OCR receipt scanner: what it reads reliably and what it guesses

An OCR receipt scanner photographs a receipt and extracts the fields you need to record it — supplier, date, total, tax, sometimes the individual lines. Interest in these tools has climbed steeply over the past year, partly because phone cameras and text recognition both got good enough, and partly because tax authorities in several countries now expect a digital image rather than a shoebox.

The technology does two separate jobs and they succeed at very different rates. Reading the characters is close to solved. Deciding which characters are the total, which are the VAT and which supplier issued it — that is the part that still gets things wrong, and it is the part that matters for your accounts.

What an OCR receipt scanner gets right and wrong

• Reliable: total amount, date, and the supplier name on a clean, flat, well-lit receipt from a major retailer.
• Usually reliable: currency, payment method, card last digits.
• Unreliable: tax breakdown where several rates appear on one receipt.
• Unreliable: line items on long receipts, especially with abbreviated product names.
• Poor: faded thermal paper, creased receipts, photographs at an angle, handwritten additions.
• Poor: foreign-language receipts and unusual layouts, where field detection has little to work from.
• Never automatic: which expense category it belongs to and whether it is a legitimate business cost. That is a judgement.

Thermal paper is the practical enemy. It fades in months in a wallet and faster in a hot car, and once contrast drops the recognition rate collapses. Whatever tool you choose, the highest-return habit is photographing the receipt on the day rather than at month end.

The accuracy claim to be sceptical about

Vendors quote figures like 99% accuracy, and the number is usually true of something narrower than it sounds — character recognition on a test set of clean receipts, rather than correct field extraction on the crumpled ones your team actually submits. The figure worth knowing is different: what proportion of receipts go through with no human correction at all, on your kind of paperwork.

Test it before committing. Take fifty real receipts of the sort you actually get, run them through, and count how many needed an edit. If a third need touching, the tool has moved effort rather than removed it — which may still be worth it, but is a different proposition from what was sold.

Fitting it into a process that works

1. Photograph at the point of purchase, not at the end of the month.
2. Require the image, and treat a claim without one as incomplete.
3. Let extraction populate the fields, then have a person confirm total, date and tax before posting.
4. Set a threshold below which a receipt is not scanned at all — under a small amount, manual entry is faster than correcting.
5. Keep the original image attached to the transaction, since that is usually what the tax rules require rather than the extracted data.
6. Check what your jurisdiction accepts as evidence, including how long images must be retained and whether the paper original can be destroyed.
7. Reconcile scanned receipts against the card statement, which catches both misreadings and missing claims.

Where the image has to live

The extraction is the interesting part and the storage is the part that matters legally. Rules on digital record keeping differ by country — some accept an image as the sole record, some require it to be legible for a set number of years, some still expect the original in specific circumstances. Check what applies to you rather than assuming a scan is sufficient, and treat retention periods as a legal question rather than a settings screen.

Ettex Books is where the resulting transaction lands: the expense recorded against the right account, with the receipt image kept attached to it, so the evidence and the entry stay together instead of the image sitting in a phone gallery.

There is no OCR in it. It does not read receipts, extract fields or categorise them automatically — if you want scanning, that comes from a dedicated tool or your phone, and what arrives here is the transaction and the image. Saying so plainly seems better than implying an extraction feature that does not exist, since that is precisely the sort of claim that wastes an evaluation.

Photograph receipts the day you get them, let extraction fill the fields but confirm the total and tax yourself, and check what your local rules say before binning the paper.