3-way matching is the process of verifying that a vendor invoice matches both the original purchase order and the goods receipt (or service confirmation) before approving payment. The three documents are: (1) the purchase order showing what was ordered at what price, (2) the goods receipt showing what was actually received, and (3) the vendor invoice showing what the vendor is billing. A clean 3-way match means all three documents agree on items, quantities, and prices within configured tolerance. Discrepancies between any two documents create an exception that requires human review before payment is released.

The agent catches: price discrepancies (invoice unit price differs from PO price by more than the configured tolerance), quantity discrepancies (invoice quantity exceeds goods receipt quantity), item code mismatches (invoice references an item not on the PO), duplicate invoices (same invoice number or amount from same vendor within a lookback period), missing PO reference (invoice arrives with no matching PO in the system), and tax or freight charges not on the original PO. Each exception type is given a reason code and routed to the appropriate reviewer.

OpenAI's vision and document parsing capabilities extract structured data from invoice PDFs — including vendor name, invoice number, invoice date, line items (description, quantity, unit price, extended amount), tax amounts, freight charges, and total amount. The extraction handles multiple invoice formats without templates because the AI reads the document layout contextually rather than using fixed field coordinates. Extraction accuracy is verified by cross-checking the extracted total against the sum of line items before any matching logic runs.

The agent integrates with any ERP that exposes a query API or can export PO and goods receipt data in a structured format. Pre-built integrations exist for NetSuite, SAP Business One, QuickBooks Online (via QBO API), and Microsoft Dynamics 365. For ERPs without direct API access, the agent can work from scheduled exports to a shared folder or Google Sheet. The key data required from the ERP is the open PO line items and any goods receipts posted against those POs.

When a discrepancy is detected, the agent classifies the exception by type and severity (dollar amount of variance), generates a plain-language summary of the discrepancy, and routes it to the designated approver via Slack or email. The summary includes the invoice number, vendor name, discrepancy type, dollar amount, and a recommended action (contact vendor for credit memo, approve and pay with variance notation, or hold for PO amendment). The approver's decision is logged back to the system and the invoice status is updated.

Yes. Because the extraction uses AI document understanding rather than template-based OCR, it handles any vendor's invoice layout — including multi-page invoices, invoices with complex line item structures, invoices with merged cells or unusual formatting, and invoices in languages other than English. The system handles both digital PDF invoices and scanned paper invoices converted to PDF. Processing accuracy improves as the agent sees more invoices from each vendor over time.

For a company processing 500 invoices per month at $12 average manual processing cost, the total manual cost is $6,000 per month or $72,000 per year. Automated matching reduces the per-invoice cost to approximately $2 for clean matches (roughly 70 to 80% of total volume) and $5 to $6 for exceptions requiring human review. Total automated cost for the same volume drops to approximately $1,500 per month — a 75% reduction. Additionally, faster processing reduces late payment penalties and enables capturing early payment discounts that manual teams miss.