Travelers Puts Agentic AI on Live Auto Claims Calls, Closes Half Its Call Centers

Travelers launched its AI Claim Assistant on February 18, 2026, an agentic voice service built with OpenAI's Realtime API for auto damage claims. The more consequential disclosure came earlier, on the Q4 2025 earnings call, when CEO Alan Schnitzer said the claims call center population was down by a third and four claim call centers would consolidate to two during 2026. Headcount comes back in months. Buildings do not.

Key Takeaways

• Four claim call centers consolidate to two during 2026, against a claims call center workforce already reduced by roughly a third.
• More than 50% of claims are eligible for straight-through processing and 66% of customers choose it, against an industry digital FNOL completion rate nearer 40%.
• 1.5 million claims processed in 2025, roughly one every 20 seconds, with payments above $23 billion, and about 50% of first notices already arriving digitally.
• $13 billion of cumulative technology investment since 2016 at a current run rate above $1.5 billion a year, alongside an expense ratio that moved from 31.5% to 28.5%.
• Facility closure is a discrete removal from the expense base, not a trend, and it will land in the 2026 and 2027 accident years.

What Launched, and What Closed

The word agentic is doing real work in the product description. Unlike an interactive voice response system following branching scripts, the assistant conducts a dynamic conversation: it looks up policy information, answers coverage questions, helps the caller decide whether to file at all, files the claim, and then hands off to a digital flow for photo upload, appraisal, repair scheduling and rental reservation, with a live specialist available at any point.

Chief Claim Officer Nick Seminara described the technology as remarkably dynamic and responsive with overwhelmingly positive early feedback. OpenAI's Olivier Godement called it one of the most sophisticated agentic voice implementations capable of consulting, advising and supporting customers.

Scale is what makes it operationally significant. Travelers processed 1.5 million claims in 2025, roughly one every 20 seconds, with payments exceeding $23 billion, and about 50% of first notices already arrived through the mobile app before the assistant existed. The system therefore inherits the phone-based residual: callers who could not or would not use the app.

Vendor choice was deliberate rather than incidental. OpenAI was selected after extensive testing and benchmarking for claims voice, while Anthropic serves engineering productivity. Each partner covers a domain rather than one provider covering everything.

Why Two Centers Can Absorb What Four Handled

The consolidation is arithmetic on three channels, not a bet on the voice assistant alone.

More than half of all claims are now eligible for straight-through processing, moving from first notice to payment with minimal human intervention, and 66% of customers take the option when offered. That adoption rate is the unusual number: industry digital FNOL completion sits nearer 40% on Datos Insights survey data. Another 15% of claims run through advanced digital tools that accelerate handling without removing every human touchpoint, so roughly 65% of claims flow through automated or semi-automated channels before the phone is involved at all.

The AI Claim Assistant then covers the remainder. That sequence, not the assistant by itself, is what makes half the physical infrastructure redundant.

The reserving consequence is specific and it is not a trend. Allocated loss adjustment expense steps down when the fixed overhead of two closed facilities leaves the cost base, and it leaves on the closure date rather than gradually. That shows up in Schedule P as a period-over-period LAE ratio improvement, most likely in the 2026 and 2027 accident years.

An LAE ratio trend fitted across that break will misread it in both directions: it will overstate the ongoing rate of improvement by absorbing a one-time removal into a slope, and it will then understate the level once the step has happened and the slope reverts. The correct treatment is to date the closure and split the series, which requires knowing the closure occurred, and a competitor benchmarking off published expense ratios will not.

Ten Years of Spend, and the Part That Only Happens Once

The assistant sits at the end of a program that has consumed $13 billion since 2016, now running above $1.5 billion a year with nearly half directed to strategic initiatives.

Metric	2016	2025	Change
Expense ratio	31.5%	28.5%	Improved 300 bps
Underlying combined ratio	~92%	83.9%	Improved ~8 pts
Cumulative tech investment	Year 1	$13B total	$1.5B+ annual run rate
After-tax underwriting income	~$850M	$3.4B	4x growth
AI tool users	N/A	20,000+	~60% of workforce

The returns predate the AI layer. The expense ratio improved from 31.5% in 2016 to 28.5% in 2025, a 300 basis point reduction achieved while technology spending rose, with the underlying combined ratio reaching 83.9% and after-tax underwriting income of $3.4 billion, up 40%.

Travelers separates the arc into Innovation 1.0 from 2016 through 2025, being infrastructure, analytics and digital products, and Innovation 2.0, which Schnitzer describes as powered by AI. The distinction matters for anyone reading the expense ratio as an AI result: most of the 300 basis points was earned by the first phase, on cloud migration and workflow digitization that any carrier could have funded and most did not.

The facility closure belongs to neither category cleanly, and it is the item most likely to be misread. It is irreversible, which is what makes it a credible signal about the automation's durability, and it is non-recurring, which means it cannot be extrapolated. A carrier can retire a building once. The expense guidance for full-year 2026 is 28.5% against 29.0% in the first quarter, which is a normal operating range, not a step change.

So the disclosure carries two things that point in opposite directions. Permanently removing physical capacity is the strongest evidence yet that first notice of loss automation holds at a top-five carrier's volume. It is also the last time that particular saving is available to book.