Document Extraction Pipeline

Extract structured data from invoices, receipts, contracts, and purchase orders — validate the output automatically and flag anything that needs human review.

Before You Start

You need the following to run this workflow:

• Claude Code (desktop app, CLI, or IDE extension). Claude reads documents and extracts fields during Steps 3 and 4.
• Documents in a readable format — PDF, image (PNG/JPG), or scanned files. Claude Code can read PDFs and images directly. For very poor quality scans, run OCR first (see tips in Step 2).
• A spreadsheet app — Google Sheets, Excel, or CSV editor for the final output.
• 20-30 minutes for a batch of 10-15 documents. Process documents in batches of the same type (e.g. all invoices together, all contracts together) for the most consistent results.

Accuracy expectations: Standard invoices with clean formatting typically see 94-97% field-level accuracy. Messy scans, handwritten notes, and unusual layouts will have lower accuracy — the workflow flags these for human review rather than guessing. That tradeoff is intentional: 95% of the grunt work disappears, and the remaining 5% gets flagged with context so a human resolves it in seconds.

Quick Reference

Step	Who	What you do	What you get
1	You	Choose document type, define extraction schema	Reusable schema template
2	You	Gather 10-15 documents in a folder	Batch ready for processing
3	Claude	Read each document, extract fields	Raw extraction table
4	Claude	Validate math, completeness, and format	CLEAN and FLAGGED record sets
5	You	Resolve flagged items against originals	All records verified
6	You + Claude	Export to CSV, import to your system	Structured data in your workflow

Going Further

Once you have run this workflow on a few batches:

• Build schema variants. Create separate schema files for each document type you process regularly — invoices, purchase orders, expense receipts, contracts. Reuse them across batches without redefining fields each time.
• Track accuracy over time. Note how many documents come back CLEAN vs FLAGGED per batch. If a specific vendor's invoices consistently flag, add vendor-specific extraction hints to your schema (e.g. "This vendor puts the invoice number in the footer").
• Scale with confidence. Once your schemas are calibrated and your CLEAN rate is above 90%, you can increase batch sizes or process more frequently. The 5% that needs review will stay roughly constant — it is the nature of messy real-world documents, not a limitation you can eliminate.
• Automate the pipeline. Schedule this as a recurring Claude Code task that watches a folder for new documents, extracts and validates automatically, and drops the CSV into a shared drive. Human review only triggers when FLAGGED items appear.