Separate confidence from business approval

A high confidence score should support only the decision for which it is meaningful. In document recognition, it may describe certainty about a recognised value. It does not thereby authorise a price change or order approval.

This is not a blanket requirement to review every document manually. It separates recognition quality from authority for the subsequent action.

A correctly read price leaves a question open

The example is fictional. Quotation K-7 states a unit price of 24 EUR. A recognition system captures 24 EUR with a field score of 0.98. The price record stored as valid contains 20 EUR.

Recognition may be entirely correct. The difference is between two business records. A higher recognition score does not explain whether K-7 contains a new agreement, the price record is outdated or the quotation first requires review. Here, 0.98 is an example value, not measured ASES performance or a guarantee of the entire transaction's correctness.

If the flow overwrote the stored price solely because of that score, it would embed a decision about agreement status and authority into a reading assessment. That additional decision needs its own rule.

The assessment needs a clear reference

Microsoft distinguishes word and field confidence in document recognition, among other levels. Even these assess different things. A single value cannot therefore stand for every document field's quality without a definition.

For K-7, the flow can capture the recognised price while holding the business decision for clarification. The responsible person receives the quotation and affected price record. They need a way to confirm the new agreement or return the case. Simply clicking “Everything is correct” offers little help if the conflicting information remains invisible.

Thresholds can still be useful

For a low-impact sorting task, a tested threshold may allow suitable automatic routing. A narrowly bounded document flow can also operate automatically with business checking rules. What matters is that the selected rule has been examined for the actual inputs and its error consequences are acceptable.

The assessment of K-7 would change if an authorised person confirmed a new agreement for 24 EUR. The price may then be valid, and the old record needs updating through the defined procedure. That confirmed basis changes the decision. A higher reading score alone does not supply it.

Sources

• Microsoft: Accuracy and confidence

• IBM: Human-in-the-loop