Automate Purchase Order Data Extraction from PDFs

When a purchase order arrives as a PDF, someone has to get the data out of it: vendor name, PO number, line items, quantities, unit prices, delivery dates. Doing that by hand is slow, error-prone, and completely unscalable once volume picks up. Whether you are processing five POs a day or five hundred, you need a repeatable method that pulls the right fields every time without manual re-keying.

The approach on this page uses a structured prompt fed to a large language model. You paste or pipe in the text content of the PDF, and the model returns a clean JSON or table with every field you care about. No custom OCR pipeline to build, no vendor-specific template to maintain, no code to write for each new supplier format.

This works because modern LLMs understand document structure well enough to find PO number on line 3 or buried in a header block, match line items to their quantities and prices, and handle the inconsistencies that come from dealing with dozens of different supplier templates. The comparison table below shows how four leading models handle the same real-world PO input, so you can pick the right one for your volume and accuracy requirements.

When to use this

This approach is the right fit when you are dealing with PDFs from multiple suppliers with different layouts, need structured output you can feed into an ERP or spreadsheet, or want to automate a manual data-entry step without writing a custom parser for every vendor template. It handles both native text PDFs and, with an OCR pre-step, scanned documents.

• Accounts payable teams receiving POs from dozens of suppliers with inconsistent formatting
• Procurement analysts who need to aggregate line-item data from POs into a single spreadsheet or database
• Developers building an intake workflow where PO data must populate an ERP like SAP, NetSuite, or QuickBooks
• Operations teams validating that received goods match what was ordered by comparing extracted PO fields against packing slips
• Small businesses replacing a manual data-entry step that currently takes hours per week

When this format breaks down

• The PDF is a scanned image with no text layer and you have no OCR step in place. The model reads text, not pixels. You must run the file through Tesseract, AWS Textract, or a similar tool first.
• You need legally defensible extraction with zero tolerance for hallucination, such as a contract PO that triggers automatic payment. Always add a human review step before money moves.
• Your POs are in a single, fixed, machine-readable format you control end to end. A simple regex or a CSV export from your own system is faster and cheaper.
• The document is over 50 pages of dense line items and the model's context window cannot hold it all. You will need to chunk the document or use a purpose-built document AI service.

What makes these work

1. 01
   Always specify your output schema

   Tell the model exactly which fields you want and in what format before you paste the document. A prompt like 'Return a JSON object with these keys: po_number, vendor, line_items (array with description, qty, unit_price, total), po_total, delivery_date' produces consistent output you can parse programmatically. Without a schema, different POs will return fields with different names and structures, breaking any downstream automation.

2. 02
   Extract text from the PDF before prompting

   LLMs read text, not PDFs. Use a library like PyMuPDF (fitz), pdfplumber, or pdftotext to pull the raw text layer first, then pass that string to the model. For scanned POs, run the file through an OCR tool first. Feeding clean text instead of a file path means better extraction accuracy and no dependency on the model's file-handling capabilities.

3. 03
   Include validation instructions in the prompt

   Ask the model to cross-check its own arithmetic. A prompt addition like 'Verify that each line_total equals qty multiplied by unit_price, and that the sum of line_totals equals po_total. Flag any discrepancy with a mismatch field set to true' catches data-entry errors in the original document and extraction errors from the model itself.

4. 04
   Handle multi-page and multi-item POs with chunking

   If a PO has dozens of line items spread across several pages, concatenate all the extracted text and pass it in one prompt if it fits the context window. If it does not, split by page and merge the line_items arrays in your code afterward. Always include a system instruction that the document may continue across chunks so the model does not try to infer a false total from an incomplete view.

More example scenarios

PO Number: PO-2024-08821. Date: 2024-11-14. Vendor: Apex Industrial Supply. Ship To: 400 Commerce Dr, Detroit MI 48201. Line 1: Steel Hex Bolts M8x30, Qty 5000, Unit Price $0.04, Total $200.00. Line 2: Nylon Lock Nuts M8, Qty 5000, Unit Price $0.03, Total $150.00. Requested Delivery: 2024-12-01. Payment Terms: Net 30.

{"po_number":"PO-2024-08821","date":"2024-11-14","vendor":"Apex Industrial Supply","ship_to":"400 Commerce Dr, Detroit MI 48201","payment_terms":"Net 30","requested_delivery":"2024-12-01","line_items":[{"description":"Steel Hex Bolts M8x30","qty":5000,"unit_price":0.04,"total":200.00},{"description":"Nylon Lock Nuts M8","qty":5000,"unit_price":0.03,"total":150.00}],"po_total":350.00}

Purchase Order from: Westfield Retail Group. PO: WRG-55293. Vendor: BlueLine Apparel Co. Order Date: Jan 8 2025. Item: Men's Classic Crewneck Sweatshirt, Style BL-CC-401. Color: Navy. Sizes: S x20, M x40, L x35, XL x15. Unit Cost: $18.50. Total Units: 110. Total Cost: $2,035.00. Deliver By: Feb 15 2025.

