Confidence as a first-class output: routing straight-through vs. review by number

Every automation business case for document extraction rests on one sentence: “high-confidence results clear without a human.” The entire ROI — fewer FTE-hours per submission, faster cycle times — depends on that being safe. Which means the confidence score isn’t a cosmetic number in the corner of a screen. It’s the control that decides what a person never has to look at. If it’s wrong, you either leak errors into production or you review everything and capture none of the savings.

So confidence in InsightXtract is a first-class output, engineered to be trustworthy and to route work automatically. Two things make it different from the confidence number most tools print.

What makes the number trustworthy

It’s grounded, not self-asserted

A model asked “how sure are you?” will happily say 0.95 about something it invented — self-reported confidence is notoriously uncalibrated. InsightXtract instead grounds confidence against the actual document evidence and the outcome of validation: does the value trace to a real span on the page, did it pass its type and cross-field checks, did it survive the re-extract loop? A value that reconciles and cites its source earns a high score; one that doesn’t, doesn’t.

It’s per field, not just per document

One score for a whole 38-field record is almost useless — it hides which values are solid and which are shaky. Every field carries its own confidence, so the system can clear the 36 fields it’s sure of and flag only the 2 it isn’t.

How confidence routes the work

Confidence and validation severity combine into an objective gate. You set the threshold; the engine routes each field and each document accordingly.

Condition	Routing
Field above threshold, all checks pass	Auto-accept — no human touches it
Field below threshold, or a warning	Flag the field for targeted review
An unresolved error-severity check	Hold the document for review
All fields clear	Straight-through — the document is done

A real document, routed

An excess-casualty submission: 38 fields extracted. Most clear; two need a look. The underwriter reviews two fields, not the whole packet.

The stamped effective_date and a total_incurred that only just reconciled are the two the agent is least sure of — so they’re exactly the two surfaced. Partial review, not all-or-nothing: the reviewer’s attention goes to the 5% of the record that carries the risk.

Why grounded, field-level confidence matters to the business

• Automation you can actually trust. Because the score reflects evidence and passing checks, “high confidence” genuinely means safe to auto-accept — so you capture the STP savings without leaking errors.
• Review effort follows risk. Humans see the handful of shaky fields, not walls of green ones. The cost of review scales with uncertainty, not volume.
• A tunable dial, measured. The threshold is a business setting: raise it for a conservative line, lower it where errors are cheap — and watch the straight-through rate and accuracy move together in the eval harness.
• Defensible decisions. Every routed field records its confidence, its evidence, and why it was flagged — the audit trail for an automated decision.