AI医疗索赔代理

Autonomous agent demos get millions of views running on benchmarks in sandboxes. This team shipped 7 production agents handling live healthcare claims under HIPAA, built in 4 months with zero patient data exposed to any model. They route calls across two model providers and swapping one for the other changes nothing in the output. The team spent more engineering hours building enrichment pipelines and eval harnesses than integrating the model. Follow @mardehaym for breakdowns of AI in production. > **引用原帖 Mark Ajzenstadt (@mardehaym):** > A PE-backed healthcare billing platform had two people hand-writing every rule in their claims system. 300+ denial codes, payer variations that change three times a year. > We built seven production AI agents that handle it. Four months. Zero PHI exposure. > I want to walk through what that actually took, because "we added AI to the product" tells you nothing. > Their system processed real claims under real HIPAA constraints. A wrong output costs money or violates a regulation. This is production healthcare, not a demo. > First month, we didn't pick models. We mapped data. Where is it, what are we NOT capturing, how do we structure it so agents reason correctly? Most teams skip this. They bolt a model onto the product, watch it hallucinate over unstructured inputs, and conclude AI isn't ready. > We built an enrichment layer. 34 dynamic variables assembled per claim before any LLM touches it. Pre-computed, versioned, auditable. The agent doesn't search for context. It receives clean, ranked facts. > Then we designed the architecture. Every agent follows a fixed pattern: pre-compute context, strip all patient data before the model sees it, call the LLM, validate output against a strict schema, let deterministic code decide the action. Our engineers wrote validation logic that rejects any output outside a bounded allowlist. If the model suggests something the allowlist doesn't cover, the system fails closed. > Seven agents, each locked to a single workflow. One handles denied claim follow-up. Another runs billing reconciliation. Each has its own enrichment payload and output schema, validated by a dedicated eval suite. > Their PE operating partner drove AI transformation across the full portfolio. The other portfolio companies now benchmark against this agent system. > You bolt a chat widget onto your product and call it intelligent. Production agents need enrichment pipelines pre-computing context and eval harnesses catching regression. They need audit trails and human-in-the-loop controls at confidence thresholds. We spent more engineering hours building that scaffolding than integrating the model. > We route calls across two model providers. Swapping one for the other changes nothing in the output. Model integration was the shortest line item in the build. > https://x.com/mardehaym/status/2080221212639101131