# Using AI to Read Receipts: Accuracy, Setup and Limits

*How-to — 2026-11-02 — by Mahmoud Zalt*

What AI actually gets right when it reads a receipt photo, which fields to trust, and where a human still has to look before the books close.

**Short answer.** **AI reads a clear photo of a printed receipt well, and reads a crumpled faded one badly.** Merchant, date and total come out reliably from a decent capture. Line items, tax breakdowns and anything handwritten are where the errors live. Set a confidence threshold, let the confident ones through, and glance at the rest before anything reaches your books.

The receipt problem is physical before it is technical. A receipt is a small piece of thermal paper that lives in a pocket, fades in a wallet, creases in a bag, and gets photographed at an angle on a restaurant table in bad light. Every one of those things costs you accuracy, and none of them is the model's fault.

That is worth knowing up front because it changes where you spend effort. Most people go shopping for a smarter tool when the thing that would actually help is photographing the receipt flat, in daylight, the moment they get it. Capture quality beats tool choice, and it is free.

## Can AI actually read a receipt?

Yes, for printed receipts photographed reasonably. Modern vision models pull merchant name, date, total, payment method and often the individual line items out of an ordinary phone photo, without anyone building a template for that particular shop. That is a real change from the template-matching tools of a few years ago.

What has not changed is that this is recognition, not reading in the way you read. The system is predicting the most likely characters given a blurry image, so it will occasionally produce a confident, well-formatted, completely wrong number. That is the failure mode to design around.

Microsoft's own documentation on [the limits of text recognition](https://learn.microsoft.com/en-us/azure/foundry/responsible-ai/computer-vision/ocr-characteristics-and-limitations) is direct about the causes: scan quality, resolution, contrast, lighting, rotation, and the size, colour and density of the text all affect the result. A receipt manages to be difficult on nearly every one of those axes at once.

## Which receipt fields can you actually trust?

Not all fields are equally hard, and treating them the same is why people either over-trust or over-review. Merchant, date and total are large, high-contrast and near the edges of the paper, so they survive a mediocre photo. Line items and tax are small, dense and in the middle, so they degrade first.

| Field | How reliable | What breaks it |
|---|---|---|
| Merchant name | High, and easy to sanity check | A logo instead of text, or a franchise trading name that differs from the legal entity |
| Date | High, but the format is the trap | Day-month versus month-day ordering, which silently produces a valid wrong date |
| Total amount | High on a clear photo | Fading, creases across the number, and receipts that print a tip line after the total |
| Currency | Moderate | Symbols shared across countries, and receipts that print no symbol at all |
| Tax or VAT breakdown | Moderate to low | Small dense print, multiple rates on one receipt, and inclusive versus exclusive pricing |
| Line items | Low to moderate | Abbreviated product codes, wrapped descriptions, and long receipts photographed in one shot |
| Anything handwritten | Low | Handwriting is the hardest case in document recognition, including a scribbled tip |

The date row deserves special attention because it fails so quietly. A receipt reading 03/04 is the third of April or the fourth of March depending on where it was printed, and both are valid dates that will pass any format check. If you travel or buy from abroad, pin the expected format per merchant country rather than trusting a guess.

## How do you set up receipt reading that works?

The setup that survives contact with real life is short. Most of it is about capture and about what happens to the uncertain ones, because the reading step is the part you do not have to build.

### A receipt pipeline that holds up

1. **Capture at the moment of purchase** — Photograph the receipt flat, in the best light available, before it goes in a pocket. Thermal paper fades with heat and time, so a receipt photographed three weeks later is a genuinely harder document than the same receipt photographed on the day.
2. **One inbox for everything** — Email forwards, phone photos, PDF receipts from online orders. Send all of them to one place. A pipeline that only handles one arrival route leaves you doing the others by hand, which is how the whole habit collapses.
3. **Extract with a confidence value per field** — Insist on a score, not just an answer. Without it you cannot separate the receipts that need a look from the ones that do not, and you end up reviewing all of them or none of them.
4. **Run the arithmetic and duplicate checks** — Line items should sum to the total before tax. The same merchant, date and amount appearing twice usually means a receipt was submitted twice. These catch errors no confidence score will.
5. **Review only what failed** — A queue showing the photo next to the extracted fields, with the uncertain ones highlighted. Confirming one is a few seconds. This is the step that makes the whole thing safe, and it is the step people skip.
6. **Keep the image attached to the record** — Tax authorities generally want the source document, not your extracted summary of it. Storing the original photo alongside the row costs almost nothing and is the difference between a clean audit and an unpleasant afternoon.

Notice how little of that is about which model reads the receipt. Capture, one inbox, arithmetic checks, a short review queue and keeping the original. Those five decisions determine whether this works far more than the choice of extraction engine does.

## Who should actually be doing this for you?

If you are a solo founder or a small team, the honest answer is that assembling a capture app, an extraction service, a rules layer and a review queue is more project than the problem deserves. The work is real but it is not complicated, and it is exactly the kind of steady admin that should belong to someone else.

