Route each field by the consequence of an error
Send an extracted field to a person when the field is required and missing, the document contains two plausible values, or a validation rule fails. Add a review rule for uncertain readings when the extraction tool returns a confidence score. Keep fields that authorize a shipment, payment, or other consequential action behind the applicable approval step even when the document is easy to read.
Decide what a wrong value could change in the next system. An optional note may be safe to leave blank. An unreadable order ID can send a valid update to the wrong record. Give each field a rule that names its purpose, required evidence and allowed destination. Keep the document itself and the extracted proposal separate so a reviewer can compare them.
Apply the rules before a destination write
- Identify the document and field. Record the document version, page and the source region or text associated with each proposed value. If the tool classified the wrong document type, hold the whole record.
- Check the value. Apply allowed formats, required fields and cross-field rules. Let a destination lookup confirm the intended record. Conflicts and missing evidence go to a person regardless of the score.
- Check reading uncertainty. If the tool returns a score, use a threshold defined for that field and model version. Define how to handle a missing score. Treat the score as a routing input. Only a relevant trial can tell you how often an accepted value is correct in your workflow.
- Release the right scope. Save an internal draft when the rules permit it. Hold the action that depends on a reviewed field. Make the approval refer to the exact document version and corrected values.
Microsoft documents field confidence and notes that some fields do not return a score. AWS documents separate human-review conditions in its Amazon Textract integration for missing keys and low confidence for existing Amazon Augmented AI customers. AWS no longer offers that review service to new customers. Use the documentation for your exact model and configuration. A source conflict such as two delivery dates still needs a rule in your workflow.
Worked example with a fictional dispatch team
Juniper Cartage is a fictional company. It extracts an order ID, requested delivery date and optional dock note from delivery-change forms. Every shipment change still needs dispatcher approval. Automatic acceptance below means the extracted fields can fill a draft. The tool scores and the 0.90 test threshold are made up. The routing results are expected outcomes for testing.
| Form | Proposed fields and evidence | Expected route |
|---|---|---|
| JC-01 | Known order O-402. One delivery date at 0.96. Optional note blank. | Draft ready for dispatcher approval. Leave the note blank. |
| JC-02 | Order ID missing. One delivery date at 0.98. | Hold for missing order ID. Do not update a guessed order. |
| JC-03 | Known order O-402. Header and amendment show different dates. The extracted date has score 0.99. | Hold for conflicting dates. Show both source regions and confirm which instruction applies. |
| JC-04 | Known order O-402. One date at 0.89. | Hold for low reading confidence. Ask the dispatcher to check the source. |
| JC-05 | Known order O-402. One date with no score. | Hold for missing confidence. Record the reason separately from low confidence. |
A reviewer may confirm, correct or reject the proposal. For JC-03, the dispatcher records the accepted amendment and the evidence it applies. Keep the original proposal, the resolved date and the review reason. If a new amendment arrives before approval, reopen the decision against the new version.
Give the reviewer enough evidence to finish
Put the proposed value beside the matching page and region. Include the original field label, the rule that blocked it, and any competing value. Give the reviewer one owner and a fallback when the source stays unclear. Log who resolved it, which version they reviewed, what they chose and what the destination accepted. Keep only the source and review history that the job requires.
Before launch, include missing scores, readable but wrong values, two conflicting fields and a new version received after approval. Follow one safe item through to the destination and confirm that every held item leaves the destination untouched. Review a sample of draft-ready items as well so you can find bad values that escaped the rules.
Use a field review worksheet
DOCUMENT FIELD REVIEW Document type and version: Field and destination action: Required evidence and page/region: Required / optional: Format, lookup and conflict rules: Confidence scale, field threshold and missing-score rule: Draft allowed or held: Approval owner and fallback: Allowed correction and reason to record: Pass case / expected result: Missing value or missing score / expected result: Conflicting high-score value / expected result: New document version after approval / expected result: Actual downstream write and evidence:
Use one worksheet row or copy for each field with a different rule. Read the required intake fields guide to define the field contract before adding extraction. Bring a fictional or redacted form to an Internal Tools fit call if you want help building the review queue.