.png)
A computer vision system replaced a manual processing team and handled 94% of freight documentation automatically without human review.
A mid-sized US freight broker managing approximately 4,200 shipment documents per day — bills of lading, proof of delivery receipts, rate confirmations, carrier invoices, and customs declarations — was processing every document through an 11-person manual data entry team. Documents arrived by email, fax, carrier portal upload, and EDI feed in 23 distinct formats from 340 carrier partners. The data entry team classified each document, extracted the relevant fields, validated the extracted data against the shipment record in the TMS, and flagged exceptions for operations staff to resolve.
Average document turnaround was 3.8 hours from receipt to TMS entry — a delay that was causing downstream problems in billing, carrier payment, and customer shipment tracking. The data entry team was operating at capacity, errors in manual extraction were running at 4.2% of documents, and the business had no path to handling volume growth without hiring proportionally into a role that was difficult to recruit for and had annual turnover of 38%.
Verttx built a computer vision document processing system covering all 23 document formats and all 340 carrier partners. The system went live 11 weeks after the initial discovery call. Within 30 days, 94% of documents were processed straight through without human involvement.
The document processing problem had three layers that each made the others worse. The first was format variability. Freight documents are notoriously unstandardised — a bill of lading from one carrier looks nothing like one from another, and the same carrier might use different templates for different lanes or customer relationships. The 340 carrier partners in this broker's network were generating documents in 23 distinct format families, with significant variation within each family. Template-based OCR approaches — the most common off-the-shelf solution — had been evaluated and rejected because they require a separate template for each document variant, and maintaining 23-plus template sets against a carrier network that regularly changes document formats was operationally untenable.
The second layer was image quality. Documents arriving by fax — still a significant channel in freight, representing 31% of inbound document volume — were frequently low-resolution, skewed, or partially degraded. Documents photographed by drivers on delivery — proof of delivery receipts — varied enormously in lighting, angle, and background. The manual team had developed the ability to read and interpret poor-quality images through experience. Any automated system would need to match that tolerance for real-world image conditions, not just perform well on clean scans.
The third layer was exception handling. Approximately 18% of documents contained discrepancies that required human judgement — weight variances between the bill of lading and the carrier invoice, delivery signatures that did not match the consignee record, or rate discrepancies between the confirmation and the invoice. These were not processing errors. They were legitimate business exceptions that required an experienced operations person to investigate and resolve. Any automation system that could not cleanly separate straight-through documents from genuine exceptions would either block the business on false exceptions or pass real exceptions through undetected.
The system's first stage is a document classification model that identifies document type, carrier, and format variant from the raw input — regardless of quality, orientation, or channel of origin. The classifier was trained on 340,000 labelled documents from 26 months of the broker's historical document archive, covering all 23 format families and all 340 carrier partners. Image preprocessing handles the real-world quality issues systematically: automatic deskewing corrects rotations up to 35 degrees, adaptive binarisation improves low-contrast fax images, and a super-resolution upscaling model trained specifically on freight document characteristics improves legibility of degraded text before extraction begins. Classification accuracy on the production document stream reaches 99.1% — sufficient to route each document to the correct extraction model without manual intervention.
Each document type routes to a dedicated extraction model trained on that format family. Rather than a single general-purpose extraction model — which performs adequately on common fields but poorly on the format-specific layouts and terminology that vary across carrier types — Verttx trained seven specialised extraction models covering the primary document categories: bills of lading, proof of delivery receipts, carrier invoices, rate confirmations, customs declarations, lumper receipts, and accessorial charge documents. Each model was trained to extract the fields relevant to that document type and to recognise the format-specific conventions that distinguish, for example, a temperature-controlled carrier's bill of lading from a flatbed carrier's.
Extracted fields are validated in real time against three reference sources: the shipment record in the TMS, the carrier's rate confirmation on file, and a business rules engine encoding 84 validation rules covering weight tolerances, date consistency, address matching, and rate variance thresholds. Documents where all extracted fields pass validation are classified as straight-through and written directly to the TMS. Documents where any field fails validation — or where the extraction model's confidence score on any critical field falls below a calibrated threshold — are routed to a human review queue with the specific validation failure highlighted and the relevant TMS record pre-loaded. Reviewers see exactly what the system could not resolve, not a cold document to re-process from scratch.
Every document reviewed by a human — whether a genuine exception or a low-confidence extraction — feeds back into the training pipeline. Reviewer corrections are captured, labelled, and incorporated into monthly model retraining cycles. This data flywheel means the system improves continuously as it encounters new carrier format variants and real-world image quality conditions. In the six months following go-live, the straight-through processing rate improved from 91% at launch to 94% through accumulated learning alone, without any manual model intervention by the Verttx team.
Within 30 days of full deployment, 94% of the 4,200 daily documents were processed straight through without human involvement — classified, extracted, validated, and written to the TMS in an average of 4 minutes from receipt. The remaining 6% — 252 documents per day — were genuine business exceptions routed to the review queue with full context. The 11-person data entry team was restructured: 3 people now manage the exception queue and carrier relationship issues that require operational judgement. The remaining 8 positions were eliminated through natural attrition over six months, reducing the processing function's annual labour cost by $640,000.
Document extraction accuracy improved from a manual baseline of 95.8% — the complement of the 4.2% manual error rate — to 99.3% for straight-through documents, driven by the consistency of automated extraction against validated reference data versus the variability of manual data entry across shift patterns and experience levels. Downstream billing accuracy improved measurably: disputed invoices attributable to document processing errors fell from 3.1% of monthly invoice volume to 0.4%, reducing the accounts receivable team's dispute resolution workload by an estimated 14 hours per week.
Document turnaround time fell from 3.8 hours to 4 minutes for straight-through documents. The operations team, previously managing a perpetual backlog of unprocessed documents, now operates against a near-real-time document state. Shipment tracking accuracy for customers improved because proof of delivery receipts are now in the TMS within minutes of driver upload rather than hours later when the data entry team reached them in the queue. The cost per document processed fell from $1.84 to $0.35 — a reduction of 81%.
All seven extraction models, the classification system, the validation rules engine, the TMS integration, and the continuous learning pipeline were transferred to the broker's engineering team at handover — the system runs entirely within the client's own infrastructure with no ongoing involvement from Verttx required.
We had looked at template-based OCR twice before and rejected it both times because maintaining templates against 340 carriers was a maintenance burden we could not sustain. Verttx built something that learns the formats rather than being told them. The straight-through rate at month six is better than the accuracy rate of our manual team. That is not something I expected to be able to say. — VP of Operations, US Freight Broker
.png)
94% of 4,200 daily documents processed straight through within 30 days of deployment. Document turnaround fell from 3.8 hours to 4 minutes. Extraction accuracy improved from 95.8% manual baseline to 99.3%. Annual labour cost reduced by $640,000 through natural attrition across 8 positions. Cost per document fell from $1.84 to $0.35 — an 81% reduction. Disputed invoices from processing errors fell from 3.1% to 0.4% of monthly invoice volume. Straight-through rate improved from 91% to 94% through the continuous learning pipeline in the first six months alone.
.png)
.png)
