Receipt OCR: how to extract data from receipts accurately
For finance, expense and AP teams handling receipts at volume: which fields to extract, how receipt OCR works, the checks that keep bad data out of your books, and how to pick a tool.
Key takeaways
- Receipt OCR turns a photo, scan or PDF of a receipt into structured data: merchant, date, line items, tax, tip, total and payment method.
- Receipts are harder than invoices: thermal paper fades, photos are skewed, layouts vary by merchant and many show several totals.
- Plain OCR only gives you text. AI-based extraction works out which number is the total, which is tax and which lines are items, and flags low-confidence values.
- Run checks after extraction: line items add up, the date falls in the claim period, the total matches the card charge and the receipt hasn't been claimed before.
- For US tax records, IRS Publication 463 asks for documentary evidence, such as a receipt, for every lodging expense and for any other travel expense of $75 or more.
On this page
Receipt OCR uses optical character recognition, usually paired with AI, to read a photo, scan or PDF of a receipt and turn it into structured data: merchant, date, line items, tax, tip, total and payment method. Finance, expense and AP teams use it to stop typing receipts into expense reports and accounting systems, and to check each claim automatically. Accuracy depends on the image and on whether the tool understands receipts or only reads their text.
This guide covers the fields to extract, how receipt OCR works, what makes receipts hard to read, the checks to run and how to choose a tool.
What data does receipt OCR extract?#
| Field | Why it matters |
|---|---|
| Merchant name, address, phone | Identifies the vendor and supports categorization |
| Transaction date and time | Places the expense in the right period and policy window |
| Line items (description, quantity, price) | Needed for itemized policies, split costs and non-reimbursable items |
| Subtotal, tax, tip, total | The amounts posted to the ledger and checked against the card charge |
| Currency | Required for foreign receipts and conversion |
| Payment method and last 4 digits | Matches the receipt to a corporate card transaction |
| Receipt or transaction number | Helps detect duplicates |
How receipt OCR works#
Receipts arrive as phone photos, emailed receipts, scans and PDFs from travel and ride-hailing apps. Automated receipt data capture runs each one through the same four stages.
- Phone photos
- Emailed receipts
- PDFs and scans
- 01Clean up the image
- 02Read the text
- 03Extract the fields
- 04Validate
Clean-up crops, straightens and sharpens the image. Extraction is where plain OCR stops: the model decides which text is the merchant, which rows are items and which of several totals to post, and scores its confidence in each field. Validation checks the values and sends anything low-confidence or failing to a person.
What makes receipts hard to read#
Thermal paper fades
Many receipts are printed on thermal paper, which fades over time, so capture them early.Photos, not scans
Shadows, curls, glare and angles distort the text. Clean-up and models trained on photos help most.No standard layout
Every point-of-sale system prints differently, with wrapped lines, abbreviations, other languages and currencies.Several totals
Subtotal, total, amount tendered, change and a tip line all look like "the total".Long receipts
Grocery and hardware receipts run to dozens of lines and are often photographed in two parts.Handwriting
Tips and notes are often written by hand, so route low-confidence handwritten values to a person.
Checks to run after extraction#
Extraction gets the data out; checks keep bad data out of your books. The same checks apply to every receipt type, from store and restaurant receipts to emailed e-receipts, hotel folios and fuel receipts.
- MathLine items add up to the subtotal, and subtotal plus tax and tip equals the total.
- TaxThe tax amount is plausible for the location.
- DateThe date falls in the claim period and isn't in the future.
- DuplicatesThe same merchant, date and amount hasn't been claimed before, by anyone.
- Card matchThe total matches a corporate card transaction.
- PolicyCategory, amount and items fit your expense policy, such as meal limits or alcohol rules.
- EvidenceA receipt is attached where required. For travel, IRS Publication 463 asks for one for every lodging expense and for any other expense of $75 or more.
In Docsumo, you add checks like these as workflow steps: an AI step, or your own Python code. Receipts that fail go to a reviewer, who clicks a flagged field to see its source line highlighted on the image. See document workflows.
How to choose a receipt OCR tool#
- Test on your own receiptsInclude faded, crumpled, foreign and long ones, and measure accuracy per field, not per receipt.
- Check line itemsMany tools return only merchant, date and total.
- Look at the review screenReviewers should see low-confidence fields on the image and fix them quickly.
- Check validation and duplicate detectionThey matter as much as extraction.
- Confirm the integrationYour expense tool, accounting system or ERP, or an API and webhooks for your own flow.
- Check securityReceipts can show card digits and personal data, so ask for a SOC 2 Type 2 report.
Docsumo isn't limited to pre-built models, so one platform reads receipts alongside invoices and bank statements. It reports 99% field-level accuracy across 250+ pre-trained document types and 95%+ straight-through processing. Data goes to your expense tool, accounting system or ERP through the API and webhooks, or downloads to Excel. Try it free for 14 days on up to 1,000 pages; see pricing and security.
The bottom line#
Receipt OCR saves the most time when it goes past reading text: picking the right total, tax and line items, checking them against the card charge and your policy, and sending only the exceptions to a person. Capture receipts early, test tools on your worst ones, and put checks after extraction. For receipts that come with supplier invoices, see accounts payable automation.
Book a demo with a batch of your own receipts, or start a free trial.
Frequently asked questions#
What is receipt OCR?
Receipt OCR is optical character recognition applied to receipts. It reads the text on a receipt image and, combined with AI extraction, returns fields such as merchant, date, items, tax and total as structured data.
How accurate is receipt OCR?
It depends on the image and the tool. Clear, flat receipts read well; faded thermal paper, crumpled or long receipts, and handwritten tips are harder. Measure field-level accuracy on your own receipts, and send low-confidence fields to a person.
Can I automate receipt data entry?
Yes. Employees photograph or forward receipts, receipt OCR extracts the fields, rules check them against the card charge and your expense policy, and the data goes to your expense tool or accounting software, with the image attached. People only look at the receipts that fail a check.
What are the types of receipts?
The ones businesses handle most are purchase receipts from stores and restaurants, emailed e-receipts from airlines and ride-hailing apps, hotel folios, fuel receipts, and payment receipts that confirm an invoice was paid. They share the same core fields: merchant, date, amounts and payment method.
Are scanned receipts acceptable for business records?
Generally, yes. The IRS allows paper records to be kept in an electronic storage system that meets Revenue Procedure 97-22, so a legible scan or photo can replace the paper. Each receipt still needs to show the amount, date, place and nature of the expense. Check with your tax adviser for your situation.
What's the difference between receipt OCR and invoice OCR?
Invoices are usually PDFs with an invoice number, due date and payment terms, and they feed accounts payable. Receipts prove a payment already made, are often photos, and feed expense reports. See invoice OCR and invoice extraction.
Sources
- IRS Publication 463 (2025): Travel, Gift, and Car Expenses
- IRS: Automated records (electronic storage systems and Rev. Proc. 97-22)
First published . Last updated .