Pilot air-waybill data extraction with field-level review

A clean extraction can still update the wrong shipment

An OCR service reads an air-waybill number correctly, but the integration attaches it to the wrong house shipment. Another upload contains an amended weight, yet the system treats it as a new booking. Neither failure is explained by a simple character-accuracy score.

A useful pilot evaluates the full path from source document to reviewed field to acknowledged destination record. Its first objective is controlled reduction of re-entry within one operational scope. Automatic extraction supplies proposals; authorised reviewers decide what may leave the staging area.

The Amazon Textract — Best practices advises choosing confidence handling according to the sensitivity of the use case. A service score can help prioritise review, but it cannot establish the correctness of a shipment relationship or provide universal permission to export a field.

Set the boundary before collecting the sample

Choose one forwarder branch, a named group of AWB layouts and one destination shipment system. Define whether the sample includes master documents, house documents or both, and how their relationships are represented. Exclude document families whose mapping has not been approved.

Use authorised copies of original documents, a restricted review store and a test destination without live booking or filing effects. Identify the operational sponsor, document-review lead, integration owner and person authorised to approve shipment changes.

The first stage should produce reviewed staging records only. A later supervised export stage can be considered after reconciliation, but the pilot does not require unattended updates or automatic creation of operational jobs.

Freeze a consecutive baseline sample that includes legible originals, poor scans, amendments, repeat uploads and relationship exceptions. Record manual entry time, checking time, corrections and unresolved cases. Keep the sample membership fixed so difficult documents cannot disappear from the comparison.

Map each field to its meaning and source

The field map should identify source location, destination field, unit, formatting rule, validation rule and approval authority. Preserve the raw extracted value alongside the reviewed value. A correction should show who made it and why.

The IATA — ONE Record describes shared air-cargo data models and exchange mechanisms. Those can inform consistent field meaning, but they do not determine the mapping for a particular legacy document or destination application.

If a field is absent or illegible, retain an explicit unresolved state. Do not fill it from a neighbouring document unless an authorised rule identifies that source and preserves its provenance.

Worked example: 150 uploads represent 120 business cases

Consider an illustrative test set with 120 original AWB document versions, ten later amended versions for ten of those same business cases, and twenty repeated uploads of versions already present.

The ten amendments add versions, not ten new business cases. The twenty repeats should be retained as intake events linked to their existing versions rather than generating additional operational jobs.

For case A75, original version V1 states 900 kg for the relevant gross-weight field. Approved amended version V2 states 940 kg for that same scope. The pilot must preserve V1, identify the 40 kg change and obtain the required review of V2. It should not mix V2's weight with an incompatible relationship or routing value extracted from another case.

For a separate field-quality assessment within this same sample, suppose reviewers check five defined critical fields on the current approved source version of each of the 120 cases. That gives 600 expected field values. If the raw extraction has 588 correct values and twelve errors, its measured accuracy on that defined set is 98%.

That percentage says nothing about whether the twelve errors include a dangerous association mistake. Report the errors by field, document condition and consequence. The proposed export gate remains review of every critical exported value and the shipment relationship, including high-confidence values. These figures are constructed test arithmetic, not observed product performance.

Design review around disagreement

The reviewer should see the original document region, extracted proposal, current system value and any relevant earlier version together. Highlight changes without concealing the rest of the source needed to interpret them.

A low-confidence field can enter a review queue, but high confidence must not bypass a critical-field gate in this pilot. A visually clear number can still refer to chargeable weight rather than gross weight, or to a master movement rather than a house consignment.

When reviewers disagree, preserve both observations and route the case to the designated operational authority. Track the time spent resolving the disagreement. Excluding that work would overstate the benefit of extraction.

Measure entry, review, amendment handling and export reconciliation separately. Compare total staff effort per completed case with the baseline, retaining unresolved cases and their age. Page-processing speed alone is not the business outcome.

Use identity and version controls before exporting

Define the business association using the complete identifiers and partner context required by the destination. Do not rely on a bare number that may be ambiguous across document types or source systems.

Use a file fingerprint to identify exact repeat uploads, while recognising that a new scan of the same document may have a different fingerprint. Resolve those cases through document identity, version evidence and reviewer confirmation. A changed image is not necessarily an amendment, and a repeated reference does not prove identical content.

A supervised export should carry the approved case, field set, source version and a stable operation reference. Record the destination's business acknowledgement. If a response is uncertain, reconcile the destination before retrying or creating another job.

The acceptance test should include a repeated upload, an amended weight, an unreadable identity, a wrong house/master association and an acknowledgement lost after the destination update. Require the expected record count and values after recovery, rather than only a successful API response.

Define stop conditions and a recoverable fallback

Stop export if a critical value lacks review, a case is associated with the wrong shipment, the same approved operation produces a duplicate record or document access crosses the authorised client boundary. Keep intake evidence and pending decisions available for investigation.

The operations lead returns new cases to the established manual entry process. The integration owner reconciles every operation that may already have reached the destination. The document-review lead checks whether a mapping error affects other records in the sample.

An exported amendment cannot be undone by deleting the OCR result. Follow the destination's authorised correction process and preserve the prior and corrected values. Resume supervised exports only after the operational sponsor accepts the reconciliation and the failing case has been demonstrated successfully.

Review the pilot as a controlled data-entry change

The decision to continue should use field-level error results, duplicate and amendment outcomes, total review effort and the reliability of downstream reconciliation. Record which layouts and operating conditions were actually tested.

A pilot that reduces typing but creates an unresolved review backlog needs a narrower scope or a different workflow. A successful sample supports the tested boundary; it does not prove that every partner layout or document family is ready.

Next actions

1. Select the branch, layouts, destination and authorised reviewers.
2. Freeze the baseline sample and approve the critical-field and relationship map.
3. Run extraction in staging, then inspect duplicates, amendments and field-specific errors.
4. Consider supervised export only after the operational team accepts the recovery tests and reconciliation.

To scope an AWB extraction pilot, describe the document families, approximate review volume, destination system and amendment problems. An anonymised process outline is enough to begin defining the sample and review boundary.