Summary
Vasant Kearney from Onlay presents at AI Engineer on using X12—the strict EDI data standard governing US healthcare insurance claims—as a structural harness for AI agents navigating the full patient journey from eligibility verification to provider payment. The talk frames the problem in terms of compounding errors: a typical claims workflow involves 50 or more discrete steps, and without hard structural constraints, mistakes in early steps propagate through the entire pipeline.
The central insight is that X12 functions similarly to TypeScript for LLMs—a bounded, strict format that constrains the output space, enables hard rejection of malformed outputs, and gives agents something reliable to ground their reasoning in. Kearney draws an analogy to early computer vision (recognizing handwritten digits was solved long before depositing a check was tractable) to illustrate how exciting AI capabilities and production-safe AI systems are different engineering problems requiring different harnesses.
The talk also addresses multimodal challenges in claims—medical imaging, EHR data, and desktop UI navigation—and the core architectural tradeoff between fully agentic reasoning (flexible but expensive, slow, and error-prone) and fully hardcoded workflows (reliable but brittle and unmaintainable at scale). Onlay’s approach introduces partner-level and user-level memory to identify repeated patterns in multi-site health organizations and selectively hardcode high-frequency paths, striking a balance between adaptability and the strict reliability that insurance processing requires.
📺 Source: AI Engineer · Published August 19, 2026
🏷️ Format: Keynote Launch
