The challenge
A mid-sized logistics firm operating out of JNPT and ICD Tughlakabad was processing 320 Bills of Lading every working day across road, rail and ocean consignments. Four data-entry executives did nothing else — keying consignor, consignee, container numbers, HSN codes, freight terms and weights into SAP. Errors propagated downstream into GST invoicing and customs declarations. A typo on a BoL meant a Rs 40,000 dispute three weeks later.
How we deployed
- Built a multi-modal extraction agent that reads scanned BoLs, photographed BoLs and EDI feeds in one pipeline.
- Schema-validated every extracted field against a JSON contract — container number checksum, HSN format, INCOTERMS enum.
- Pushed clean records straight into SAP via the firm idoc bridge, with the source PDF attached as document object.
- Routed any field below 92% confidence to a Next.js exception dashboard for one-click human review, never silent failure.
- Logged every accepted and corrected extraction back into a feedback loop to fine-tune the prompt monthly.
What changed
- Per-BoL processing time fell from 9 minutes manual to 14 seconds end-to-end.
- Extraction accuracy hit 99.6% on a daily blind audit of 50 random BoLs.
- 52 hours of weekly data-entry labour recovered across the 4-person team.
- Downstream invoice disputes dropped 87% as upstream typos disappeared.
- Two of the four executives were redeployed to billing-exception management, which the firm had previously outsourced.
"My team used to retype BoLs eight hours a day. Now they actually catch the disputes before they become disputes. Same headcount, completely different job."
— Operations Director · Logistics Firm