How to Extract Data from Receipts with Claude

To extract data from receipts with Claude, attach a photo of your receipt to a claude.ai conversation and type a specific instruction — for example, 'Extract the date, restaurant name, each line item with price, subtotal, tax, tip, and total. Return a markdown table.' Claude analyzes the image and returns structured data you can paste directly into a spreadsheet or expense tool. This works on every plan, including the free tier, with no special setup required.

What Is Claude's Receipt-Reading Capability?

Claude's vision feature — also called image input — lets it receive, read, and reason about static image files alongside text. When you attach a receipt photo, Claude analyzes the visual content and responds in natural language, extracting text, interpreting layout, and organizing data into whatever format you request. As the Anthropic vision documentation describes, practical applications range from simple tasks like reading a receipt to advanced document-processing pipelines.

Importantly, this is a one-way capability: Claude can understand images but cannot generate, edit, manipulate, or create them. It is purely an analysis and extraction tool.

How Do You Extract Receipt Data Step by Step?

1. Open claude.ai in your browser and start or continue a conversation.
2. Attach your receipt photo by clicking the paperclip or attachment icon in the chat input bar, dragging and dropping the image file directly into the chat window, or pasting from your clipboard.
3. Supported formats are JPEG, PNG, GIF, and WebP. Each image must be within the per-image size limit on claude.ai.
4. Type your extraction instruction alongside the image. Be explicit about what fields you want and how you want them formatted. For example: Extract the date, restaurant name, each line item with price, subtotal, tax, tip (if shown), and total. Return a markdown table. If any value is unclear, note it in parentheses.
5. Press Enter or click Send. Claude will analyze the image and respond with the structured data.
6. Review flagged fields. If Claude notes that a value is partially obscured or unclear, zoom into that section of the receipt and ask a follow-up question to confirm the value.

What Does a Good Extraction Prompt Look Like?

Specificity is the key to clean output. Compare these two prompts:

• Vague: What's on this receipt?
• Precise: Extract the date, vendor name, each line item with price, subtotal, tax, tip, and total. Return a markdown table. If any value is unclear, note it in parentheses.

The precise prompt tells Claude exactly which fields to find, what format to use, and how to handle uncertainty. The result is a table you can paste into a spreadsheet without reformatting.

What Does the Output Actually Look Like?

For a restaurant receipt, Claude might return something like this:

| Field             | Value           |
|-------------------|-----------------|
| Date              | 2025-11-04      |
| Restaurant        | Luigi's Trattoria |
| Pasta Carbonara   | $18.00          |
| Caesar Salad      | $14.50          |
| Espresso          | $4.50           |
| Subtotal          | $37.00          |
| Tax               | $3.28           |
| Tip               | $7.00           |
| Total             | $47.28          |

*Note: The tip amount is partially obscured; value shown is best estimate.*

Notice that Claude proactively flags the uncertain field rather than silently guessing. This built-in transparency is what makes it practical for real expense workflows where accuracy matters.

What Are the Most Common Pitfalls — and How Do You Fix Them?

Blurry or Low-Resolution Photos

The single biggest cause of extraction errors is a blurry or poorly lit receipt photo. Ensure the image is sharp and well-lit before uploading. For documents with fine text, zoom in or crop to the relevant section rather than uploading a full-page photo from a distance. Re-photograph in better lighting if the original is blurry.

Rotated or Skewed Images

Rotate the image to the correct orientation using your phone's photo editor or a desktop tool before uploading. Claude can sometimes compensate for slight skew, but upside-down or sideways images significantly degrade accuracy.

Putting the Image After a Long Block of Text

Put the image before your text instructions whenever possible. While Claude will still process images placed after text, leading with the image and following with the question produces more accurate and focused responses.

Expecting Claude to Edit or Enhance the Image

Claude is an image understanding tool only — it cannot create or alter images. If your receipt photo is too dark or blurry, fix it in your phone's photo editor before uploading, not by asking Claude to improve it.

When Should You Use Vision vs. Other Approaches?

The guiding principle: if the content can be copied as text, paste it as text. Vision is for content that only exists as an image.

How Does This Work for Developers Building Expense Pipelines?

If you are building an automated expense or invoice processing system, Claude's vision capability is available through the Messages API. You can include image data as base64-encoded strings or as URL source blocks in the content array of your API request, paired with a text instruction. The Files API lets you upload an image once and reuse it by file ID across multiple requests, which keeps payloads small and avoids redundant encoding — a significant advantage when processing large volumes of receipts.

For general API examples, claude-sonnet-4-5-20250929 is a solid model choice that balances capability and efficiency for document extraction tasks.

What Kinds of Receipt and Invoice Tasks Can Claude Handle?

Vision handles a wide range of document extraction scenarios beyond simple restaurant receipts:

• Restaurant and retail receipts — date, vendor, line items, subtotal, tax, tip, total
• Vendor invoices — invoice number, billing address, itemized services, payment terms
• Scanned paper forms — insurance applications, onboarding documents, compliance checklists
• Business cards — name, title, company, email, phone, address
• Handwritten receipts — Claude can read printed and many handwritten formats, though printed text yields higher accuracy

The same structured-extraction approach applies to all of these: attach the image, specify the fields you want, specify the output format, and ask Claude to flag anything it cannot read with confidence.

Is Receipt Extraction with Claude Accurate Enough for Real Expense Reports?

For standard printed receipts photographed in good lighting, accuracy is high enough for practical use — especially when you ask Claude to flag uncertain values rather than silently guess. The workflow described in the source material explicitly builds in a human-review checkpoint: "Review the table Claude returns and note any low-confidence fields it flags. Correct any flagged values by zooming into that section and asking a follow-up question."

This makes Claude a strong first-pass extraction tool rather than a fully autonomous one. You get the speed of automated OCR with the transparency of knowing exactly where to double-check. For high-stakes financial documents, always treat Claude's output as a starting point for human review.

Vision is available on every plan tier — Free through Enterprise — so there is no paywall between you and this capability. You can test it immediately on claude.ai without any configuration.