Accounts Payable (AP) departments are drowning in paper. The average cost to process a single invoice manually is $12.90, and it takes 9.7 days.
DocsRouter allows you to bring that cost down to pennies and the time down to seconds.
The Challenge: Line Item Extraction
Getting the "Total Amount" is easy. Even legacy OCR can do that. The real challenge is line items.
An invoice might have a table that spans 3 pages. It might have merged cells. It might be handwritten. Most legacy tools (like AWS Textract or Azure Form Recognizer) struggle with complex, nested tables.
The Solution: Vision LLMs
Models like GPT-5.2 and Gemini 3 Pro treat the document as a visual scene. They understand that a row continues on the next page because they "read" the context, not just the pixels.
How to Implement 2-Way Matching
2-way matching verifies that the Invoice amount matches the Purchase Order amount.
- Ingest: Send the PDF to
POST /v1/ocrwithmodel: "openai/gpt-5-2". - Extract: Request a schema that includes
line_items[]andpo_number. - Match: DocsRouter returns structured JSON. Your backend simply checks:
if (ocrResult.total === databasePO.total) { approveInvoice(); }
Why accurate line items matter
It's not just about paying the bill. It's about data.
- Spend Analysis: "How much did we spend on 'Screws' vs 'Bolts'?"
- Fraud Detection: "Why is the unit price for this item 50% higher than last month?"
You can't answer these questions if you only extract the Total. DocsRouter unlocks the data trapped in your PDFs.