Why receipt scanning apps still get receipts wrong

If you have ever photographed a receipt, watched the app think for a while, and then had to correct the total, the date and the category anyway, you already know the problem with most receipt scanning apps. They are wrong just often enough that you stop trusting them — and once you check everything yourself, the time saving disappears.

While building Costrix's capture we read through around 35 public sources of employee feedback: App Store and Google Play reviews, Trustpilot, Capterra, GetApp UK and vendor community forums. The single most common complaint about expense apps was not slow reimbursement or clunky menus. It was that the scan read the receipt wrong and the person still had to fix it. This article lists those failure modes, explains why they happen in plain language, and sets out what good receipt capture looks like — for anyone who wants to scan receipts for expenses in the UK or Ireland without doing the work twice.

The ways receipt scanning goes wrong

The failures are remarkably consistent across products, which suggests the problem is the approach, not any one vendor.

Faded thermal paper. Most receipts are thermal prints. They fade in a wallet, crease in a pocket and grey out in sunlight. One Capterra reviewer of Zoho Expense wrote that it "misreads amounts or vendor names, especially on faded bills".

The category is almost always wrong. A poster on SAP Concur's own community forum reported that ExpenseIt got the category "almost always wrong" and the currency "almost always wrong". A Trustpilot reviewer of Concur wrote in August 2026: "Every single one of the 'Automatic' (AI) receipt or invoice reads miscategorized things completely."

The date shifts by a day. An App Store reviewer of Spendesk reported that "the app sets the date of the expense a day behind the date on the receipt… every time". Concur forum users described meals landing on the wrong date about half the time, which tripped per-diem limits.

Cash mistaken for card. Another Trustpilot reviewer of Concur, in June 2026, complained the AI makes "frequent mistakes (e.g. not categorizing correctly, mistaking cash for credit card)". For the employee, that is the difference between being reimbursed and not.

Matching that misses half the time. A Software Advice reviewer of Payhawk estimated the "receipt match is not very good (probably get a 50% match rate)". A Pleo reviewer on Trustpilot described 120-dollar bills matched to 20-dollar bills.

"Quicker to type it in." A Trustpilot reviewer of Concur put it plainly in July 2026: "The app is so slow at recognising receipts it's quicker to input the data by hand."

Forced crops and vanishing uploads. Payhawk's UK App Store reviews describe a forced crop on e-receipts where it was "genuinely impossible to select the bottom half of the image". Concur forum threads describe ExpenseIt uploads getting stuck and the app "deleting the receipt altogether". A GetApp UK reviewer noted of Dext that "it isn't always clear whether an upload has been submitted ok."

Add it up: the amount, date, category, currency and payment method can each be wrong, the photo can be cropped badly or lost, and the person may not even know whether the upload went through.

Why this happens

None of this needs a technical background to understand — receipt OCR accuracy is limited by physics and by design choices.

Thermal print is genuinely hard to read. Low contrast, uneven fading and creases mean the characters the reader sees are not quite the characters that were printed. A 3 becomes an 8; a decimal point vanishes.

Background clutter confuses edge detection. A pale receipt on a wooden desk, marble counter or car seat is a rectangle against texture and shadow. If the app guesses the edges wrongly and crops silently, part of the receipt is gone before anything is read.

Models guess instead of asking. A receipt often carries a subtotal, a VAT line, a card total and change given. A system forced to return one "total" will pick one — and if it reports that pick with false confidence, the error hides until finance queries it.

Dates and time zones. Many apps store the purchase date as a timestamp and display it in another time zone, so the 14th becomes the 13th.

No confirmation step. The deepest problem is design, not accuracy. Any reader of photographs will sometimes be wrong. Without a clear "here is what I read — is it right?" moment, every error flows straight into the claim.

Crop handled badly. Forcing a crop on an e-receipt PDF, or crop handles that cannot reach half the image, adds friction without adding accuracy.

What good capture looks like: a checklist

Whether you are an employee or a finance manager evaluating expense app receipt capture, this is the list worth testing against:

1. Auto-crop with a fallback when unsure. Crop silently only when the app is confident it has the whole receipt; otherwise show a crop screen seeded with the full photo, so a quick OK can never lose part of it.
2. Read, then confirm, on one screen. Merchant, date, total, VAT and VAT number pre-filled, with the fields the app was unsure about clearly marked — so you check those, not everything.
3. Ask a plain question rather than guess. If the total or the VAT is ambiguous, say "check this" instead of asserting a number.
4. Handle "no VAT" receipts. Parking, trains and many small purchases carry no VAT. Record that and move on; never block the claim.
5. Remember cash vs card. Default to the person's last choice.
6. Never lose the photo. A crash mid-attach or a stuck upload must never delete the receipt.
7. Show status afterwards. Draft, submitted, approved or rejected — with the reason — visible without asking anyone.

If you are weighing this against staying manual, our breakdown of the true cost of manual expense claims covers what the spreadsheet route really costs.

How Costrix does each of these

We built Costrix around that checklist, and we would rather be precise than impressive about it.

• Auto-crop with a fallback. The receipt is detected and cropped in the browser before upload. When the detector is confident, the crop happens silently; when it is not, a crop screen appears seeded with the whole photo, so OK keeps everything. An "Adjust crop" button lets you redo it from the original photo until the receipt is saved. In our benchmark round on roughly 475 public receipt photos, no confident crop was wrong on a genuine receipt.
• Read, then confirm. An AI vision model reads merchant, date, total, VAT amount and VAT number, with a local tesseract OCR pipeline as a silent fallback whenever the AI cannot run. Every value is re-validated on the server and carries a confidence score. You confirm the lot on one pre-filled screen.
• Ask rather than guess. Any field the reader was unsure about is marked in amber with a "Looks right" confirmation beside it, so your eye goes to the one or two fields it doubted rather than all six. Where the merchant is recognisable the category is filled in as a suggestion and flagged the same amber way, because it is a guess — and some categories ask a plain follow-up such as "how many people?" for a meal.
• "No VAT" receipts are fine. VAT amount and VAT number are independent; a receipt with no VAT is recorded as exactly that, and nothing is blocked. For what counts as a valid VAT receipt, see our guide to VAT receipts for employee expenses in the UK and Ireland.
• Cash vs card remembered. The payment method defaults to your last choice on that device.
• The photo is never lost. The file is stored before any reading is trusted; cropping is a separate step on the same photo, and replacing a photo never overwrites fields you have already corrected.
• Status is visible. Every expense shows draft, submitted, approved or rejected with the reason, and every month shows a Submitted or Not submitted flag.

What this is worth in time

Here is our estimate, with the assumptions on show so you can substitute your own numbers.

That is about 8 minutes back per receipt. For one person doing 3 claims a month, roughly 24 minutes a month — about 5 hours a year. For a 50-person firm where 25 people claim, it is 75 claims a month: roughly 10 hours a month, or about £280 a month at £28 an hour. These are estimates for a typical UK SME, not measurements of your business, and they exclude manager and finance time.

What to do next

If your team has been burned by a receipt scanning app that guesses, try one that asks. Costrix has a 7-day free trial on every plan — start your free trial, photograph the most faded receipt in the bottom of your bag, and see what it reads and what it asks you to check.