Focus 02 — Reference Architecture for Regulated AI

How do we get the speed of LLMs without the E&O exposure of probabilistic decisions?

The technical wedge

The core technical problem in insurance AI is separating probabilistic understanding (LLMs reading documents, classifying intent) from deterministic action (coverage triggers, eligibility, pricing). Standing up a reference architecture — Intake · Understanding · Decision · Action · Truth · Guardrail — with auditable rules engines is what makes AI deployable in a regulated, E&O-exposed environment.

This is the sharpest differentiator versus generalist CAIO candidates: most ship demos that work in the happy path. The hard part is documenting why a coverage decision happened, in a form a department of insurance, a plaintiff’s lawyer, or an internal MRM committee can audit.

What I’ve shipped

• Six-layer architecture — documented and used in production at SWIA across multiple workflows. Pattern lives behind unifi.ai, the 5gVector endorsement analyzer, and the carrier-accounting MCP.
• GIA — Georgia Insurance AI — three-LOB FastAPI agent serving as a working reference implementation.
• 5gVector EpicConnect MCP — 18-tool MCP server bridging Applied Epic AMS and BigQuery. Under Anthropic directory review.
• ClaimSense AI + Claims-Env — RL environment for insurance claims workflows. Working example of the Decision layer with explicit reward signals.
• Nova Insurance Claims AI — Amazon Nova Hackathon entry showing the same six-layer pattern under a different foundation model.
• Power of 15 — assertion-based coding standard whose safety primitives crosswalk cleanly to MRM controls. See Focus 04 for the governance crosswalk.

What’s coming next

Market signal I track

Anthropic, OpenAI, Google insurance case studies and partnerships · Federato, Akur8, Coalition, Sixfold, Indico, Tractable, Lemonade engineering blogs · NAIC AI Model Bulletin adoption · NIST AI RMF Insurance Profile · ICAIF, NeurIPS Workshop on AI in Finance · LangChain / LlamaIndex insurance integrations.

Success metrics

• Reference repo: 200+ GitHub stars by Q4 2026
• 2 MCP servers accepted to Anthropic directory by Q3 2026
• 1 conference talk delivered, 1 podcast appearance by Q4 2026