Teams often compare tools using one headline percentage. To make a real decision you need to know what that percentage counts.
Field-level accuracy
This counts each extracted value separately. If a model reads 98 of 100 fields correctly, field-level accuracy is 98%. It is useful for finding weak spots, such as handwriting or low-quality scans.
Document-level accuracy
This counts a document as correct only if every required field is right. It is stricter, and it is the number that predicts how many documents can flow straight through with no human review.
Straight-through rate
The metric operations teams care about most is the share of documents that pass validation and reach your system untouched. Track it weekly and watch it climb as reviewers correct the model.
A simple test plan
- Collect 100 real documents across your messiest sources.
- Label the fields that matter to your workflow.
- Run each candidate tool and compare field-level, document-level and straight-through results.
- Note how much effort it took to get each tool running.
The tool that wins on your files is the right one, whatever the marketing page says.
