Why every field gets its own confidence score
A 94% on a bill does not tell you which of its eight fields is the shaky one — so you check all eight anyway.
The averaging problem
Most document AI reports one confidence number per document. It looks reassuring and it is nearly useless, because it is an average over fields that are not remotely equally certain.
A bill where the vendor, date and total were read cleanly off a sharp PDF, and the invoice number was read off a smudge, might average to 94%. That number describes no field on the document. The three settled values are near-certain and the fourth is a coin toss, and the average hides exactly the one you needed to see.
So you do what any careful reviewer does with an unspecific warning: you check everything. Which is the work the extraction was supposed to remove.
What per-field scoring changes
Score each field separately and the picture inverts. Seven values are settled, one is flagged, and your attention has somewhere specific to go.
That is the entire mechanism by which review actually gets faster. Not better OCR — better reporting of what the OCR was unsure about. The extraction quality can be identical; what changes is whether you can act on it selectively.
It also changes what a low score means. Under averaging, a low number is a vague signal that the document is bad. Per field, it is a precise instruction: look here, not there.
The score has to be visible where the work is
A score that lives in an export, or behind a detail view, or in a log, does not change anyone's behaviour. It has to sit next to the value, in the queue, at the moment of approval.
We put the flagged field where you cannot miss it and leave the settled ones quiet. A reviewer should be able to tell, without reading, roughly how much of their attention this document is about to want.
What we deliberately do not do with it
We do not auto-approve above a threshold by default. The gate stays: every document reaches a human review queue regardless of how confident the extraction was.
You can set thresholds for what gets flagged and what needs a second pair of eyes, because risk appetite genuinely differs between firms. What is not configurable is whether a human sees it at all. A confidence score is a tool for directing attention, not a licence to skip the review.