That is the shape of what we built at [Sistava](/hire-ai-employees). You hire an AI Employee, forward receipts to it, and tell it what you want in plain English: pull the merchant, date, total and tax, check the sums, file the image, and ask me about anything that looks off. The rules are a conversation, not a configuration screen.

The practical difference is that an employee handles the awkward middle. A receipt in a foreign currency, a duplicate submission, a photo where the total is under a crease. Those are the ones that break a fixed pipeline and quietly end up as a wrong number, and they are exactly the ones worth asking a person about.

## Where does receipt reading go wrong?

Say this part plainly, because expense data feeds tax filings and nobody enjoys discovering an error at the worst possible time. Automated receipt reading is a very good assistant and a poor final authority. Here is where it slips.

- **Faded thermal paper. **Contrast is one of the documented drivers of recognition accuracy, and thermal receipts lose contrast with heat, sunlight and time. A receipt found in a coat pocket in spring may simply be unreadable, by any tool and by you.
- **Handwritten additions. **A tip written on the slip, a note about who the meal was for, an amended total. Handwriting recognition is materially weaker than print, so treat any handwritten figure as needing eyes.
- **Tax on mixed-rate receipts. **A supermarket receipt with items at different rates is one of the harder documents there is. If you reclaim tax, check these rather than trusting the split, because the error is invisible until someone reconciles.
- **Category guesses. **Whether a purchase is a business expense is a judgement about intent, not something printed on the paper. A tool suggesting a category is helping you sort. It is not deciding your tax position and should never be treated as if it were.
- **Very long receipts. **A metre of supermarket till roll photographed in one shot gives every line too few pixels. Photograph it in overlapping sections, or accept that only the total will be reliable.

The workable rule is that anything feeding a tax return or a legal record gets a human on the numbers that matter. Not every field, and not every receipt. The total, the date, the tax, and anything the system flagged. That is a minute a week for most solo businesses and it removes the only risk worth worrying about.

One habit is worth more than any of the tooling above: do it weekly. Receipt automation fails for most people not because extraction was wrong but because the photos never got taken and the review queue never got cleared. Fifteen minutes on a Friday keeps the whole thing honest.

The reason weekly beats monthly is the paper itself. A receipt photographed the day you got it is a clean, high-contrast document. The same receipt found in a bag five weeks later has faded, creased and picked up a coffee ring, and no amount of clever software recovers information that is no longer on the page. If remembering is the hard part, that is exactly the sort of nudge [Sistava](/personal-ai-assistance) can own for you, so the reminder arrives without you having to set it every week.

However you end up doing it, keep the originals. Extraction improves, your accountant may want to see the source, and a tax authority is entitled to ask. The photo is the record, and the extracted row is a convenience built on top of it. That is also worth holding in mind when you compare what these tools charge. Reading one receipt costs very little, at [Sistava](/pricing) or anywhere else, so let accuracy and the review step decide your choice rather than a per-page price.

## Frequently asked questions

## FAQ

### How accurate is AI at reading receipts?

On a clear photo of a printed receipt, merchant, date and total come out reliably. Line items, tax breakdowns and anything handwritten are noticeably weaker. Accuracy tracks image quality more than anything else, since contrast, resolution, lighting and rotation all affect what can be recognised. Measure it on your own receipts rather than trusting a vendor figure measured on clean samples.

### Can AI read a crumpled or faded receipt?

Sometimes, and less well than a fresh one. Thermal paper loses contrast with heat and time, and contrast is one of the main things recognition depends on. Flattening the receipt and photographing it in good light helps more than switching tools. If the print has faded to the point where you struggle to read it, no tool will do better.

### Do I still need to check receipts if AI reads them?

Yes, but not all of them. Use a confidence threshold so the clear ones pass without you, and review the flagged ones plus anything that failed an arithmetic or duplicate check. For expenses that feed a tax return, confirm the total, the date and the tax figure. That is seconds per receipt and it removes the failure that actually costs money.

### Can it handle receipts in other languages and currencies?

Printed receipts in common languages are usually read well, and the currency is often inferable from the merchant and the format. The reliable practice is to store the currency explicitly rather than assuming, and to pin the expected date format by merchant country. A day-month receipt read as month-day produces a valid date that is silently wrong.

### What is the best way to photograph a receipt?

Flat on a plain surface, in even light, filling the frame, straight on rather than at an angle, and as soon as you receive it. For a very long receipt, take overlapping sections instead of one distant shot. This costs nothing and improves results more than any change of software.

### Can AI categorise expenses as well as read them?

It can suggest categories, and suggestions are genuinely useful for sorting a pile quickly. Treat them as suggestions. Whether a purchase counts as a business expense depends on intent and on your local rules, neither of which is printed on the receipt. Keep that decision with you or your accountant.

**Tags:** receipts, expenses, data-extraction, bookkeeping, ai-employee