Confidence thresholds, field by field

The most common mistake we see in document processing pipelines is setting one confidence threshold for an entire document, instead of one per field. It seems like a simplification. It’s actually where straight-through processing rates go to die.

Here’s the problem with document-level thresholds: a single low-confidence field — a smudged date, an unusual formatting choice on one line — drags an entire otherwise-clean document into manual review, even though every other field extracted perfectly. At enterprise document volumes, that adds up to thousands of documents routed to a human reviewer for a single ambiguous character.

Field-level thresholds fix this. A patient name extracted at 99% confidence and a referring-provider fax number extracted at 61% confidence on the same document should be treated differently: the name gets accepted, the fax number gets flagged, and a reviewer corrects one field instead of re-keying the whole form.

This requires more setup work than a document-level threshold — you’re tuning per field, sometimes per document type, and you need enough sample volume to know what “normal” confidence looks like for each field before you can meaningfully flag “abnormal.” It’s also the single biggest lever we’ve found for straight-through processing rates that hold up once a pipeline meets real, messy documents instead of a clean sample set. On the Alderleaf Health Network engagement, this distinction was most of the difference between a pipeline that looked good in testing and one that hit a 91% straight-through rate in production.