Travelers Deploys AI Claims Assistant Nationwide After Eight-State Pilot

For Travelers, the first notice of loss is not an administrative prelude. Erik Roen described it as the moment that “sets the tone for the entire claim process”: a customer may have just been in an accident, may have had a car damaged in a storm, may be under stress, and may not know whether to file a claim or what should happen next.

That is why Travelers chose the workflow as an early production target for an AI claims assistant. Roen, Travelers’ senior vice president and chief information officer for claim, said the company handles about 1.5 million claims a year, and first notice of loss crosses all lines of business. The volume made it a meaningful place to deploy AI; the variability made it a useful test of whether non-deterministic agents could do more than route users through a rigid script.

The workflow has two kinds of importance. For the customer, the first interaction determines whether they understand their options and feel guided through a stressful event. For Travelers, accurate and timely information at the beginning affects downstream triage, assignment to a claims professional, self-service scheduling, body shop coordination, rental car setup, and other services that depend on the initial claim record being right.

Roen said the company began with auto physical damage, but selected first notice of loss because it could later scale to other lines. The process is structured enough that Travelers knows what information it needs to start a claim, but varied enough that customers ask different questions, need different explanations, and may not even know whether filing is the right action.

That combination is where Roen said AI agents can create value. They can adapt to different responses, clarify the process, answer questions, and guide a customer through the filing decision in ways that a fixed digital flow cannot.

Before the AI claim assistant, Travelers customers already had several ways to file a claim: through the website, the mobile app, or by calling a contact center. Roen said many customers used digital channels, but many still called. During events such as hurricanes, call volume can arrive in a compressed period, creating peak demand that is difficult to staff for and sometimes leading to longer hold times at the moment customers most need help.

The new flow gives callers an option. When a customer calls to file a claim, they can choose to interact with the AI claim assistant. Roen emphasized that the customer opts in, and that at any point they can speak to a contact center specialist instead.

Once a customer opts in, the experience is not a single monolithic bot. Roen said multiple agents work together in a way that is “totally undetectable to the customer.” The system listens, understands intent, provides explanations, answers questions, and eventually creates the claim in Travelers’ legacy system. It can also trigger follow-on activities such as scheduling a body shop or rental car.

The most distinctive part, in Roen’s account, is a “loss consultation agent.” Many customers who call are not simply trying to submit known facts into a form. They are trying to decide what to do. They may not know what coverage applies, what their deductible is, what the impact may be on their pricing if they file a claim, whether they are at fault, or whether they should file with Travelers or with another insurer. Roen said the loss consultation agent is designed to handle those situations and “many more.”

Denise Dresser said the “stat now” was that 80 to 90 percent of customers were completing their claim through the assistant. The source does not define the denominator more precisely. Roen said adoption was strong “right from the pilot” rather than requiring a long ramp. He attributed that to the amount of evaluation and testing done before release, including AI-generated synthetic callers that helped Travelers identify needed improvements before customers encountered the system.

Separately, Roen said about 35 percent of people who are given the option to use the AI claim assistant still want to talk to a person. He framed that less as a failure of the system than as a behavioral issue: many people have not yet had a good agentic voice experience. He expects openness to increase as customers encounter better agentic voice systems in other parts of their lives.

Dresser characterized Travelers’ approach as treating AI not merely as a technology or application layer, but as an operating layer. Roen agreed that the operating model was central to getting the deployment into a critical workflow.

In traditional software development, Roen said, about 80 percent of the work might sit with technology teams and 20 percent with business teams. A business group would provide requirements, participate in user acceptance testing near the end, and then the system would go live. He said the AI claims assistant required something closer to a 50/50 split between technology and business resources.

That meant bringing together, early, the people who would normally appear at different points in a program: data engineers, software engineers, data scientists, legal teams, architects, subject matter experts, and business stakeholders. Roen said this group needed to be involved not only in requirements, but in evaluations, building LLM judges, iterating with the team, and deciding how the system would be tested and deployed.

Dresser noted that this differs from the older pattern of kickoff, design, conference-room pilot, user acceptance testing, go-live, and then “ta-da.” In the AI deployment she described, the process remained iterative: the agent is given context, receives feedback, is tested and retested, and improves through repeated evaluation.

Roen said Travelers was making changes daily. Business people became familiar with the agents, the prompts, and the evaluations. That cadence contrasted with older development patterns in which business stakeholders might be brought into a room once a week to check progress.

The change-management work was also deliberate. Roen said many people’s default reference point for an automated customer interaction is a disappointing chatbot. To counter that, Travelers brought people from senior leadership down to contact-center staff “into the room” to see how the system worked. They examined the agents, the observability model, and the experience itself. Roen said that transparency helped employees and leaders get comfortable with the direction.

Travelers did not approach the deployment as a first encounter with AI. Roen said the claim organization has used AI and machine learning for 15 years and already had a responsible AI framework. That framework sat alongside the company’s technology governance process.

