A POC I recently built — grounded in years on a group-benefits insurance product platform — showcasing how agentic systems can be built around the insurance pipeline: quote → underwriting → servicing → claims. Mock data, full pipeline, honest scope.
Auto quote generation and agentic underwriting. A customer finishes a property quote in minutes; behind the desk, AI agents assemble the risk picture, run the underwriting playbook, and draft the decision — and a human underwriter approves, refers, or declines. Everything in the demo exists to serve those two promises.
This is a proof of concept, not a production system — a use case turned into a runnable proof of concept. The insurance thinking is real (it draws on years building a group-benefits product platform — rules engines, underwriting workflows); the POC shows how agentic systems wrap that pipeline end to end: messy inbound leads, missing and contradictory data, escalation rules — built as something you can click through, not a slide deck.
The full property/fire lifecycle across role workspaces — applicant/producer quote-to-submit, distribution channels, underwriter referral queue, case management & claims — over an event-sourced data spine, with a climate cat-model scoring every address and clickable architecture walkthroughs (GCP vs AWS) for the engineering view.
Where the agents work: pull a morning queue of imperfect leads, watch intelligent prefill fill every service-owned field (assessor, hazard/GIS, aerial, permits, cat-model), then submit — most auto-quote, runbook declines auto-send, conflicts land in human review, and customer-only gaps produce exactly one well-scoped outbound follow-up.
Each line of business is a playbook; a runbook is that playbook resolved for a context (state, business line); every field and rule carries effective/expiry dates — a new state is an activation, not a release.
The model drafts; deterministic gates decide what needs a person. High-risk actions queue for approval, every decision is logged and replayable, and escalation is a first-class terminal state — not an exception.
Agent output quality is measured against fixtures with known failure modes, so iteration is driven by pass/fail evidence instead of eyeballing — the same discipline the platform applies to underwriting rules.
Reset, seeding, and routing are deterministic and replayable; LLM calls are a governed, per-session capability — so the demo is always walkable, and the agentic parts are additive, never load-bearing for the tour.