Building the First AI Claims Agent in US Insurance

Building the First AI Claims Agent in
US Insurance

Timeline

Timeline

3 months

Year

Year

2025

Contributors

Contributors

1 PM, 2 Developers, 1 Marketing manager

Role

Role

Lead designer

What Fairmatic Is

Fairmatic uses AI to price commercial auto insurance by how safely someone actually drives, not just their age and zip code. Their customers are fleet operators: trucking companies, delivery services, gig platforms. When a driver gets into an accident, a claim goes through Fairmatic. When I joined, that entire process was handled by human agents on the phone, during business hours, with nothing in between.

The Problem

86% of claims at Fairmatic were lost to slow, manual, outdated processes . 45% of it was handled entirely offline, a third of drivers who started a claim in the app never finishing it.

86% of claims at Fairmatic were lost to slow, manual, outdated processes . 45% of it was handled entirely offline, a third of drivers who started a claim in the app never finishing it. We built the first AI voice agent for claims and roadside assistance in the US. Direct claims reporting went up 89.2%.

Four Ideas, One That Stuck

We explored different ways to close the gap through a design sprint. Four stuck: automated emails, proactive notifications, a redesigned reporting flow, and a voice agent. The redesign was the safe bet -cleaner questions, guided photo capture, a real improvement over what existed. It was still a form, though, and a lot of drivers didn't want to fill one out one-handed after an accident.

The idea that kept winning arguments was voice. What if you could just talk? Nobody had time to build it properly, so I took it into a hackathon instead, built a working version, and won. That's what turned a concept nobody had bandwidth for into a prototype nobody could dismiss, and got it real engineering time on the roadmap.

What Driving Taught the Design

Early tests surfaced the obvious problems first - too many questions at once, no way to skip ahead, drivers wanting a transcript even after hearing a spoken summary. The insight that actually changed the design came from watching, not listening: drivers barely looked at their phones while talking. So the whole interface shifted from screen-first to audio-first. A checklist that ticks off by itself, motion only when the agent is actually speaking, a structured summary instead of a wall of text.

A second round of testing surfaced what we didn't have time to fully solve - no Spanish support yet, sessions running long when drivers added unnecessary detail, some hesitant to speak freely to what still felt like a stranger. We shipped anyway, with what we'd gotten right, and kept the form and app running as a fallback for anyone who didn't want to talk to an AI after a crash.

What Happened

89.2%

increase in direct claims reporting

4.5/5

average driver satisfaction rating

50+

agent-hours per week reallocated from intake to complex claims

The bigger shift was in the data. Claims used to run almost entirely through a third-party administrator and Fairmatic had no visibility into why a claim was made, so pricing couldn't respond to it. With direct, structured reports, underwriting could finally tell a routine claim from an unusual one and adjust terms accordingly. That data flow is what started moving loss ratio, not just the app experience.


The agent was also WCAG 2.1 compliant and built around insurance regulation that guards against leading the reporter - nothing in the flow could nudge a driver toward one account of what happened over another.


The best proof wasn't a metric. A driver in testing didn't realize they'd been talking to a bot at all.

The bigger shift was in the data. Claims used to run almost entirely through a third-party administrator and Fairmatic had no visibility into why a claim was made, so pricing couldn't respond to it. With direct, structured reports, underwriting could finally tell a routine claim from an unusual one and adjust terms accordingly. That data flow is what started moving loss ratio, not just the app experience.


The agent was also WCAG 2.1 compliant and built around insurance regulation that guards against leading the reporter - nothing in the flow could nudge a driver toward one account of what happened over another.


The best proof wasn't a metric. A driver in testing didn't realize they'd been talking to a bot at all.

“I didn't realize it was a bot!”

Kenneth Gregory

Driver | Hop Skip Drive

What I'd Do Differently

Some drivers' phones flagged the agent's calls as “Insurance Scam.” A working agent means nothing if caller ID can't tell it apart from an actual one - the kind of failure that's invisible until real users hit it at scale, and the thing I'd test for earlier next time.