Guides · 2026-07-24

Invoice OCR API accuracy: what the numbers mean, and why almost nobody publishes them

Every invoice OCR API claims high accuracy. Almost none tell you what was measured, on which documents, or field by field. If you are choosing an API to feed invoice data into Xero, MYOB, or your own AP workflow, that gap is where the cost hides — because the errors that hurt are not evenly distributed.

“99% accurate” is not a measurement

An accuracy claim only means something when three things are stated with it: the benchmark (which documents, how many, how representative of real scans), the unit (per character, per field, or per document), and the breakdown (which fields, not just the average). A model can read printed totals almost perfectly and still misread a third of handwritten vendor names — a single blended percentage hides exactly that.

For accounts payable the distinction is not academic. A wrong total becomes a wrong payment. A wrong date lands GST in the wrong BAS period. A wrong vendor name is mostly an annoyance — unless it masks a look-alike vendor, which is a fraud problem, not an OCR problem. You need to know the error rate of each field you act on, not the average of all of them.

Five questions to put to any invoice OCR vendor

1. Which benchmark, exactly? A named, public dataset of real scanned documents can be re-tested by anyone. “Internal testing” cannot.
2. What are the per-field numbers? Including the weak ones. A vendor willing to publish its worst field is telling you something a marketing average never will.
3. Is there a confidence score on every value? Accuracy is a population statistic; confidence is what lets you route this invoice to a human or straight through.
4. What is the straight-through rate? The share of documents needing no human touch is the number that actually sets your labour cost.
5. Where are the numbers published? If they live on a page that can quietly change — or nowhere — they will not survive the vendor's next model update.

The numbers we publish

DVAP is measured on SROIE, a public benchmark of real scanned receipts, and the results are served from the live API — the same numbers are shown on the product page, sourced from the same endpoint:

Vendor name at 88% is the honest part of that table. Free-text fields are harder than amounts and dates for every extraction model on the market; the difference is whether the vendor tells you. Because every DVAP value carries its own confidence score, the 12% of vendor names the model is unsure about arrive flagged — routed to a reviewer instead of flowing silently into your ledger.

Accuracy is the floor, not the product

Extraction is where invoice automation starts, not where the money is lost. The invoices that cost real money are the ones that read perfectly: the duplicate submitted twice, the tampered bank details, the line items that never matched a purchase order. DVAP runs that whole pipeline in one call — extract, validate (ABN checksums, GST arithmetic), screen, match, audit-trail — which is why we measure and publish the pipeline, not just the OCR.

Frequently asked questions

What accuracy should I expect from an invoice OCR API?

It depends on the field. On DVAP's published SROIE benchmark, dates reach 100% and totals 99%, while vendor names sit at 88% — free-text fields are genuinely harder. Any vendor quoting one site-wide number without per-field detail is averaging away the fields most likely to hurt you.

What is the difference between field-level and document-level accuracy?

Field-level accuracy measures how often a single field (total, date, ABN) is extracted correctly. Document-level — often reported as a straight-through rate — measures how many invoices need no human correction at all. A vendor can have 99% field accuracy and still need a third of documents touched by a human.

What is a straight-through rate?

The percentage of documents processed end to end with no human touch. DVAP's published straight-through rate is 87%: 87 of 100 benchmark documents needed no correction, and the rest were routed to review by their confidence scores rather than failing silently.

Why don't most invoice OCR vendors publish accuracy numbers?

Because a published number is a commitment. It names the benchmark, exposes the weak fields, and invites re-testing when the model changes. Unpublished 'up to 99%' claims carry no such risk — which is exactly why they tell you so little.

How is DVAP's accuracy measured?

Against SROIE, a public benchmark of real scanned receipts, scored per field. The results are served live from the API itself — not a marketing page updated by hand — so what you read is what the current model actually does.