Three cloud vendors sell invoice extraction. None of them publishes how accurate it is.
Microsoft, AWS and Google all sell invoice extraction and none publishes an accuracy figure for it; each returns a confidence score per field and leaves the measuring to you. The academic benchmarks that do publish numbers are receipts, not Australian tax invoices. The independent tests are small. Here is all of it, and how to test on your own bills.
Start with what the vendors do not publish
Microsoft’s documentation for its prebuilt invoice model says it "extracts key information such as customer name, billing address, due date, and amount due" and "currently supports invoices in 27 languages". It gives no accuracy figure. Microsoft, Document Intelligence invoice model documentation, updated 15 August 2026. AWS documents 44 standard field types for its expense analysis, including VENDOR_ABN_NUMBER and RECEIVER_ABN_NUMBER, and returns "the confidence score associated with each piece of data detected". It gives no accuracy figure. AWS, Textract invoice and receipt documentation, fetched September 2026. Google’s Document AI processor list was not retrievable when we checked in September 2026, and we found no published accuracy figure for its invoice parser anywhere else.
This is not evasion so much as honesty of a limited kind. Accuracy depends on the documents, and a vendor cannot know yours. A confidence score per field is the right primitive: a well-built system posts the confident fields and routes the doubtful ones to a person. But it means the question this page was commissioned to answer has no vendor answer, and anyone quoting one has either measured it or made it up. The rest of this page is everybody who has measured something, with the size of the sample beside the number.
The academic benchmarks are receipts, and the good scores are old
The benchmark most extraction papers cite is SROIE, from a 2019 competition: 1,000 scanned receipts, 600 for training and 400 for testing, with four key fields — company, date, address and total. Eighteen valid entries were submitted to the extraction task. Only one scored above 90%, at 90.49%. Huang et al., ICDAR 2019 SROIE competition report, competition 2019, report March 2021.
The other is CORD: 1,000 receipts with 30 labels in four categories, split 800, 100 and 100. LayoutLMv3, the 2022 model from Microsoft Research, reached an F1 of 96.56% in its base size and 97.46% in its large size on CORD, and 90.29% and 92.08% on the FUNSD forms benchmark. Huang et al., LayoutLMv3, April 2022.
Those are real, reproducible numbers, and neither describes an Australian tax invoice. A receipt has one merchant, one date, one total. A supplier bill has an ABN to validate, GST to reconcile against a total, a due date that may be a term rather than a date, and a table of line items that the receipt benchmarks barely test. More recent work argues that combining OCR, layout models and language models handles variable layouts and poor scans better than template-based OCR (Automated Invoice Data Extraction: Using LLM and OCR, arXiv 2511.05547, November 2025, revised January 2026), but publishes no field-level figures in its abstract that we could quote.
The independent tests that exist, and how small they are
An analyst firm ran six services — Amazon Textract, Google Document AI, Microsoft Document Intelligence, Rossum, Docsumo and Claude Sonnet 3.5 — over "more than 400 key value pairs from 20 publicly available invoice samples", scoring correct pairs over total pairs, and reported Claude with the highest overall accuracy without publishing the percentages in text. AIMultiple invoice OCR benchmark, tests run December 2024, page updated 4 August 2026. Twenty invoices is a demonstration, not a benchmark.
A software consultancy tested scanned invoices with no text layer, across at least three layouts and dates from 1971 to 2020, and published field-level and line-item scores. GPT-4o with OCR reached 98.0% on fields and 57.0% on line items; Azure Document Intelligence 93.0% and 87.0%; GPT-4o on the image alone 90.5% and 63.0%; Google Document AI 82.0% and 40.0%; AWS Textract 78.0% and 82.0%. Businessware Technologies benchmark, January 2025. It did not state how many invoices.
The pattern inside that table is the most useful thing on this page. The tools best at header fields were worst at line items, and the reverse. A model that gets the total right and the items wrong reconciles to the bank and mis-codes every job. Whether that matters depends on what you do with the data next, which no benchmark knows.
| Header fields | Line items | |
|---|---|---|
| GPT-4o with OCR text | Yes 98.0% | No 57.0% |
| Azure Document Intelligence | Yes 93.0% | Yes 87.0% |
| GPT-4o, image only | Partly 90.5% | No 63.0% |
| Google Document AI | Partly 82.0% | No 40.0% |
| AWS Textract | No 78.0% | Yes 82.0% |
Field accuracy is not document accuracy, and the vendors know it
Here is the arithmetic every accuracy claim hides. If a model gets each field right 98% of the time and a bill has eight fields you need, the chance every field is right is 0.98 to the eighth power — about 85%. At 95% per field it is 66%. So a "98% accurate" system sends one bill in seven to a human, and a "95% accurate" one sends a third of them. Neither number is on the brochure.
The vendors measure this as straight-through processing — the share of documents needing no human touch — and even they describe the gap. One vendor writes that "a 95 percent field accuracy does not translate directly to a 95 percent straight-through rate", puts vendor claims at 95 to 99% on clean digital invoices against 80 to 92% straight-through in production on mixed documents, and cites a competitor at 75 to 85%. Hypatos, a vendor in this market, updated 5 May 2026. That is a vendor describing its own market, and we quote it as one.
