How to extract data from a receipt photo or PDF (3 methods)

Receipts are the messiest documents to extract data from: thermal paper that fades, wrinkled photos, rotated scans, and a total lack of consistency between merchants. Here are three approaches that work in practice.

Method 1: OCR apps (for one-off receipts)

For occasional receipt capture, dedicated apps work well: Adobe Scan, Microsoft Lens, or Google Drive's built-in OCR (upload a photo, right-click → Open With Google Docs) all extract text from a receipt photo. You'll get the raw text, which you then manually pull fields from.

The limitation: you still have to read the OCR output and find the merchant name, total, and line items yourself. For a single receipt that's fine; for twenty receipts a week, it's unsustainable.

Method 2: Python with pytesseract or easyocr

For batch receipt processing in a pipeline (expense management systems, accounting automation), Python lets you run OCR at scale.

• pip install easyocr pillow
• import easyocr; reader = easyocr.Reader(['en'])
• result = reader.readtext('receipt.jpg', detail=0)
• text = ' '.join(result)

The hard part isn't the OCR — it's parsing the output into structured fields. Merchant names appear in different positions, totals may be labeled 'Total', 'Amount Due', 'Grand Total', or just be the last dollar figure. Tax shows up as 'Tax', 'GST', 'VAT', 'HST', or percentage notation. Writing robust regex or parsing logic for this is a significant engineering project.

If you're processing receipts from a narrow set of known merchants with consistent layouts, the parsing logic is tractable. If you're processing receipts from any restaurant, retail store, or service provider, you'll spend more time on edge cases than the automation saves.

Method 3: AI extraction (structured output from any receipt photo)

Multimodal AI extraction reads the receipt the way a person does — it understands that 'SUBTOTAL' and 'Sub-Total' and 'Sub Total' all mean the same thing, that the merchant name is usually at the top even if it's in a large font, and that line items are the rows between the merchant header and the subtotal. You get a clean structured object back.

What you can extract from a receipt

• Merchant name and address
• Date and time
• Line items — description, quantity, unit price
• Subtotal, tax (GST/VAT/HST), tips, discounts
• Total and payment method (cash, card, card type)
• Receipt or transaction number

Receipt photos vs PDFs

Phone photos of receipts are harder than PDFs — they may be rotated, poorly lit, or have the paper curled at the edges. Multimodal AI models handle this because they process the image visually rather than relying on a clean text layer. pytesseract struggles with rotated or low-contrast thermal paper; AI extraction doesn't.

Building a receipt processing pipeline

A common setup: employees forward receipt photos to a shared inbox, a Zapier/Make automation routes each attachment to ExtractFox's API, and the structured JSON lands in a Google Sheet or accounting tool. The engineering surface is one API call per receipt rather than a custom OCR pipeline.

Receipt data extraction vs expense apps

Expensify, Concur, and Ramp capture receipts inside their own workflow — great if your whole company lives on that stack. Receipt data extraction is the layer underneath: turning a photo into merchant, date, line items, and total as structured rows you can import anywhere. Use an expense app when you want policy enforcement and approvals; use extraction when you need the raw data in your own spreadsheet, ERP, or custom pipeline.

Exporting to Excel for expense reports

Finance teams usually want one row per receipt (merchant, date, category, total) or one row per line item for VAT reclaim. Excel export should return amounts as numbers, not strings with dollar signs — otherwise SUM() and pivot tables break. Map merchant → vendor, date → transaction_date, and total → amount once; the schema stays the same whether the receipt is from a coffee shop or a hotel chain.

Common receipt extraction mistakes

• Using raw OCR text as structured data — you still need to find which number is the total.
• Ignoring tax as a separate field — GST/VAT lines matter for reclaim workflows.
• Storing amounts with currency symbols — keep currency at header level, amounts as numbers.
• Waiting until thermal ink fades — extract while the receipt is still readable.
• Treating tip and total as the same field — many restaurants print them on separate lines.