Logistics · AI Document Intelligence
Sample writeupThree-day BOL backlogs cleared the same day.
A mid-market US freight broker was three days behind on BOL and invoice processing, every day. Errors cost real money downstream. We embedded with ops, deployed a document pipeline against their real worst-case docs, and turned a 3-day backlog into same-day processing.
- Client
- Anonymized · US Freight Broker
- Industry
- Freight & Logistics
- Region
- United States
- Duration
- 10 weeks
The challenge
12 people processing BOLs, invoices, and rate confs by hand across emails, PDFs, EDI, and photos from drivers. Error rate around 3%, costly when the average load is $1,800 and a single mistyped invoice number delayed payment by 2 weeks.
The approach
Identify. Build. Adopt.
Identify
Week one: 'send us your 50 worst documents.' We benchmarked the pipeline against their actual edge cases (handwritten amendments, blurry trailer-cab photos, multi-page rate confs) and showed the math: 1,400 hours/year of pure data entry, with measurable error cost.
Build
Multimodal extraction (Claude/Gemini vision) + specialized OCR for the harder docs, structured to a Pydantic schema, validated against the TMS, queued for human review when confidence dropped below threshold. Integrated straight into their existing ops dashboard so the team never left their workflow.
Adopt
Two-week parallel run. Every doc went through both the pipeline and a human; we logged every disagreement and tuned. By week 8, the team trusted the queue enough to flip the default from 'review every doc' to 'review only the flagged ones.'
Outcomes
What changed, measurably.
We stopped chasing missing data and started running the brokerage again. The pipeline catches what we used to catch on a good day.
Got a similar leak?
30 minutes with the founders. Bring us the workflow that hurts most. We'll cost the leak live and tell you whether the same playbook deploys here.
More work