Measure the fields your workflow depends on
Test invoice OCR or AI extraction by comparing its output with manually checked answers from your own documents. An overall text-recognition score can hide a wrong total or missing item. Measure critical invoice fields separately, then measure the time needed to review and correct the result.
This is a proposed evaluation method, not a completed Parzable benchmark. No accuracy percentage, processing-time guarantee, or supplier-coverage claim is established by this page.
The FinCriticalED research paper examines financial OCR at the fact level, reinforcing the value of checking the facts that matter rather than relying only on character-level similarity. Its results are not measurements of Parzable.
Build a reference set before running the tool
- Select documents that represent your normal suppliers, page counts, currencies, and scan quality.
- Include difficult cases deliberately: unclear dates, multiple totals, credits, and tables crossing pages.
- Have a reviewer record the expected values from each source. Mark genuinely unreadable fields as unresolved rather than inventing answers.
- Keep some documents aside while you tune schema descriptions. Evaluate the final schema on those unseen examples too.
- Use the same files, fields, and review rules when comparing runs or tools.
Start small enough to inspect every result. The purpose is to discover the failure modes in your workflow, not to find only the clean examples that make the tool look good.
Use a scorecard with separate error types
Download the blank extraction evaluation scorecard. Record filename, field, expected value, actual value, outcome, and review seconds. Suggested outcomes are correct, wrong, missing, unresolved-source, and processing-failed.
| Metric | How to compute it | What it reveals |
|---|---|---|
| Field accuracy | Correct comparable values ÷ all comparable expected values | Report separately for number, date, currency, and total. |
| Document pass rate | Documents with every required field correct ÷ documents evaluated | How often the whole required record is correct. |
| Missing item count | Expected item rows absent from output | Errors hidden by correct header fields. |
| Review effort | Total review and correction time ÷ documents reviewed | Work remaining after extraction. |
Include processing failures in the report rather than dropping them from the sample. Report unresolved source fields separately so unreadable input is visible without silently excluding hard documents.
An illustrative calculation, not a product result
Suppose a fictional test has 20 readable invoices and seven required fields each: 140 comparable field values. If 133 values match the reviewed reference, field accuracy is 133 ÷ 140 = 95%. If the seven errors occur on seven different invoices, only 13 of 20 invoices have every field correct: a 65% document pass rate.
Both figures describe the same fictional output. That difference explains why a high field average can coexist with substantial review work. Break down the seven errors by field and cause before changing the schema or upload process.
Set your acceptance criteria before the pilot: which fields must be checked, how exceptions will be handled, and how much review effort is acceptable. Rerun changed descriptions on held-out documents and retain the results. A useful decision is grounded in your measured workload, rather than an accuracy promise borrowed from a different dataset.
Put the workflow to work
Try your own invoice in Parzable
Create an account, define a schema, and test a small sample. Review extracted values before using them. See current plans and page allowances.
Try Parzable