A freight forwarder was drowning in 800+ shipping documents a day across English, Georgian, Russian, and Turkish. Seven months later: 95% automated, 0.3% error rate, ops team redeployed to exception handling and customer success.
less manual processing time
Across BoLs, customs forms, multilingual invoices
Client
Mid-sized freight forwarding & customs brokerage
Timeline
7 months to production + ongoing monthly retainer
Team
4 engineers (2 ML, 2 full-stack), 1 designer
Engagement
Discovery → POC → phased production rollout → monthly improvement retainer
A mid-sized freight forwarding and customs brokerage had a 12-person ops team whose entire day was data entry — pulling fields off bills of lading, customs declarations, commercial invoices, and certificates of origin, then keying them into the routing system. Volume had doubled in two years; the team was at breaking point. Errors were costing real money — incorrect HS codes alone were generating customs holds, fines, and merchant chargebacks.
We treated this as a workflow-automation problem with AI inside it, not as an 'AI project.' The hardest decisions were about where humans stay in the loop, not which model to use.
01
Spent two weeks with the ops team — sitting next to them, watching what they actually did, not what the org chart said they did. Found seven document types and 23 distinct field-extraction patterns. About 40% of their time was spent on judgement calls (HS code disambiguation, route exceptions); 60% was pure transcription.
02
Vision OCR (AWS Textract) for raw document scanning, LLM extraction (GPT-4o + Claude as fallback) for structured field parsing, RAG over the client's historical filings for context-aware classification (especially HS code disambiguation), and a review queue for anything the pipeline scored below confidence threshold.
03
Built a Next.js review console where the ops team could approve, correct, or reject the pipeline's output in seconds. Every correction fed back into a quality dataset that improved retrieval and prompt context — the system got measurably better month over month from real-world feedback, not synthetic training data.
04
High-confidence extractions ($>$0.95 across all fields) routed automatically to the TMS. Mid-confidence (0.85–0.95) flagged for one-click ops review. Low-confidence routed to full manual review. The thresholds are tunable per document type and have been adjusted twice since launch.
05
BoLs first (highest volume, most standardized). Customs declarations second. Multilingual invoices last. Each phase ran in parallel with manual processing for two weeks before cutover, with daily accuracy comparisons.
06
We worked with leadership on the team transition early — three people moved into customer success, four into exception handling, three into a new compliance review function, two left for unrelated reasons. No layoffs were necessary, which was an explicit goal of the engagement.
reduction in total manual processing time
data-entry error rate
down from 4%
of documents processed without human touch
peak-season backlog
down from 2 days
fewer customs hold incidents post-launch
The financial picture flipped within three months. The ~$40K/month in customs holds and chargebacks dropped to under $8K. The freed ops capacity moved into customer success — a function the company genuinely needed but had never been able to staff. Customer satisfaction scores moved up measurably and they won two RFPs in the following year that explicitly cited their document-handling speed as a deciding factor.
The pipeline continues to improve. Every ops correction feeds the retrieval layer and the prompt context library. We run a quarterly review with the operations director to retune confidence thresholds based on actual performance, and add new document types (we are currently piloting Letters of Indemnity).
The most important outcome was cultural. The ops team — who could have been hostile to a project that automated most of their old job — became its biggest advocates, because nobody lost their job, the work that remained was more interesting, and the system listens to them.
01
Discovery & process mapping
Time on the ops floor, document-type taxonomy, field-extraction backlog, ops-team transition planning with leadership.
02
POC on bills of lading
End-to-end pipeline on the highest-volume document type. Demonstrated >95% accuracy on a held-out test set of 500 docs.
03
Production hardening
Review console, confidence routing, monitoring, security review, TMS integration.
04
Phased rollout
BoLs → customs declarations → multilingual invoices, two-week parallel-running per document type.
05
Ops redeployment & handover
Team transitions, runbook authoring, on-call training, performance baselining.
06
LLM features, RAG systems, document AI, and workflow automation — integrated where they pay for themselves, not bolted onto everything.
Learn more02
Bespoke business systems built around the workflow you actually have, not the one a generic SaaS forces on you.
Learn more05
Infrastructure that doesn't keep you up at night. AWS, GCP, Azure. Kubernetes when it earns its place. CI/CD with rollback. Cost-aware from day one.
Learn moreStrangler-fig migration from a PHP monolith to a Node.js microservices platform on Kubernetes. p95 latency dropped from 1.8s to 320ms, PCI-DSS audit passed, transaction headroom 4×.
Stagnant 18-month organic plateau, broken Core Web Vitals, JS-heavy SPA blocking crawlers. Migration to Next.js with comprehensive schema, performance overhaul, and indexation cleanup.
Most engagements start with a 30-minute discovery call. No pitch deck, no NDAs on day one — just an honest conversation about your situation.
Schedule a Call