Teaching AI to Read Payer Edits
Extracting a UnitedHealthcare Smart Edit into structured JSON, and why X12 loop hierarchy is the hard part.
First published on LinkedIn.
After years in the trenches, I finally updated my About section last week. If you’ve ever lived in the 837/835/270/276 world at scale, take a look and let’s connect.
One thing I mentioned there: in my spare time the past few months I’ve been experimenting with LangChain to see if it can meaningfully reduce the manual pain of keeping up with payer-specific edits.
Latest test this weekend — pointed it at the current 2025 UnitedHealthcare 837 Smart Edits guide. This particular edit (CMJW2) is one example of the challenge: it references NDC Invalid Drug Discard Modifiers, mentions JW and JZ modifiers, talks about single-dose drugs, and points to an external policy document.
Image - UHC Smart Edits Guide
After several refinement passes, I extracted it into structured JSON that captures the loop context (2400), segment paths (SV1 elements), the core validation logic, and even the policy reference. Not perfect, but a solid starting point.
It's still very experimental and completely separate from any day-job work. But the potential time savings on manual rule creation feels real.
The hard part isn’t the text extraction — it’s teaching the model to respect X12 loop hierarchy (2000A → 2000B → 2300 → 2400) and map natural-language requirements to the right segments and elements.
More details coming in future posts. For now — anyone else experimenting with LLMs to help tame companion guides? Curious what’s working (and what definitely isn’t).