An AI tool I designed for insurance advisors — built around the one moment that matters: an advisor live on a call, a client asking something specific, and no room to be vague. Designed from the scenario up, not the chat down.
PolicyAdvisor
Designer + build (design & front-end)
Designer-first · scenario-led
Shipped · in production
A prospect asks: "if I'm diagnosed with early-stage prostate cancer, does my critical illness policy actually pay out?" The advisor can't say "generally, it depends." They need the exact answer, the wording that proves it, and — thirty seconds later — that same answer in the client's inbox.
A generic ChatGPT search can't do this. It produces a fluent paragraph the advisor can't stand behind. So I didn't design a chatbot. I designed the tool around this exact moment — every screen below exists to serve it.
The live-call moment a generic ChatGPT search can’t survive.
Verdict → proof → source → depth-on-demand. The order the advisor’s brain needs on a live call.
Exact policy wording
Source
Locked decisions: depth control is a button under the answer (not a top-right toggle); "Draft client email" posts the drafted email as a new message in the thread; "Show full detail" grows the answer inline. Nothing pops or slides — one scrollable column, because the advisor is on a call.
A specific verdict, the exact policy wording, the source — designed so the advisor never reads a guess to a client.
Every good call ends with this. One tap closes the loop.
Email posts as a new message; Send is the heaviest control; the clause is deliberately absent.
Hi {client name},
Following our call — here's the part you asked about in writing. If you were ever diagnosed with early-stage prostate cancer, your critical illness policy would pay out, but as a partial benefit: 25% of your coverage amount, up to $50,000, payable once. The full benefit applies to more advanced diagnoses.
Happy to walk through what that means for your specific plan whenever works.
Based on your Sun Life critical illness policy — full wording available on requestThe decision made visible: the email says "full wording available on request" instead of pasting the clause. The advisor keeps the exact wording as their proof; the client gets the plain answer. Same fact, two documents, on purpose.
The advisor’s proof and the client’s email are two different documents — on purpose.
When the client pushes, the answer expands in place.
Verdict stays put; detail unfolds beneath as structured blocks; the exclusion is flagged in amber.
Early-stage prostate cancer classified as T1a or T1b under TNM staging. Confirmed by histological diagnosis.
"Invasive" (T2 and above) triggers the full benefit, not this partial one — a common point of client confusion worth pre-empting.
The detail earns its space: it only exists after a tap, it's structured (qualifies / pays / definitions / exclusion), and the exclusion — the thing a client actually challenges — gets the amber treatment so the advisor sees it before the client raises it.
Progressive disclosure isn’t a nicety — it’s fast on every call instead of slow on all of them.
The advisor catching a wrong answer is the best signal there is.
Thumbs-down opens fixed reason chips, never a text field. Two seconds, back on the call.
The interaction, locked: thumbs-down opens fixed reason chips, never a text field. One tap captures a correction the model can learn from, and the advisor is back on the call before the client notices the pause.
The earlier version asked for an essay mid-call. Nobody wrote it — so the best signal was discarded by its own input.
Not an error screen — how the engine answers when sources conflict.
What I'm confident about
What's unresolved
Same components, honest behaviour: no separate error screen — the answer card itself splits into "confident" and "unresolved", swaps the verdict to amber, and routes to the sources instead of inventing a number. The advisor never reads a guess to a client.
Real honesty is the answer behaving differently — not a disclaimer in grey footer text.
Designing from the advisor's live-call scenario instead of from a chat window changed what every screen had to do. The answer leads with the specific fact, carries the exact wording and the source, drafts the client-safe email in one tap, learns from a two-tap correction, and tells the truth when it isn't sure. The same scenario-led pattern carried into the other PolicyAdvisor AI tools — this is the one that set the language.
A chatbot answers questions. This had to survive a live call.
The whole design came from refusing to start at the chat window. Starting at the advisor's worst four seconds — client waiting, question specific, can't be vague — produced a different tool than "an AI assistant for insurance" would have.
The decision I'm proudest of is the smallest: the client email drops the policy wording. It looks like a deletion. It's actually the whole thesis — the advisor's proof and the client's answer are two different documents, and a tool that respects that earns trust a generic one never will.