Human review workflows and confidence calibration

Overall accuracy can hide weak document types

97% overall accuracy can conceal one document type at 60%, or one field that is often wrong. Common document types have the greatest effect on the overall average, so weak performance on a rare type may be difficult to see.

Analyze accuracy by document type and by field before reducing human review. Check both the overall score and the performance of each important group.

The same 97%, broken out:

document type         volume   accuracy
standard invoice       9,400        98%
credit note              380        94%
handwritten receipt      220        61%
all types             10,000        97%

The arithmetic works: 61% on 220 documents barely moves an average dominated by 9,400. Two percent of your volume is wrong four times out of ten, and the headline number reports that as healthy.

Sample the confident cases

Reviewing only low-confidence extractions leaves high-confidence results unmonitored. Some of those automatically accepted results may still be wrong.

On an exam you re-check the questions you struggled with and skip the ones you were sure about. The marks you lose to carelessness sit in the second group, because certainty is what stops you looking. Sample the confident cases for the same reason.

Stratified random sampling of high-confidence extractions measures the error rate where you are not looking, and — more valuable — detects novel error patterns as documents change. A supplier changing their invoice layout produces confident, wrong extractions, and sampling is what surfaces it.

Stratification matters as much as randomness. A format representing 0.5% of traffic contributes about one document to a random sample of 200. Tripling the sample produces only about three examples at triple the review cost. Sample each important document type separately instead.

Two cheaper-looking substitutes don't work either. Reviewing whatever a downstream system rejects catches only errors that break something, and a plausible wrong value passes every structural check. Re-running old documents tests whether results are stable, and a stably wrong extraction looks perfectly healthy.

Validate confidence scores before using them

Have the model output field-level confidence scores, then compare those scores with known correct and incorrect results in a labeled validation set. This process, called calibration, shows what each confidence range means in practice and helps you choose review thresholds.

A calibration describes the documents it was fitted on. When a supplier redesigns its layout, the validation set holds no example of the new format, so refitting on that same set says nothing about these documents. New formats need newly labeled examples before their thresholds mean anything.

Route by capacity

Reviewer capacity is limited. Use calibrated confidence, source ambiguity, and business impact to decide which results need human review. Also review a sample of results that would normally be accepted automatically. This sample can reveal errors that the routing rules missed.