ShelterPoint Insurance AI Assistant

Overview

ShelterPoint is a fictional New Jersey property and casualty carrier, created so a demonstration could be built against a full product set and real state regulation rather than a sanitised sample. I designed and built it for a leading specialist insurance IT services and implementation firm, to show a large US carrier what hyper-personalisation looks like when a digital experience platform and an AI assistant work from the same policyholder data. It is live at shelterpoint-demo.vercel.app, and the training architecture behind it is published at shelterpoint-demo.vercel.app/training.

The Problem

Most carrier assistants read the FAQ page. Ask one about flood and it explains what flood insurance is. A policyholder does not want the definition; she wants to know whether her basement is covered, and the answer depends on her form, her endorsements, her FEMA zone and her state. The same is true of the site around the assistant. A home page that greets everyone with the same three products is not personalised, whatever the platform is called.

The demonstration had to show both halves working together for one person, and it had to do so inside the compliance boundaries a regulated carrier actually lives with.

One Policyholder, Everywhere

The demo is rendered for a single returning policyholder: an HO-3 at $485,000 in Princeton, two vehicles, the verbal tort threshold selected, a renewal in August, FEMA Zone X, no flood cover, no umbrella, and a named agent at the Princeton office.

The home page is built from that record. The hero states her renewal and her dwelling limit against her county's construction costs. The coverage cards are the gaps in her policy, not the carrier's product list: flood for a Zone X property with none, umbrella for her liability exposure, a dwelling limit review, and the water backup endorsement her HO-3 does not carry. A severe weather alert for her county sits above it all.

The Assistant

The assistant answers from the same record. Asked whether she is covered if the storm floods her basement, it separates a burst pipe from rising water, names flood as excluded from every homeowners form, notes that sewer backup needs the endorsement she does not hold, adds the NFIP basement limitation she would face even with a flood policy, and then offers to route her to her agent by name. That is one answer; it is what every answer does.

It responds in English and Spanish as native generation rather than translation, and when it hands off to a person the handoff carries the full transcript, the customer's intent, the data collected and a recommended next action.

Five Layers of Knowledge

The assistant is trained in five layers, and the layering is the point.

Educate, Never Advise

The compliance layer is the part an insurance audience tests first. The assistant explains options and never says which to choose. It presents both New Jersey tort thresholds without recommending one. It will not compare carriers, disclose underwriting criteria, give legal advice, bind coverage, process a payment or touch another policyholder's data. Every coverage statement carries the policy terms caveat. At any of those boundaries, and on any sign of distress, dispute or uncertainty, it routes to a person with context.

Technology Stack

Status

Live at shelterpoint-demo.vercel.app with the architecture page at /training. ShelterPoint, its licence number and its policyholder are fictional; the New Jersey rules, forms and thresholds are real.