Define the decision before the test
Specify the document family, required fields, expected repeating rows, desired output and next business step. Agree what constitutes a correct value, a permissible missing value and a review-required exception.
Prepare authorised samples reflecting real variation. Separate synthetic evaluation files from production customer jobs.
Compare expected and actual values
Record expected values independently before inspecting the extraction. Count checked required fields, correct values, missing fields and ambiguous cases. Keep document-level acceptance separate from field-level correctness.
For example, 98 correct values out of 100 checked required values is 98 percent for that test set under its stated rules. It is not a product-wide accuracy guarantee and says nothing by itself about unchecked rows.
Measure the complete workflow
| Measure | What to record |
|---|---|
| Completeness | Required fields and expected table rows present |
| Correctness | Agreement with independently reviewed values |
| Review effort | Time and number of corrections before acceptance |
| Reliability | Failures, retries and ambiguous outcomes |
| Output fit | Whether the receiving process accepts the format |
| Repeatability | Results after mapping or template revision |
Test exceptions, not only the happy path
Include missing identifiers, page-spanning tables, unclear number formats and known rule failures. A useful system leaves the exception visible rather than silently inventing a pass.
Agree a bounded pilot and record evidence of both success and failure. Retest the same representative set after a change.
Decision checklist
- The documented capability matches the tested environment.
- Security, retention and integration questions have named owners.
- Important exceptions have a clear review path.
- Commercial allowances match expected usage.
- No unsupported performance or compliance claims remain.
Evaluate your document workflow
Start with representative permitted files, the fields you need and a clear review decision.