United StatesRetail and e-commerce2025 – ongoingPilot Company
Generative AI processing of handwritten fuel bills of lading
Pilot Company cut manual fuel-bill processing by 99% with a multimodal AI pipeline
Problem
Pilot Company, a North American travel center operator with 850 locations, had to process handwritten fuel bills of lading for billing. The forms came in more than 300 layouts, and earlier OCR and RPA attempts failed.
Approach
Pilot built a generative AI pipeline on Databricks. A multimodal model reads photos of the documents with few-shot prompting, and fuzzy matching against reference tables corrects product names and terminal addresses. Accuracy rose in steps from plain OCR at 76.2% to 98.6%. The team tracked prompts and experiments with MLflow.
Results
Takeaway
Plain OCR scored 76.2% on these forms, below the 98.6% reached once a multimodal model and fuzzy matching were combined. Pilot's human annotators scored 71% on the same test.
Categories
- Size
- Large enterprise
- Industry
- Retail and e-commerce
- Business function
- Production and quality
- Innovation type
- Process innovation
- AI technology
- Generative AI
- Adoption
- Off-the-shelf tool or platform
- Evidence
- Vendor case study
Sources
Source summaries
Databricks describes how Pilot Company automated fuel billing by reading handwritten bills of lading with a multimodal model and fuzzy matching. It reports 98.6% accuracy, a 99% cut in manual processing and a 90% drop in processing costs, and it puts the value in the low seven figures. Agentic data enrichment and internal chatbots are listed as planned, with no figures.
Full details are in the original sources.