# Invoice Data Extraction Prompt Pack

Turn a stack of invoice PDFs into clean, structured data using any capable AI assistant. ChatGPT, Claude, and Gemini all work. No coding, no new software. Two prompts, a field checklist, and a verification pass that catches mistakes before they hit your books.

## What you need

- An AI assistant that accepts file uploads (ChatGPT, Claude, and Gemini all do)
- Invoice PDFs, or clear scans or photos of paper invoices
- A spreadsheet to paste results into
- About 15 minutes to set up and test

## Setup

1. Decide which fields you actually need. The prompt below covers the common set: vendor name, invoice number, invoice date, due date, PO number, currency, subtotal, tax, and total. Delete any column you don't use. Fewer fields means fewer errors.
2. Open a fresh chat with your assistant. One chat per batch keeps the context clean.
3. Upload one to five invoices at a time. Larger batches tend to degrade accuracy, especially with scanned documents.
4. Paste Prompt 1. Then run Prompt 2 on the same output before you trust anything.

## Prompt 1 — Extraction

```
You are extracting data from the attached invoice(s). For each invoice, return one row in a markdown table with exactly these columns:

vendor_name | invoice_number | invoice_date | due_date | po_number | currency | subtotal | tax | total

Rules:
- Copy values exactly as printed. Do not reformat amounts or "fix" anything.
- Dates: convert to YYYY-MM-DD. If a date format is ambiguous (e.g. 03/04/2025), write it as printed and flag it with [AMBIGUOUS].
- If a field is not present on the invoice, write MISSING. Never guess or infer a value.
- If a value is partially unreadable, write what you can read plus [UNCLEAR].
- If both a brand name and a legal/billing entity appear, use the legal entity as vendor_name.
- After the table, list every invoice where subtotal + tax does not equal total, and every flagged field.
```

## Prompt 2 — Verification pass

Run this in the same chat, immediately after the extraction:

```
Now verify your own extraction. For each invoice:

1. Re-read the document and check every field in the table against the source, one field at a time.
2. Confirm the arithmetic: subtotal + tax = total. Show the calculation.
3. Confirm the due date is on or after the invoice date.
4. Check the invoice number against the others in this batch for duplicates.

Output a corrected table, then a short list titled "NEEDS HUMAN REVIEW" containing every field you are not fully certain about, with the reason. If everything checks out, say so explicitly.
```

## Field checklist

Scan each row before it goes into your spreadsheet:

- [ ] Vendor name matches the entity you have on file
- [ ] Invoice number is unique, not a repeat from a prior batch
- [ ] Dates are real dates and the due date follows the invoice date
- [ ] Math checks: subtotal + tax = total
- [ ] Currency is what you expect from this vendor
- [ ] Everything marked MISSING, AMBIGUOUS, or UNCLEAR got a human look

Spot-check rows against the original PDFs for at least your first few batches. Once you learn where your invoices trip the model (handwritten notes, stamps over the totals, multi-page layouts), add a rule for it to Prompt 1.

## Tips

- Scan quality matters. If a page is skewed or low-resolution, re-scan it before blaming the prompt.
- If one PDF contains several invoices, say so explicitly ("this file contains 3 invoices") or split the file first.
- For recurring vendors, keep a short note of each vendor's quirks (where the PO number hides, which date is which) and paste it above Prompt 1.

## When this outgrows DIY

This workflow holds up for a handful of invoices a day. It stops being the right tool when someone on your team is spending real hours uploading, checking, and pasting, or when the data needs to land in your accounting system automatically instead of a spreadsheet. That is the point where a built pipeline earns its keep: it watches the inbox, extracts and validates against your vendor list and PO records, posts to your accounting software, and only asks a human to review exceptions.

That is the kind of system agentclaw builds. Start with a free AI opportunity audit at https://agentclawhq.com/book

---

Free resource from agentclaw — AI agents and workflow automation for growing businesses. https://agentclawhq.com