Solving document fragmentation to power freight AI

The mission: Unlock spend visibility with better data

A global audit and payment-solutions provider in the logistics industry set out to transform how enterprises understand and control transportation spend. With proprietary AI models as their strategy, they needed to train these systems on vast amounts of real-world logistics data, particularly freight invoices and shipping documents. To succeed, they sought a labeling partner who could handle the high variability, unstructured formats, and handwritten inputs found across global carrier networks.

The challenge: Navigate complex data for precise labeling

The client encountered major roadblocks in scaling their labeling operations.

Key obstacles

• Highly fragmented documents across carriers, formats, and systems
• Handwritten fields that made OCR unreliable and required human interpretation
• No standardized templates, increasing the complexity for labelers
• Scalability limitations that slowed down model training and iteration cycles

The goal

Build a high-volume, high-accuracy labeling pipeline for diverse logistics documents that could fuel proprietary models with clean, reliable data consistently.

The solution: Deploy domain experts to tackle format diversity

We deployed a dedicated team of 80 trained analysts that were equipped to handle nuanced logistics documents across a diverse set of tasks. From understanding diverse freight information to annotation guidelines, the operation was tailored to keep up with labeling velocity and quality.

Our approach

1. Assembled a skilled team of 80 logistics-savvy annotators
2. Labeled 80+ complex task types, including:
   1. invoice field extraction (invoice number, PO number, addresses, payment terms)
   2. handwritten data classification (delivery notes, signatures, manual corrections)
   3. document type classification (invoice, proof of delivery, bill)
   4. anomaly detection (corrupted and mismatched documents)
3. Validated quality of tasks with human-in-the-loop reviews and cross-checking label consistencies
4. Built an iterative feedback loop with the client to improve task definitions and maintain alignment
5. Accelerated learning cycles, enabling multiple tasks to reach data saturation where additional labeling was no longer needed

The results: AI model saturation achieved

• 80 expert analysts onboarded
• Over 20M tasks labeled within 9 months
• More than 80 task types covered, with multiple achieving saturation
• Model accuracy improved to above 90%

The client’s AI models delivered consistent insights that enabled more transparent and efficient transportation cost management for the industry.