Back to Blog

OCR Accuracy in Logistics and Human Review

A practical way to evaluate field accuracy, document-level corrections and reviewer effort in BOL and invoice extraction.

Esnaj Team

An OCR accuracy figure is only useful if you know what was counted. Field accuracy measures extracted values; document pass rate measures whether every required value on a document was acceptable without correction. Those are different questions. In Esnaj’s DHL Aruba BOL and invoice case, the reported accuracy exceeds 96% across extracted fields. That does not say 96% of documents need no review. This is an operational figure supplied by Esnaj, not an audited dataset.

Measure OCR accuracy before human review

MeasurePractical definitionWhy it matters
Field accuracyCorrect extracted fields divided by evaluated fields.Shows where extraction works or fails by field type.
Document pass rateDocuments with all required fields accepted without correction divided by evaluated documents.Shows whether a whole document can move on without an edit.
Review timeMinutes spent checking and resolving fields per document.Captures the human work left after extraction.
Exception rateDocuments needing an explicit decision about missing or ambiguous data.Helps plan reviewer capacity.

Define “correct” before sampling: exact string match may be wrong for a date that has an equivalent normalized format, but an incorrect invoice number must not be accepted because it looks similar. Separate OCR reading errors from field interpretation errors. Keep a reviewer-approved reference value for each tested field and record what changed during review.

A hypothetical illustration: if each document has several critical fields, even high field accuracy can leave some documents with at least one wrong value. The actual document pass rate depends on field count, error distribution and which fields are mandatory; do not derive it from 96% alone. A confidence score is a cue for review, not a substitute for testing accuracy against reviewed values.

Design review for the hard cases

Gather examples from each supplier layout, including skewed scans and alternative labels. In the DHL workflow, image preparation addresses orientation; OCR and AI interpret the text and varied layouts; staff verify extracted data and correct or decide on problematic fields. Reviewed data normally remains in the replacement platform and can be exported. The platform replaced an older system; this example does not establish a customs or transport-management integration.

For an evaluation, stratify a held-out sample by supplier, document type and scan quality. Record field-level correctness before human edits, then separately measure time from upload through final human review. Check a sample of apparently clean documents too: otherwise you will miss plausible-looking errors that were never flagged. Track false alerts as well as missed problems, because both consume capacity or create risk.

Agree escalation rules for illegible originals, absent fields and conflicting values. Name who can approve a correction, and retain a way to trace reviewed values back to the source image. Before expanding to new suppliers, rerun the sample: an overall figure can hide poor performance on a small but important document group.

Read the BOL and invoice workflow walkthrough for the concrete sequence, or explore document extraction services. To assess your own exception queue, contact Esnaj.

Frequently Asked Questions

Q: Does 96% field accuracy mean 96% of documents need no correction? A: No. The DHL figure is for extracted fields. A document with one incorrect required field still needs a correction, regardless of how many other fields were right.

Q: Should reviewers only inspect flagged fields? A: Flags help prioritize work, but sample unflagged results too. A plausible incorrect value may escape a flag; define the checking policy according to the consequences of each field.

Q: What should a pilot measure besides accuracy? A: Measure document-level pass rate, correction and exception counts, and full processing time including human review on a representative sample.