Rossum’s own case studies claim "90% accuracy after only 10 documents" for one customer, 92.6% after 20 and a 93.4% average for another, and a 71% straight-through rate for a third. Rossum customer case studies on its own site, page dated 21 May 2026. Those are honest-sounding numbers precisely because they are not 99%. Any pitch that promises more than the vendor’s own website does is a pitch.
For scale, recall the ATO’s model of the manual process it is trying to replace: purchase-invoice exceptions on 3.6% to 48% of invoices, costing 5 to 20 minutes each. ATO, Peppol eInvoicing value assessment — cost calculations, page updated 17 October 2025. An extraction system that routes 15% of bills to a person is not failing. It is doing roughly what the office already did, faster, and the question is whether the other 85% were posted correctly.
What an Australian tax invoice demands that a receipt benchmark never tests
A tax invoice for a sale under $1,000 must show seven things: that it is intended to be a tax invoice, the seller’s identity, the seller’s ABN, the date of issue, a description of what was sold with quantity and price, the GST amount, and the extent to which each sale is taxable. For $1,000 or more it must also show the buyer’s identity or ABN. A tax invoice is not required for sales of $82.50 or less, and a seller must provide one within 28 days of being asked. ATO, Tax invoices, updated 25 August 2025.
Two of those are checkable, and a system built for Australia should check them rather than trust the read. An ABN can be validated against the ABN Lookup register, which turns a plausible eleven-character string into a confirmed supplier or a flag. A GST amount can be checked against the total, and if the two do not reconcile the bill goes to a person regardless of what the confidence score says. Neither check appears in any benchmark above, and both catch more errors on real bills than another point of model accuracy would.
The line-item problem is the third Australian wrinkle. For a trade business, posting a supplier bill is not one number to the bank; it is materials to a job so the job’s margin is right. A system that reads the total perfectly and the items poorly reconciles cleanly and quietly wrecks job costing. That is why the field-versus-line-item table above matters more than the headline, and why accounts payable treats coding to job as the hard part, not the reading.
- KNDR·01Seven mandatory fields, eight over $1,000The ATO’s list. If the model cannot find one, the document is not a valid tax invoice for your BAS, whatever the confidence says.
- KNDR·02Validate the ABN, do not read itABN Lookup confirms or flags. A misread digit becomes a lookup failure, not a wrong supplier.
- KNDR·03Reconcile GST to totalArithmetic catches misreads the model is confident about. Route mismatches to a person.
- KNDR·04Line items to jobsThe part receipts never test and trade businesses depend on. Score it separately.
How to test it on your own bills in a day
Take 100 of your own supplier bills from the last quarter — real ones, including the photographed dockets and the handwritten ones. Label eight fields by hand: supplier, ABN, invoice number, date, due date, subtotal, GST, total. Add the line items for twenty of them. That is your ground truth and it takes a morning.
Run them through whatever you are considering, and score three things separately. Field accuracy: fields right over fields attempted. Document accuracy: bills with every field right over bills. Straight-through rate: bills that would have posted with no human touch given whatever confidence threshold you choose. Keep the twenty worst failures in a folder; they are worth more than the average, because they tell you which suppliers to fix at source.
Then compare document accuracy to what your office does now, estimated from the manual data entry page and the ATO exception rates. If the machine posts 80% clean and routes 20% to a person who previously handled 100%, that is the number the decision rests on. The Xero page covers how a draft bill differs from a posted one, which is where confident-but-wrong fields should land. What not to automate is the page to read if your worst-failures folder turns out to be one supplier.
What we have not measured, and what would change this page
Nothing above is a Kindra measurement. This page was specced as one — first-party data on field-level and document-level extraction accuracy on Australian supplier bills, licensed for anyone to reuse with attribution. We have not collected it, so the page publishes what is published instead, with the dates on.
What would change it is not a survey. It is a labelled set of real Australian tax invoices — including the bad scans — scored field by field and by straight-through rate against each service, with the failure cases published, on real Australian businesses, for long enough that the numbers stop moving. When that exists it gets published here with the sample size, the method and the cases where it failed, and the borrowed figures above come out.
If our own numbers turn out smaller than the ones vendors quote, they still get published. That is the point of measuring instead of asserting. Until then, the arithmetic section above runs on your figures, which are worth more than anybody’s average.
The questions people actually ask.
01Why is there no Kindra figure on a page filed under research?
Because we have not measured one, and a number you cannot show the working for is a marketing claim with a chart. What we could do honestly was find everything that is published, date it, put the sources that disagree next to each other and mark where the data stops. That turned out to be more useful than the page we set out to write.
02A vendor told me 99% accuracy. Are they lying?
Probably not, but ask which 99%. Per character, per field, per document, and on whose documents? Clean digital PDFs from a large supplier score very differently from a photographed docket. The vendor whose own website quotes 90 to 93% for named customers is being straighter with you than a salesperson quoting 99%, and the arithmetic section above shows why a per-field figure overstates what lands clean. Ask for straight-through rate on documents like yours, or test it yourself in a day.
03Have you built this and measured it?
We have built against Xero, and we have not published an accuracy measurement for extraction on Australian bills, so there is no Kindra number on this page. What we can say is how we would measure it — the hundred-bill test above — and that the thing we score hardest is line items to jobs, because that is where trade businesses lose money quietly. When we have a measurement worth publishing it will appear here with the sample and the failures.
Stop bending.
Start shipping.
Bring the one flow that keeps costing you a Tuesday afternoon. We map it live and tell you what we'd build, what we wouldn't, and what it costs.