Healthcare AI postmortem: model integration took days, data work took months

alex_verem · x · 2026-10-05

An engineer shares lessons from a four-month build deploying production AI agents inside a PE portfolio company's medical billing system under HIPAA:

The claims engine's rules were hand-written by two people: 300+ denial codes and payer logic that shifts every quarter
Month one involved no modeling—instead mapping where data lives and what's missing, so agents reason from structured facts rather than guessing
Key artifact: an enrichment layer pre-computing and versioning 34 variables per claim before any model sees it
Core takeaway: teams that bolt a model onto messy inputs and conclude "AI isn't ready" have it backwards—"AI was ready. Their data wasn't."

More from coding & agent