Extraction is only the first step
A document pipeline needs to preserve meaning as well as text. Keep the source page and evidence span for each extracted field. Validate the output against a schema before it reaches a downstream system. A missing purchase order is an exception, not an invitation to invent one.
Choose the simplest tool that fits
Use OCR and rules for stable document layouts. Use a language or vision model where wording and layout vary enough to make rules brittle. Keep arithmetic, duplicate checks, and master-data matching deterministic. The model can explain a mismatch; it should not decide that the mismatch no longer matters.
Design the review queue
Send reviewers the original document, extracted values, failed checks, and proposed correction together. Measure corrections by field and document type. A pipeline that processes easy invoices quickly but hides difficult ones has not solved the operational problem.