Turning paper into rows you can check

The job, and the error rate

A shoebox of receipts. A stack of delivery notes. Three years of handwritten job cards. Turning these into something searchable is one of the most satisfying things you can do with a phone camera, and it comes with an error rate you should know before you start.

Optical character recognition on clean printed text (a laser-printed invoice, photographed flat in decent light) is very good. Well above 99% of characters, which means a typical page has a few errors or none.

On a thermal till receipt that has faded, or a page photographed at an angle in a dark stockroom, accuracy falls sharply. Numbers suffer worst, because 3 and 8, 5 and 6, 1 and 7, 0 and O are exactly the pairs that degrade first.

On handwriting it depends enormously on the hand. Neat block capitals do reasonably well. Ordinary cursive is far worse, and one person’s habitual way of writing a 7 can produce a consistent, systematic error across three years of records.

Modern tools that combine OCR with a language model do better than old OCR, because context helps: a model reading “Total: 1,2SO” can work out what was meant. That same helpfulness is a hazard, which is the point of this lesson.

The specific danger

Old OCR gave you garbage that looked like garbage. T0ta1: 1,2SO is obviously wrong.

A model reading the same receipt gives you Total: 1,250: clean, plausible, and possibly not what the paper says. It has repaired the text using what usually follows, which is the same mechanism as everything else in this course. Sometimes the repair is right. Sometimes it turns 1,280 into 1,250 and there is nothing on the screen to indicate it.

This means the check has to be against the paper, not against the plausibility of the output.

How to do it so the checking is cheap

Photograph properly. Flat, no shadow, whole document in frame, in daylight if you can. Ninety seconds of care here removes most of the errors downstream.

Ask for structure, not prose. Request specific fields (date, supplier, total, tax, invoice number) as a list. Structure is easier to check than a paragraph, and easier to paste into a spreadsheet.

Insist on a blank rather than a guess. If a field is unreadable, write UNREADABLE. Do not estimate. Without this, illegible fields come back filled in.

Check totals against something external. The supplier statement, the bank feed, the card statement. This is the real safety net, and it catches errors regardless of where they came from.

Sample the rest. For a batch of 200 receipts, check 20 against the paper properly. If all 20 are right, your error rate is probably low. If two are wrong, that is roughly one in ten across the whole batch, and you have a decision to make about the other 180.

Free tools that do the job

• Your phone. Google Lens, Apple’s Live Text, and the scanning built into Google Drive and OneDrive all extract text at no cost.
• Tesseract, free and open source, runs on your own computer and never sends anything anywhere. Excellent on clean print, poor on handwriting.
• Your accounting software. Zoho, QuickBooks, Xero, Wave and others have receipt capture built in, usually included. This is nearly always the best option, because the extracted data lands directly where it needs to be with no second copy.

Paid document-processing services exist and are more accurate on messy input. They are worth it at thousands of documents a month and are hard to justify below that.

Keep the paper

Or at least keep the images. Most tax authorities accept digital copies, and most require you to keep something for a number of years: commonly five to seven, differing by country.

The practical reason is narrower than compliance. When a figure is queried in eighteen months, the photograph is the only way to settle it, and by then nobody remembers whether that receipt was faded.

Two jobs, not one

It helps to separate them, because they have different failure modes.

Extraction is turning marks on paper into characters. This is where photograph quality, faded thermal paper and handwriting decide your accuracy, and where the errors are mechanical.

Interpretation is deciding that the number next to the word TOTAL is the total, that 12/03 is a date, and which of the three numbers on the receipt is the tax. This is where a language model helps enormously and where it also guesses.

Keeping them separate tells you where to look when something is wrong. A total that is out by one digit is usually extraction. A total that has been read from the wrong line is interpretation, and it repeats across every receipt from that supplier, which makes it both more damaging and easier to spot once you know to sort your check by supplier.

What to do with the output

Straight into a spreadsheet or your accounting software, with two extra columns: the source image filename, and a blank column for checked.

The filename column is what lets you settle a query in eighteen months without hunting through a phone gallery. The checked column is what stops the sampling exercise turning into a vague memory that most of them were fine.