A useful implementation detail to add after extraction: Textract's blocks describe geometry and relationships, but they do not automatically become a stable editable document. For multi-column pages, sort LINE blocks within detected layout regions instead of globally by Y coordinate, or copied text can interleave columns. For tables, follow CHILD relationships into CELL blocks and preserve row/column spans before exporting; flattening everything into strings loses the structure users expect in Word or Excel. I also recommend storing confidence beside every critical field and routing low-confidence names, dates, totals, and handwriting for review rather than silently accepting them. A successful API response proves extraction completed, not that the result is correct.