Invoice OCR and Three-Way Matching: What to Ask

Most AP teams don't have an OCR problem; they have a matching problem. The invoice gets read reasonably well, then someone still opens the PO in one tab and the GRN in another and retypes line items to check them. If your team clears more than a few hundred vendor invoices a month, that gap between extraction and matching is where the hours quietly go — and it is the part most vendor demos skip past in favour of a slick upload screen.

What does three-way matching actually compare?

Three-way matching compares the purchase order, the goods receipt note (or service completion record) and the invoice, line by line, before payment goes out. Practically, the software should check quantity ordered against quantity received against quantity billed, unit price against the PO rate, tax treatment and freight, and flag differences outside a tolerance band you configure — say 2% or ₹500, whichever is lower. Ask one direct question: "Do you match at line level or header level?" Header-level matching only compares invoice totals to PO totals, which passes a lot of wrong invoices. If the vendor cannot show you a screen with mismatched line quantities and the exception routed to a person, the feature is a checkbox, not a control.

How should you test OCR accuracy before you sign?

Test on 50 to 100 of your own invoices with known-correct values, not on the vendor's curated sample set. Build the set deliberately: multi-page invoices, a rotated scan, a bad photocopy, a 25-line-item GST invoice, a credit note, and one handwritten GRN. Score field by field — supplier GSTIN, invoice number and date, HSN/SAC, line quantity, taxable value, the IGST or CGST/SGST split, and total. A vendor claiming 99% accuracy usually means 99% of characters, not 99% of invoices cleared straight through. Watch closely how the system behaves at low confidence: does it route the field to a human with the source image highlighted, or silently guess? That workflow design matters more than the model behind it — see how document AI pipelines handle low-confidence extraction.

Which numbers should a vendor commit to in writing?

Get three numbers defined on your sample and written into the contract: field-level accuracy, straight-through processing rate, and duplicate catch rate. Field-level accuracy is the share of extracted values that are exactly right. Straight-through processing (STP) is the share of invoices that reach "ready to pay" with no human touch at all — 60-80% is a realistic first-year range for messy mid-market data, and anyone promising 98% on day one has not seen your scans. Duplicate catch rate matters because duplicate payments are the expensive failure. Tie the STP commitment to a remedy clause: if the vendor promised 70% and delivers 45% at month four, you want a documented price or exit conversation, not a support ticket.

How is this priced, and where do costs creep in?

Quotes in India typically land between ₹5 and ₹20 per invoice, or roughly ₹1 to ₹4 per page, and the per-unit price is rarely the whole bill. Ask separately about implementation fees, per-vendor-format template setup, ERP or accounting connector charges, reprocessing of failed documents, and storage. Run the arithmetic on your own volume: at 4,000 invoices a month and ₹10 each, you are at ₹40,000 a month, or ₹4.8 lakh a year. Your team, meanwhile, spends around 7 minutes per invoice on manual keying and matching — 460 hours a month, close to three full-time people. That comparison is the real business case, so make vendors price against it rather than against a competitor's per-page rate.

What India-specific compliance questions matter?

Ask how the system handles GST e-invoice IRN capture and GSTR-2B reconciliation, because matching your purchase register to 2B is half the reason finance teams want this at all. Then cover TDS — does it read and apply the right section (194C, 194J) per vendor, and does it flag the MSME 45-day payment deadline before interest starts accruing? On data, the DPDP Act applies to the vendor contact details sitting on invoices, so ask where documents are stored, whether they are used to train shared models, and what the deletion process looks like. If the vendor cannot answer these three in plain language, budget for GST billing automation work as a separate line item.

Where to start

Pick 60 invoices you have already processed, write down the ten fields you care about and your tolerance rules on a single page, then run two vendors cold on the same set and score the output yourself. You will learn more in a week than from a month of demos. If you would rather have someone run that bake-off alongside you, that is the kind of work Maysan Labs does with finance teams before the build starts.