GenAI invoice matching
Co-architected KIP.ai, a cloud-native, three-way invoice-matching Generative AI product, with AWS, reducing manual labour by over 40%.
The problem
Before an invoice gets paid, three documents have to agree: the purchase order (what was ordered), the delivery order or goods receipt (what arrived) and the invoice (what is being charged). Often they do not, because of a partial delivery, a price that moved, or a line split across two shipments, and someone in accounts payable has to work out which one is wrong. Done by hand, that means retyping every field and cross-checking three documents before month-end close. Rules and templates cannot keep up with how differently suppliers lay out their documents: a moved total or a scanned PDF breaks them quietly, and the failure shows up as a growing backlog rather than an error.
The solution
Built at Axrail, which sells it as KIP.ai, the product runs the whole job:
- Collect documents from wherever they already arrive: email, WhatsApp, shared folders, SFTP or an API, as PDFs, scans, Excel or CSV.
- Extract every field and line item, each with a confidence score. Anything low is held for a person before it reaches the ERP.
- Match the purchase order, delivery order and invoice in one pass. A mismatch becomes an exception with its reason: a quantity shortfall, a price variance or a missing reference.
- Approve under the customer’s delegation of authority, with every match, override and approval in an audit trail, then export clean data to Xero, QuickBooks, Oracle NetSuite or another ERP.
- The result: manual labour cut by over 40%, measured by the drop in staff hours spent on invoice matching each day. For the product as a whole, Axrail publishes 80% less time per document and 90% extraction accuracy.
- In production at FamilyMart Malaysia, turning supplier invoices that arrive by email into purchase orders for Microsoft Navision.
Textract first, then Bedrock Data Automation
The first version read invoices with Amazon Textract’s expense analysis alone, and it did well on clean, conventional invoices. Real supplier documents are messier: line items spread across columns, totals in odd places, poor scans, and judgement calls such as which of two dates is the due date. So the reading moved to Amazon Bedrock Data Automation, AWS’s generative AI service for turning documents into structured data. A blueprint for each document type sets the fields to pull, and every value comes back with a confidence score and its place on the page, so every figure can be traced back to where it came from. Claude on Amazon Bedrock then matches every line to its purchase order and goods receipt and explains in plain words why anything does not match. Rules in code check the totals, and anything uncertain goes to a person rather than into a payment run.