How accurate is AI receipt scanning?

Accurate enough that checking a scan is faster than typing one, and not accurate enough to leave unchecked. That is the honest answer, and it is more useful than a percentage, because accuracy on receipts is not one number — it varies enormously by field and by the condition of the paper.

Anyone quoting you a single accuracy figure is quoting a figure measured on their own test set, over receipts of their own choosing, counting whichever fields they chose to count. It tells you nothing about your pile.

Accuracy is per field, not per receipt

Break a receipt into its fields and the difficulty ranges over an order of magnitude.

Total is the easiest and most reliable. It’s printed large, usually labelled, usually at the bottom, and it can be cross-checked against arithmetic. If a scanner gets one thing right, it’s this.

Date is harder than it looks. Formats are ambiguous (03/04 is two different days depending on continent), receipts print times next to dates, and some slips carry two dates — transaction and settlement.

Vendor is easy to read and hard to normalise. The slip says “NWTS HOLDINGS LTD #4412”; you think of them as Northwest Tools. Getting the characters right is not the same as getting the vendor right.

Tax depends on the receipt’s own clarity. Multiple tax lines, tax-inclusive pricing, and slips that print a tax rate but not a tax amount all make this genuinely ambiguous rather than merely difficult.

Line items are the hardest by a wide margin. Abbreviated product names, quantities and unit prices aligned with spaces, and multi-line items that wrap. Expect materially lower reliability here than on the total, and expect it to vary by retailer.

If you need one mental model: fields that appear once, are labelled, and can be arithmetically verified are reliable. Fields that repeat, are unlabelled, or depend on convention are not.

Condition matters more than the model

The strongest predictor of whether a receipt scans correctly is not which software you use. It’s the state of the paper and the photo.

In descending order of damage: a faded thermal slip where the print has gone grey, a crumpled slip with creases crossing the total, glare from a flash directly on the reflective coating, a photo taken at a steep angle, and a receipt photographed against a same-coloured surface so the edge detection can’t find it.

Any of those turns a field that is normally near-perfect into a coin flip. Fix the input and you have improved accuracy more than switching tools ever will.

The two error types, and why one is much worse

Refusals — the scanner marks a field low-confidence or leaves it blank. Annoying, honest, and safe. You fill it in.

Confident mistakes — a plausible wrong value with no flag on it. A tip read as the total. A subtotal read as the total. A card’s last four digits read as a date. These are the errors that reach your books, because nothing about them looks wrong until someone reconciles the account.

Judge software by its second-type error rate, not its first. A tool that flags more and guesses less is better for accounting even though it feels worse to use.

What “verified” should mean in your workflow

Confidence scores are useful and frequently misunderstood. A confidence value is the model’s estimate of its own correctness, and it is well calibrated for text recognition and considerably less so for field selection — the stage where the confident mistakes come from.

So use confidence to order your review queue, not to decide whether to review at all:

• Anything below the threshold: check it.
• Anything where line items don’t sum to the subtotal, or subtotal plus tax doesn’t equal the total: check it, regardless of confidence. Arithmetic disagreement is a better error detector than any score.
• Anything from a vendor you’ve never posted before: check the category.
• Round numbers and unusually large amounts: check them, because these are the ones that cost you if wrong.

What accuracy is actually for

The point of scanning is not to eliminate human attention. It’s to move human attention from transcription to judgement.

Typing a receipt requires you to read six fields and type six fields, and your own error rate on that task is not zero — it’s just invisible, because nobody audits their own typing. Reviewing a scan requires you to read six fields and confirm six fields, which is faster, and it directs your attention at the values the software already flagged as shaky.

That’s the real gain, and it doesn’t depend on any accuracy claim. Even at a modest hit rate, review-and-fix beats read-and-retype on both speed and total error, as long as you actually review.

A reasonable expectation to hold

Totals and dates on decent photos of intact receipts: reliable enough to trust with spot checks. Vendor names: reliable characters, unreliable normalisation, so maintain your own vendor list. Line items: treat as a convenience, verify when they matter. Faded or crumpled paper: expect to key it manually and capture sooner next time.

Hold that expectation and receipt scanning is a straightforward time saving. Expect zero review and it will eventually cost you an afternoon of reconciliation to find one transposed total.