When Ai Agents Meet Insurance Agents

17 June 2026

State Farm just rewrote the contracts for 19,000 of its insurance agents to fit AI into the work, and it’s making me think more about Agent Experience.

The word “agent” carries the tension on purpose. In software, an agent acts. In insurance, real estate, law, talent, and public service, an agent is a person authorized to act for someone else, holding authority, accountability, and trust. We need to get precise about what it’s going to mean to manage AI agents because they’re not just synthetic “employees” we’re deputizing software to make judgments based on human action. Credit scores are a good example of an opaque change that happened over time, not overnight.

One of the underrated things about people working with people is context switching and tone shifts. Whether you’re confident in what you’re saying changes how you show up to an exchange, and the person across from you reads that signal and adjusts before either of you says anything explicit. A good agent hears hesitation in a customer’s voice and slows down. A customer hears certainty in an agent’s answer and trusts the next thing they say. That back-and-forth is real information, and it moves the outcome. In a world where every B2B exchange runs through a single script, perhaps translated, perhaps personalized at the margins, we lose the read.

Building trust is really what Agent Experience has to do if we’re going to deliver an agentic web that gives customers a more frictionless, responsive, smooth experience. The trust has to run in both directions. The customer trusts that the agent acting for them understands their intent and recognizes the nuance that comes from dealing with people in stressful, real-life situations. A human agent talking to someone a few days late on their bill knows how to treat that customer with dignity while explaining the options. They exercise judgment about what to escalate and which levers to push. We’re still picturing Agent Experience with the same myopia service design had, where we assume every user has roughly the same experience and design around the average.

Like a lot of people, I think there are real risks to handing so much of our civil infrastructure over to proprietary software in the name of cost savings, and I suspect the consequences of this rush will keep surfacing in ways we haven’t priced in yet.

Yet, AI agents are being deputized into real work and have been for a while. Our obligation is to keep building the tools that let human managers exercise judgment and keep visibility into agent decisions, to understand at a granular level what’s happening.