Confidence Threshold
The line in the sand: above it, automation proceeds; below it, a human takes a look.
A confidence threshold is the cutoff score that determines what happens to an AI system's output: results scoring above the threshold flow through automatically, while results below it are flagged for human review, retried with a different model, or rejected. In document processing, thresholds are the mechanism that converts a model's self-reported uncertainty into an operational decision — they are where statistics meets workflow.
Setting the threshold is a business decision disguised as a technical one, because it trades off automation rate against error rate. A high threshold means fewer mistakes slip through but more documents queue for human review; a low threshold means higher straight-through processing but more errors reaching downstream systems. The right answer differs by field, not just by document: a misread invoice line description may be tolerable at 85% confidence, while a misread account number or date of birth on a KYC document is not tolerable at 99%. Mature systems therefore set thresholds per field, calibrated against measured accuracy on labeled evaluation data rather than chosen by intuition.
One caveat matters in practice: raw model confidence is not the same as probability of being correct. Models can be systematically overconfident, particularly on inputs unlike their training data — a degraded fax, an unusual layout, a new language. Calibration techniques align scores with observed accuracy, and ongoing monitoring detects drift, so that "0.97 confidence" continues to mean what the threshold assumes it means. Without that discipline, a threshold provides a false sense of control rather than a real one.
Knowing what you don't know — the models that estimate how likely each extraction is to be right.
The model does the reading; a person checks its work — but only where the model isn't sure.
The document arrives, gets read, gets validated, and lands in the system of record — and no human ever touched it.
Proof Perimeter runs document AI inside your own perimeter — with a provenance record on every field.
Book a demo