Our e-commerce rule engine holds risky invoice orders for human review, and the reviewers were drowning in false alarms. So we added an LLM second opinion — advisory only: a verdict, a confidence score, and a written rationale next to the hold reasons. The human still decides, and the model can never touch the ship path.

The surprise wasn't the triage speedup. It was that when reviewers disagreed with the model, about half the investigations ended in our code: a years-old off-by-one that flagged invoices as overdue on their own due date, "outstanding credit" that was actually unshipped prepayment orders, a payload field that had silently been zero forever, and a loyalty-credit loophole that had been quietly farmable for years.

The full post covers the architecture (de-identified payloads, prompts versioned in the database, eval fixtures that deliberately freeze time) and why a model that must explain itself turns out to be a continuous audit of your feature pipeline.

Originally published at andreasbergstrom.dev — read the full post there.