In a GEICO claims pilot, agents research and ask while people approve the queue

In a pilot run three times in GEICO's claims group, agents did the research and asked the questions. People approved the research queue and the worksheets that fill the gaps. Ford Prior, a senior staff engineer at GEICO, described it on the AI Engineer show.

What was said

Prior said GEICO sees itself as a risk management company, and that this thinking shapes how it builds software. His problem was that coding had sped up while the work around it, such as decisions and sign-offs, had not.

His answer was a war room: a chat channel holding the people relevant to the project, with a fleet of agents working alongside them. A markdown rulebook called FleetKit steers the agents through a nine-step process.

The agents first put their research findings in a queue. People can comment on it and approve it. Then the agents move on to building.

If product details are missing, Prior said, the agent finds the owner from a roster and mentions that person. The worksheets the agents create need human approval. Prior also named judgment as a GEICO input: who can decide what, and at which gate.

He said the agents' questions were good. As he put it, "there wasn't ever a question of, like, why is the agent asking me dumb questions?"

Why it matters

Our reading is that this design speaks to a hard question: who answers when an agent-assisted change goes wrong? Prior did not frame it that way. But a roster of owners and set approval points would let a failure be traced to a person, not just a tool.

For leaders, the lesson is to write down who can decide what before agents arrive. Prior treated that decision map as a core input, alongside the tools and the company's own methods. In claims, those decisions touch real customers.

The other side

Prior did not say, in these excerpts, that this keeps accountability with people. That link is our inference. He also gave no data on failures, and he did not say every step has a named approver.

His own slide showed heavy human input early and less near the end. He named stronger gate enforcement as the biggest win still to come, which suggests the gates are not yet as firm as he wants.

The evidence is also narrow. It rests on three runs across rentals, damages and salvage, led by one driver on his own laptop. Prior said GEICO is still working out where to put humans in the loop and how to deploy this across the organization.