But Roen also described a claims-specific hierarchy that governed what the AI system could be allowed to optimize. Travelers refers to it internally as the “claim three laws.” First, the company must always pay what it owes on every claim. Second, it should provide a great experience to the customer or agent, as long as doing so does not violate the first law. Third, it should operate efficiently and effectively internally, as long as that does not violate the first two.

Those three laws, the responsible AI framework, and the technology governance process were, in Roen’s phrase, the “three legs of the stool” that allowed Travelers to innovate quickly while still treating the workflow as a controlled production system.

That structure matters because the assistant is not merely answering generic questions. It is interacting with customers about coverage, deductibles, fault, filing decisions, and claim setup. Roen’s account of the governance approach suggests that Travelers did not rely on model capability alone. It paired the agent with evaluation, observability, escalation paths, and claim-specific rules about what outcomes mattered most.

Travelers went from piloting the assistant in eight states to deploying it countrywide within two months, according to Roen. The mechanism that supported that expansion was an internal capability called Mission Control.

Mission Control streamed data in near real time, in 15-minute increments. Roen said it gave the team visibility into business outcomes, system performance, model performance, customer experience, and intervention monitoring. It was used during testing, pilot, and production.

The evaluation system had multiple layers. One was a synthetic caller created with AI. That caller could dial into the IVR and run through thousands of claim-call scenarios. LLM judges then assessed the resulting interactions for tone, accuracy, whether the system collected the right information, and related quality measures. Roen said this helped the team iterate quickly before and during rollout.

Another layer was production fail-safe monitoring. Once the pilot began, Travelers used LLM judges to look for specific problems: whether models or agents were hallucinating, creating inaccurate information, or making promissory statements they should not make. If a trigger fired, the team would be alerted. Roen said the team “could actually turn the agent off within 10 minutes” if needed.

That ability to observe, judge, alert, and shut down was central to Roen’s explanation of how the company gained confidence. He did not describe confidence as a belief that the system would be perfect. He described it as a function of instrumentation: the company could see what was happening, evaluate it continuously, and intervene.

Dresser framed this as a practical example of what it means to treat AI as an operating layer. The system was not just a deployed model attached to a workflow. It was surrounded by monitoring, evaluation, escalation, and cross-functional operating practices.

Erik Roen said Travelers selected OpenAI after “pretty extensive benchmarking and testing.” He said OpenAI performed well in that process, but Travelers was also looking for a partner willing to work closely with its team rather than simply provide technology.

The work Roen described was hands-on: building evaluations, creating judges, testing, and working through concrete model behavior. He said Travelers wanted a partner “in the trenches” that would be hungry for feedback. According to Roen, OpenAI researchers used feedback from Travelers’ testing work and the behaviors the team was seeing to inform the next versions of the model, which he said helped both organizations move faster.

Denise Dresser asked what OpenAI could do more broadly to help large enterprise customers meet the AI moment. Roen’s answer was less about a specific feature and more about commitment and operating discipline. He said companies need dedicated people who can pick a problem, work closely with OpenAI partners, and be comfortable with a test-and-learn approach.

He also warned against waiting for perfect information. In his view, evals, LLM judges, and observability are what give a team the confidence to go faster. Without them, AI systems can remain suspended too long before reaching production.

Dresser raised the common concern that AI may take jobs away. Erik Roen said Travelers sees the issue differently, and he linked the workforce approach to the same change-management discipline used in the technical rollout.

Travelers invested time not only with the central project team, but also with contact-center employees. Roen credited contact-center leadership with guiding the organization through the change and said local leaders were given time to understand what the deployment would mean.

The workforce response, as Roen described it, was upskilling, reskilling, and redeploying talent into other parts of the claim organization. He said Travelers believes AI will become part of all functions at the company, so the organization is leaning into training people for that environment rather than treating the claims assistant as an isolated automation project.

Roen did not present this as an abstract cultural principle. It was tied to a specific operational shift: some claim intake work can now be handled by an AI assistant, while the organization looks for ways to move people into other claim functions and prepare them for AI-supported work across Travelers.

Travelers is continuing to expand the first-notice-of-loss assistant beyond auto physical damage into other lines of business. Erik Roen said the goal is to handle different kinds of claims through the same general approach.

The company is also looking beyond intake. Roen said Travelers has about 20 other AI initiatives underway within claim, with more across the broader company. He framed those initiatives through the same three laws: pay what Travelers owes, provide a strong customer or agent experience, and operate efficiently and effectively.

Roen’s account of the deployment rests on concrete operating choices: cross-functional staffing, daily iteration, synthetic testing, LLM judging, near-real-time monitoring, fail-safes, and a claims-specific hierarchy for deciding what the system is allowed to optimize. The result, as he described it, was not simply an automated intake channel, but a first-notice workflow Travelers believed it could expand because it could test, observe, and intervene as usage scaled.