DHL Aruba: BOL and Invoice Automation
How Esnaj replaced DHL Aruba's older BOL and invoice system with image preparation, OCR, AI and human review; what the reported results mean.
BOL and invoice automation at DHL Aruba combines image preparation, OCR, AI and human review in a custom platform built by Esnaj. A Bill of Lading (BOL) records shipment information; the platform extracts its fields alongside those on invoices. It replaced an older internal system where staff entered fields manually. The reported processing time is about four minutes per document including human review, versus approximately 15–20 minutes before. These are approximate operational figures supplied by Esnaj, not audited benchmarks or promises for another operation.
What changed in the workflow?
Before the replacement, a person read each document and typed its fields into the older system. A fixed template could not reliably handle varying supplier layouts, alternative names for equivalent fields or scans with the wrong orientation. The new platform handles document preparation and extraction while leaving verification and decisions about problematic fields to people.
| Stage | What happens | Where people remain involved |
|---|---|---|
| Upload | Staff submit BOLs and invoices to the replacement platform. | Select and submit the appropriate documents. |
| Prepare | Image preparation corrects angle or orientation before extraction. | Check inputs when a scan is unclear. |
| Extract | OCR reads text; AI interprets fields across different layouts and labels. | Verify the resulting values. |
| Resolve | Problematic fields are flagged. | Correct the field or decide how to handle it. |
| Store | Reviewed data stays in the platform, with optional export. | Choose whether export is needed. |
The platform replaced the older internal system: staff upload documents, review results and keep the approved data in one place. Any connection to a separate system would need its own scope. For the full implementation account, read the DHL Aruba BOL and invoice case study.
Read the reported figures carefully
Esnaj reports around 1,000 documents processed per day and accuracy above 96% across extracted fields. Field-level accuracy does not mean that more than 96% of entire documents pass without correction: a single document may contain many fields, and one problematic value still needs attention. The four-minute processing figure already includes the time people spend reviewing. From the approximate before and after times, the estimated reduction is 11–16 minutes per document, or roughly 73–80%. It is not an automation-only runtime.
A new team should not multiply those figures into guaranteed savings. Document mix, scan quality, number of fields and review rules change the work. A pilot should count its own documents, time the entire process from upload through review and record which fields staff correct most often. Compare like-for-like documents, including exceptions, rather than choosing only clean examples.
When to use this pattern
Start with a narrow document family and a named owner for review. Collect representative supplier formats and poor scans, identify essential fields, then agree how missing or ambiguous values should be handled. Confirm where approved data lives and whether an export is actually needed. If a separate system must receive it, scope that connection explicitly rather than assuming it is part of OCR.
For a closer look at measuring field-level results and review work, see OCR accuracy and human review in logistics. Learn about logistics document extraction or discuss your own document workflow with Esnaj.
Frequently Asked Questions
Q: Does the DHL platform still need human review? A: Yes. Staff verify extracted values and resolve flagged fields. The approximately four-minute processing time includes that review.
Q: Where does the reviewed data go? A: Reviewed data normally stays in the platform. Staff can export it when needed; a connection to another system would need to be scoped separately.
Q: What does 96%+ accuracy describe? A: It describes accuracy across extracted fields, as reported by Esnaj. It does not measure the proportion of documents requiring no correction.