Brokers kept asking the same questions, and leadership wanted a chatbot to answer them. This is how we built an assistant around trust instead.
*Some content has been redacted due to PII and confidentiality.
Brokers had nowhere to go for help but other people.
They answered questions themselves or passed them to customer service, and the same ones kept coming back: commission statements, census files, quote status. Leadership saw generative AI as the fix. I saw the risk creating a chatbot that looked engaged but generated more follow-ups than it saved.
Shared layer
Owned by me
Eight capabilities
Each owned end to end by me or a senior designer, with junior and novice designers paired across them
To keep many capabilities feeling like one assistant, I split the work into two layers.
Seven of us designed it, working alongside 35 engineers, 6 PMs, and an AI/ML team. I owned the shared layer, the response anatomy, context switching, controls, and handoff patterns that every capability would reuse. The three senior designers and I each took one to three capabilities end to end, with our junior and two novice designers paired across them. A weekly sync, where every designer brought work in progress, let us catch divergence before it reached engineering.
Before designing anything, we read what brokers were already asking.
Six months of support transcripts, tagged by intent, sorted the noise into eight capability areas an agent could plausibly own. Then I sat down one on one with 12 power users, casual users, and internal operators. Two asks came up in nearly every session: remember context across turns, and show where answers come from.
Nobody said “trust,” but that was the need.
Is it accurate?
Example prompt
What’s my premium tax credit total for this group?
Can brokers self-serve?
Example prompt
Where’s my latest commission statement?
Is it low effort?
Example prompt
Check my commissions, then take me back to this quote.
But an ML team can’t build to a feeling.
So I translated trust into eight signals, grouped into three questions, each specific enough for the ML team to build an evaluation harness around.
Leadership wanted
We proposed
Leadership wanted one endless chat for everything. My team pushed back.
Our brokers run complex, multi-step workflows, and a single thread couldn’t hold them. Rather than argue, we tested early. Users found the pure chat jarring and asked questions we had no agent for yet, which meant a stream of “please try again” messages and calls to the service team. I brought that evidence to leadership, and they adopted our hybrid.
First direction
What replaced it
My own design had already made part of the case.
Following leadership’s direction, I’d first designed commission information as a fully conversational experience. In testing it was slower than the statements brokers already used. So I replaced it with charts that surfaced trends and patterns, and the assistant started answering “what changed and why” instead of reciting numbers.

What shipped was one assistant built from specialists.
A routing layer sat in front of eight capabilities, and every answer shared the same anatomy: a scoped action, source chips, and a confidence band. A “What I can do” overview showed brokers the edges before they hit them, and a transparent handoff brought in a human rep when the agent should step back. Weekly critiques with ML, fortnightly tests with 4–6 brokers, and post-launch dashboards kept it honest.
Before
With Claude and Codex
Building with AI changed how my own team worked, too.
Tool approval at UnitedHealth Group is slow, for good security reasons. Figma Make hit token limits and Copilot lacked deeper reasoning, so I made the case for Claude and Codex, and got them approved. Research now synthesizes in hours instead of weeks, competitive audits start from an AI-drafted baseline, and kickoffs open with rough prototypes instead of a blank page.
What I coach now is that the bottleneck is no longer execution speed. It is defining the right problem, directing the AI, and applying taste to judge what works.
The support queue got quieter.
Support tickets dropped 42%, long-tail questions got answered in about six seconds on average, and resolution per capability reached 56%. The source-attribution pattern spread to adjacent products facing the same credibility problem.
People didn't need the assistant to be right more often. They needed to see why it was confident. Trust came from showing the work, and my job was making that a shared pattern rather than seven designers' separate interpretations.


