Accuracy Benchmarks
Traditional OCR engines like Tesseract achieve 70-80 percent character accuracy on clean typed documents. Accuracy drops significantly with low-quality scans, unusual fonts, and handwritten content. AI vision models achieve 95-99 percent field-level accuracy because they understand document context and structure, not just individual characters.
2-Pass Verification
Cargoffer OCR implements a unique 2-pass verification system. After the first extraction pass, the model re-examines each field against the original document. Fields below the confidence threshold are flagged. This process catches extraction errors that single-pass systems miss. In production testing, invoice verification achieves 0 field-level errors.
Real-World Results
A 41-page Repsol invoice with 65 fuel cards, 15 product types, and 263,000 EUR total was extracted with zero errors across all line items and totals. A 95-station fuel price list was extracted with all stations and 5 product types correct. These results come from our standard extraction pipeline without manual correction.
Factors Affecting Accuracy
Document quality is the primary factor: 300 DPI scans achieve highest accuracy, while camera photos of documents may show lower results. Table complexity matters simple row-column layouts extract best. Handwriting accuracy varies by legibility but modern AI vision handles most handwritten CMR and delivery note fields.