Extraction demos are easy. You feed a model an invoice or an RFQ, it returns structured fields, everyone nods. Then you try to put it into a real quoting or accounting workflow and the question changes from “can it extract?” to “can we trust this without a human re-checking every field?”
That gap is validation, and it’s most of the work.
Validation is the product
On both an invoice pipeline and a client RFQ parser, the extraction was maybe 30% of the effort. The rest was making the output safe to act on:
- Schema checks — the output has to match the shape the downstream process expects.
- Business constraints — totals reconcile, required fields exist, values fall in plausible ranges. On the invoice pipeline that was 9+ distinct criteria.
- Fail loud, not silent — a bad extraction gets flagged for a human, never quietly passed through as if it were fine.
Keep the human where it counts
The goal was never to remove the estimator or the accountant. It was to remove the re-typing. The system drafts; a person approves. That single design choice is what makes a business willing to put AI output in front of a customer.
The takeaway
If you want AI in a real workflow, budget for validation like it’s the main feature — because to the person relying on it, it is.