AI drafts, a person confirms — never auto-post extracted data

Context

LLM-based document extraction is fast — a purchase invoice comes back parsed in seconds — but it isn't perfect. Even a well-tuned pipeline leaves a meaningful share of documents needing a correction before the numbers are right.

Decision

Every extracted document sits in a review state. Nothing downstream — a posting, a saved record, anything else — happens until a person looks at it and approves it. The model drafts; it never commits on its own.

Why

A wrong number in a financial posting is expensive to unwind after the fact — tracing it, reversing it, re-entering it correctly. A short review before it goes anywhere is cheaper than a correction after it's already downstream. Even at a 94% zero-correction rate on the best-performing document type, a meaningful share still needs a human to catch it.

Trade-off

Every document gets a manual step and some latency it wouldn't have with straight auto-posting. That's the deliberate cost — the alternative is trusting an unreviewed model with money.

The same decision has been made independently a second time, in a separate engagement, for a completely different kind of document — not an accounting record at all. Same reasoning, same conclusion: AI drafts, a person confirms.

purchase invoice zero-correction rate

94%

bank statement zero-correction rate (hardest document type)

89%