence is: the document is read, a draft journal is created, a person reviews and codes it, an approver signs off, and only then does anything post to the ledger.
Measuring it honestly
The useful metric is not “percentage of invoices automated.” It is the proportion of drafted invoices that a reviewer accepts without correction, tracked by supplier. That number tells you where the technology is genuinely earning its place, and where someone is quietly re-typing everything the system got wrong.
Automate the typing. Keep the judgement.