What AI agents actually mean for insurance

Every few years a technology shift opens a window where people who build things have an advantage over people who talk about them. We’re in one of those windows with agentic AI. Insurance sits at the center.

I’ve spent two decades building technology in industries where bad architecture gets punished: healthtech, fintech, venture infrastructure. The pattern repeats. A real capability shows up. The market floods with slide decks. The builders figure out what works while everyone else argues about terms.

Here’s what I’m paying attention to.

The shift

LLMs got good enough to follow multi-step instructions. That single change opened everything else. You can give a model a workflow: gather information, check it against rules, decide, handle exceptions, loop back when something goes wrong. That’s an agent. Software that handles variable inputs and makes reasonable decisions within defined boundaries.

Insurance runs on this kind of work. A claim comes in. Someone gathers documentation, checks it against the policy, approves or escalates. Someone communicates the decision. Someone handles the appeal. Each step involves judgment, but bounded judgment. The rules exist. The edge cases are catalogued. This is where agents do better than traditional automation (too rigid) and human-only processes (too slow to scale).

Where agents are working now

First notice of loss intake is the furthest along. A customer calls to report a claim. The agent handles the conversation, gathers details, pulls up the policy, creates the record. It knows what’s missing and asks for it. The implementations I’ve worked with cut intake time while capturing more complete information than phone-only intake.

Document processing is close behind. Insurance runs on documents: policy applications, medical records, repair estimates, legal filings. Agents read them, extract what matters, flag inconsistencies, route them to the right person. An agent processing medical records for a health claim can match diagnoses against coverage and identify what needs human review. Everything else flows through.

Fraud investigation is the one that excites me most, though it’s earlier in maturity. Traditional fraud detection flags suspicious claims using rules and statistical models. Agents pick up from there, cross-referencing information across systems, finding connections between parties, assembling a case for why something looks off. They do the legwork that used to take an investigator hours.

What I get asked about

Commercial lines underwriting comes up constantly. Can agents handle it end-to-end? The honest answer: the data is the bottleneck. Commercial underwriting depends on information that doesn’t exist in structured form. How stable is this business? What’s the management team’s risk appetite? Is this industry generating more claims or fewer? Those questions need context that’s hard to feed an agent reliably. The models can handle the reasoning. Getting them the right inputs is the hard part, and it’s solvable.

Customer service is another common question. Agents handle routine inquiries well. But most insurance customer service involves someone who is frustrated or scared. An agent that works for 80% of volume but fails on the cases that matter most creates liability. The implementations that work treat agents as triage, routing the difficult conversations to people who can handle them.

Claims adjudication is getting pilot attention. For simple auto-adjudicated claims, the results look promising. For anything more complex, the regulatory and reputational cost of a wrong decision is too high at current accuracy levels. That will change as models improve, and I expect it to change faster than most people think.

Where this gets exciting

What excites me most about agents in insurance is what they do to human productivity.

An experienced claims adjuster carries judgment that’s hard to replicate. Pattern recognition built over years. Intuition about when something is off. Relationship knowledge about providers and policyholders. That judgment is worth a lot. What’s wasteful is having that person spend four hours gathering documents and two on data entry so they can apply twenty minutes of real judgment.

Agents flip that ratio. The adjuster walks into a claim that’s already organized: documents extracted and summarized, policy terms flagged, similar claims surfaced, red flags identified. All judgment time, almost none of the assembly.

This is where the investment should go. Make the experienced people far more productive. A carrier with 50 good adjusters backed by solid agent infrastructure will outperform one with 100 adjusters doing everything by hand. I’d bet on it.

What’s worth building

Vertical agents for specific claim types carry defensible value. Generic agent platforms are becoming commodities fast. An agent trained deeply on auto claims, workers comp, or professional liability learns the edge cases that matter. That knowledge compounds over time.

Most carriers can’t build agent infrastructure themselves. They need orchestration layers, integration tooling, compliance monitoring, audit frameworks. That picks-and-shovels opportunity is real.

The UX of working alongside an agent is wide open. How does an adjuster see what the agent did and why? How do they override or redirect it? The companies that solve this interaction layer will own the workflow.

Explainability and audit tooling will be required by regulators. Every agent decision in a claims process needs to be traceable. This is infrastructure work, which is exactly why it’s a good business.

What I’m watching

Claims outcomes matter most to me. Are agent-assisted processes producing better decisions, or just faster ones? Speed alone doesn’t count for much in insurance.

I’m also watching regulatory response. The first major agent-related compliance action will tell us a lot about how the industry approaches autonomy.

Talent markets will be the real tell. If experienced claims professionals start commanding premiums because their judgment pairs well with agent systems, the model works. If carriers just use agents to cut headcount, the gains will be smaller than people expect.

This is builder territory. The slide decks will keep coming. The winners will be the ones who shipped.