The temptation when building an AI-native billing product is to lead with the model and let the rules backfill. We tried it. It does not work.
Deterministic rules — CPT/ICD crosswalks, NCCI edits, modifier compatibility, duplicate detection — are still the only way to guarantee a specific class of denial never ships. They are cheap, fast, and explainable to your auditor.
What an LLM is genuinely better at: scoring documentation risk, summarizing payer policy drift, and drafting first-pass appeal letters with the right tone. That is a real, important contribution — but it is additive, not foundational.
Our architecture: rules block first, the model adds risk + reasoning on top, and every decision logs the model version + input hash. You get the speed and quality of AI, with the safety net auditors require.