Scans, forms and OCR at scale

OCR turns pixels into characters. Document intelligence turns a page into structure: reading order, headings, tables, key-value pairs, checkboxes, signatures and typed fields, each with a location and a confidence score. Enterprise documents (Halden Freight's bills of lading and customs declarations) are scanned, photographed, multilingual and partly handwritten, so you need both, plus the discipline around them: pre-processing, an engine choice per document class, deterministic validators, confidence thresholds calibrated on a labelled set, a human review queue for what fails, and throughput and cost controls for volume that arrives in bursts. The output metric that matters is not "OCR accuracy". It is the error rate of fields that were accepted without a human looking.

Key takeaways:

• Measure at three levels: CER/WER for text, field-level accuracy for extraction, and error rate among auto-accepted fields for the business.
• Confidence alone is not enough. Confident-and-wrong values exist; deterministic validators (check digits, formats, cross-field sums) catch many of them.
• Language and script support differ sharply between services. Verify per language before you promise anything.
• Vision-language models read messy documents well but can produce plausible values that are not on the page. Treat their output as untrusted until validated and grounded.
• At 95% per-field accuracy, a 20-field document is fully correct only about 36% of the time. Plan review capacity from that number.