01Problem
Supplier documents arrive in every format imaginable. Rule-based extraction plateaued: each new supplier template meant more brittle rules, and exceptions still landed on humans.
02Context
Accuracy targets in finance are unforgiving — a wrong amount is worse than no answer. Any AI step needed confidence signals and a human fallback, not silent best guesses.
03Solution
LLM extraction with structured output schemas, validated against business rules (totals must reconcile, dates must parse, suppliers must match records). High-confidence results flow straight through; anything below threshold routes to a human with the model's reasoning attached.
04Tradeoffs
Per-document inference costs money and latency; batching and caching help, but the honest tradeoff is paying pennies to avoid minutes of human keying. The threshold dial is a business decision, not a technical one — we made it configurable.
05Outcome
Straight-through processing rose sharply while error rates fell, because the model abstains where rules used to guess.
06Lessons Learned
LLMs are best where rules are worst: messy inputs, bounded outputs.
Design the abstention path first; the happy path takes care of itself.