Building an AI Platform Around the Agent's Workflow

I managed this project from exploration to release. I ran the user research, wrote the requirements for every product change, and designed the UI, working closely with the CEO on direction.

I also built. Working with Claude Code, I turned my designs into working UI and functional changes and submitted them as pull requests, so engineers reviewed finished work instead of building it from specs. And I did the prompt engineering behind the AI insights, so they sounded human and were clear about what was a prediction and what was fact.

Atidot builds AI analytics for the insurance industry. Its original platform was built for carriers; Magnet is the broker-facing sister platform I led product for, built for the agents at brokerages who sell and service policies. It flags which policyholders are likely to lapse, buy more, or need attention. The predictive models were strong. Like many deep-tech products, the challenge was turning that strength into an experience users could act on.

The original platform surfaced valuable predictions, but workflows were fragmented, and many features showed what was possible without building a clear path to doing it. The core problem was simple to say and expensive to ignore: users had to connect the dots themselves.

Users struggled to answer practical questions: Who should I call today? Why is this customer being flagged? What should I say to them?

One constraint shaped every decision that followed: engineering capacity was limited, so Magnet had to be built on the original platform's models and data rather than from scratch.

The data underneath was strong. The interface left the interpretation to the agent.

Agents came to the same data as carriers, but with different goals, so Magnet had to answer different questions.

Rather than start from the interface, I started from the business. I sat in on prospect meetings and sales demos, ran client workshops, led onboarding calls, and presented the platform directly to agents. With limited engineering capacity and agents loyal to their own tools, a feature wish list was never an option, so I set one prioritization principle and ran every decision through it.

I made the call to cut before adding. I audited every flow in the original platform against usage data and feedback, then aligned engineering on what Magnet would leave behind and what it would carry over. The redundant and unfinished flows that created cognitive load without driving anything stayed behind, and the navigation labels changed with them, from what the system stored to what the agent was trying to do. Then I built Magnet's interface: one set of components instead of several near-duplicates, a type scale that held across pages, and a palette chosen for contrast and color-blind legibility. The outreach status set came out of the same work, giving agents the record of who they had already called that the original platform had never kept.

The status dropdown was quietly the most important thing we shipped. Every time an agent set one, we found out what happened to a prediction. Did they pick up? Did it convert? Was the model right, or just confident? We had a call log before, but unfortunately no one used it. Those outcomes fed back into the model, so every call an agent logged helped sharpen the next round of predictions.

On sales calls, the same objection kept coming up: clients did not trust "black box" predictions. I treated trust as a product problem and prioritized three changes across the platform.

The same gap surfaced in nearly every customer conversation: users trusted the insight but froze on the next step.

I prioritized an outreach composer that generates personalized outreach (email, SMS, or call script) directly from the platform's segmentation, drawing on the exact signals that flagged each policyholder. I scoped it to data we already had, so it needed no new integration and was the most direct way to make the product useful day to day. In the end, 85% of active agents copied an AI-drafted message into their own tools.

One of the most requested capabilities was a way to act on a whole segment at once. I designed a campaign page where users view a segment of flagged policyholders and copy or download the list straight into the marketing tools they already use. Users could go from a segment to a running campaign in minutes.

While we were building Magnet, there were a lot of discussions about which features would match the "expected" AI experience. I championed an AI chat assistant for the dashboard, built a prototype with Claude Code, and pushed for it internally. My hypothesis was that users expected it, now that chat assistants were everywhere, from ChatGPT and Microsoft Copilot to Salesforce and even the McDonald's app, and that it could become the interface for expanding our capabilities. Once we tested the idea in client interviews, it was clear the demand wasn't there. Most interviewed clients preferred a platform that was easy enough to use on its own over one they had to ask questions. We were about to build a feature users didn't need.

Our interviews caught it before we built it, so I moved that engineering time to the AI outreach composer. The chat answered questions clients weren't asking. The composer did the step they were stuck on: turning an insight into a message they could send a policyholder.

Over Magnet's first six months, the principle held up: agents moved from insight to action, and kept coming back.

I'm now managing the next product: an AI-powered policy servicing platform for an insurance distribution network, built against a signed client commitment for about 50 advisors who manage books of business across 20 carriers.