The aggregate-metric trap: a 97% overall accuracy figure can comfortably mask genuinely poor performance on one specific document type or field. Aggregate numbers hide exactly the segment that later embarrasses the team that shipped on the headline figure alone. Always analyze accuracy by segment before using it to justify reducing human review.
Three techniques that address this properly:
- Stratified random sampling of high-confidence extractions, not just low-confidence ones, for ongoing error-rate measurement and novel-pattern detection. Sampling only the low-confidence bucket tells you nothing about how the automated (high-confidence) path is actually performing — which is precisely the path carrying the most volume.
- Field-level confidence scores calibrated against a labeled validation set. Raw, uncalibrated model-reported confidence numbers are not usable as thresholds on their own — a "confidence" score only means something once it's been checked against ground truth and calibrated to it.
- Routing to human review on low confidence and on ambiguous or contradictory source documents — using both signals to prioritize limited reviewer capacity toward the cases most likely to actually be wrong.
The discriminator: segment before you automate. An aggregate accuracy number is not evidence that every segment is safe to automate — it's an average that can bury exactly the segment you'd most want to catch.
Traps to recognize:
- Sampling only the low-confidence bucket to measure error rate — this measures nothing about the high-volume automated path.
- Using uncalibrated, self-reported confidence as a threshold directly — the same underlying trap as escalation's unreliable self-reported-confidence proxy (§2) and Module 4's confidence-reporting caveat (§4.6), appearing here a third time.
- Reducing human review based on the headline accuracy figure alone, without a segment-level breakdown.