{"po_number":"WRG-55293","buyer":"Westfield Retail Group","vendor":"BlueLine Apparel Co","order_date":"2025-01-08","deliver_by":"2025-02-15","line_items":[{"style":"BL-CC-401","description":"Men's Classic Crewneck Sweatshirt","color":"Navy","sizes":{"S":20,"M":40,"L":35,"XL":15},"total_units":110,"unit_cost":18.50,"line_total":2035.00}]}

Purchase Order. Issued by: Thornton Financial LLC. PO #: TF-2025-0042. To: CloudBase Software Inc. Date Issued: March 3, 2025. Line 1: Enterprise License, 50 seats, $120/seat/year, Subtotal $6,000. Line 2: Onboarding Professional Services, 10 hours, $250/hr, Subtotal $2,500. Payment Terms: 50% upfront, 50% on delivery. Total: $8,500.

{"po_number":"TF-2025-0042","buyer":"Thornton Financial LLC","vendor":"CloudBase Software Inc","issue_date":"2025-03-03","payment_terms":"50% upfront, 50% on delivery","line_items":[{"description":"Enterprise License","qty":50,"unit":"seats","unit_price":120,"period":"year","subtotal":6000},{"description":"Onboarding Professional Services","qty":10,"unit":"hours","unit_price":250,"subtotal":2500}],"po_total":8500}

PO Date: 04/02/2025. PO Number: RC-0882. From: Harvest Table Restaurants. To: Pacific Coast Distributors. Items: Olive Oil Extra Virgin 4L x 24 units @ $14.00 = $336.00; Canned Tomatoes 2.5kg x 48 units @ $3.20 = $153.60; Dried Pasta 5kg x 30 units @ $6.50 = $195.00. Delivery Required: 04/07/2025. Total: $684.60.

{"po_number":"RC-0882","po_date":"2025-04-02","buyer":"Harvest Table Restaurants","vendor":"Pacific Coast Distributors","delivery_required":"2025-04-07","line_items":[{"description":"Olive Oil Extra Virgin 4L","qty":24,"unit_price":14.00,"total":336.00},{"description":"Canned Tomatoes 2.5kg","qty":48,"unit_price":3.20,"total":153.60},{"description":"Dried Pasta 5kg","qty":30,"unit_price":6.50,"total":195.00}],"po_total":684.60}

Purchase Order No: CPO-2025-1147. Project: Riverside Commercial Build. Project Code: RCB-7. Contractor: Summit Build Co. Supplier: BuildRight Materials Inc. Date: 05 May 2025. Ready-Mix Concrete 30MPa, 18 cubic meters, $210/m3, $3,780. Rebar 12mm D500N, 2000kg, $1.85/kg, $3,700. Delivery Address: 88 River Rd, Auckland. Required On Site: 12 May 2025. PO Total: $7,480.

{"po_number":"CPO-2025-1147","project":"Riverside Commercial Build","project_code":"RCB-7","contractor":"Summit Build Co","supplier":"BuildRight Materials Inc","po_date":"2025-05-05","required_on_site":"2025-05-12","delivery_address":"88 River Rd, Auckland","line_items":[{"description":"Ready-Mix Concrete 30MPa","qty":18,"unit":"cubic meters","unit_price":210,"total":3780},{"description":"Rebar 12mm D500N","qty":2000,"unit":"kg","unit_price":1.85,"total":3700}],"po_total":7480}

Common mistakes to avoid

• Assuming the PDF has a text layer

  Many POs are scanned documents with no extractable text. If you pass raw bytes or a file path without first confirming there is a text layer, the model gets nothing useful and will hallucinate or return empty fields. Always check programmatically with pdfplumber or similar before deciding whether OCR is needed.

• Not pinning the output format

  Without an explicit schema in the prompt, the model may return line items as a flat table in one response and a nested JSON array in the next, depending on how the PO is laid out. This inconsistency breaks parsers downstream. Define the exact field names, data types, and nesting structure in every prompt.

• Ignoring extraction confidence on messy documents

  Poorly scanned POs, faxed documents, or files with overlapping text can produce garbage OCR output that the model will try to make sense of, sometimes confidently wrong. Ask the model to return a low_confidence flag on any field it is uncertain about, and build a human review queue for those records.

• Skipping totals reconciliation

  A model might correctly read each line item but miss that the document has a freight charge or tax line at the bottom, making the extracted total wrong. Always instruct the model to extract every numeric total it finds, including subtotals, tax, freight, and grand total, and never assume po_total equals the sum of visible line items.

• Running extraction without a test suite across vendor formats

  A prompt tuned on your top supplier's PO format will likely fail on a supplier who uses a completely different layout. Before deploying to production, test against at least ten real POs from different vendors. Identify the failure cases and add examples or clarifying instructions to the prompt until accuracy is acceptable across all formats.

Related queries