Automation

Receipts straight into the books

Own project · 2 min read · Otto Boström

Receipts are a good example of a task everyone has and nobody wants to do. I built a workflow where a photo from the phone becomes a finished row in the household ledger, with amount, date, shop and each item on its own.

The problem

Typing in receipts by hand takes a few minutes each, so it often turns into a growing pile. By the time they are typed in, half of them have faded.

The solution

• Photograph the receipt from a simple page on the phone. The picture is cropped to the receipt itself before it is sent.
• An AI model reads the shop, date, total and each line.
• The tool checks the sums. If the lines do not add up to the total, it tries again with a stronger model.
• The result lands in a review step and goes into the ledger with one tap.

What I measured

• The total was read correctly on 42 of 42 test receipts.
• The cost was between 1 and 3 öre per receipt (an öre is a hundredth of a krona).
• The errors that did occur were on single lines: bottle deposits, repeated items and discounts. That is why the review step is still there.
• A cropped photo 800 pixels wide gave the same accuracy as the original picture, at a fraction of the cost.

Why the numbers matter

It is easy to say that AI can read receipts. What decides whether such a tool can be trusted is that someone has measured where it goes wrong and built around it. Here that means the total is checked automatically and a person approves the lines.

For your company

The same workflow works for delivery notes, supplier invoices, forms and registrations: anything that today is typed from paper or pdf into a system.