OCR

Pre-processing, deskew, denoise, contrast, determines OCR accuracy more than the recognition engine does.

Modern vision-language models substantially outperform classical OCR on poor scans, handwriting and complex layouts, at higher cost per page.

Output structure matters as much as character accuracy. A perfectly transcribed financial table flattened into a paragraph has lost the information that made it useful, so layout preservation belongs in the requirements alongside the accuracy target.

Set the confidence threshold against the cost of a mistake. On a financial document, routing an uncertain field to a human is cheap; on a bulk archive it may not be worth it, and that is a business decision rather than a technical default.