Soteris launches AI tool to identify insurance policies that reduce profits

Insurers can have a profitable portfolio overall while losing money on a portion of the policies within it. That is the problem Soteris is looking to address with a new artificial intelligence tool.

The US machine learning startup focused on the insurance industry announced the launch of a profit optimization product designed to identify, at the individual level, the policies that reduce the profitability of property and casualty (P&C) insurance portfolios.

The launch comes as the sector faces changing growth conditions. US P&C insurers posted a $31.7 billion net underwriting gain during the first half of 2026, nearly triple the $11.6 billion recorded during the same period in 2025, while net written premium growth slowed to 2.1%, according to data from Verisk and the American Property Casualty Insurance Association.

Against this backdrop, Soteris is targeting a less visible part of the business: individual policies that may be reducing the performance of a portfolio that looks healthy on average.

Insurers have traditionally assessed risk by grouping policies into segments with similar characteristics. This model allows companies to obtain statistically reliable metrics from a sufficiently large pool of policies.

The problem emerges when those averages hide differences between individual policies.

A portfolio can perform well overall while some policies within the same segment generate losses. For Soteris, that represents an opportunity insurers have not been able to identify with enough precision.

"We help insurers see them as unique individuals and unique policies," said Sunit Shah, founder and CEO of Soteris, in a press release.

"The insurers already collect enough information in the application process to do this, they just didn't have the tools to individualize the analyses before now," he added.

The company began working on this approach with its first product, which focuses on predicting the expected loss ratio of an individual policy. The product has been live with insurers and MGAs since 2020 and, according to Soteris, has analyzed more than 100 million submissions representing more than $180 billion in premiums.

Finding the policies that reduce profit

Analyzing that volume of data led the company to identify another problem: some insurers were unknowingly carrying a significant share of policies that reduced their profitability.

According to Soteris, those policies could represent between 20% and 30% of a book in certain portfolios, even though segment-level averages could hide them.

Its new product aims to move from predicting loss ratios to directly estimating each policy's contribution to an insurer's profit. The results are expressed through financial metrics such as EBITDA and profit.

The proposition is to identify policies that are not profitable and allow insurers to act on them while maintaining the rest of the portfolio.

"Every insurer knows they're writing policies that will lose them money. They just can't find those policies with the resources currently at their disposal," Shah said.

"That's the blind spot we built Soteris to close. For the first time, an insurer can look at a single policy and know exactly what it's worth, in time to act on that information," he added.

In its initial proofs of concept, Soteris says it has seen EBITDA increases of between 70% and 125% across the books analyzed.

"In our initial proofs of concept for this product we just launched, insurers can make 70%-125% more bottom line profit by using our product to take their analysis from segment-level to policy-level," Shah said.

Those results come from company-run proofs of concept and do not represent an industry-wide outcome.

An AI layer for insurance underwriting

Soteris' technology uses proprietary machine learning models to analyze different combinations of characteristics contained in policy histories.

The company says its platform can generate millions or even billions of segmentations, depending on the available characteristics and the volume of historical data. It then analyzes the intersections between those combinations to evaluate individual policies, effectively turning each one into what Soteris calls a "segment of one."

The tool can be integrated at different points in the policy lifecycle, including when a quote is generated, when coverage is bound or later in the policy's life. Soteris says implementation takes less than 90 days and that, once connected through an API, the system can deliver results in less than 250 milliseconds.

The company also says its customers have achieved improvements of between five and 15 percentage points in loss ratios during the first year of implementing its first product.

McKinsey has estimated that improving underwriting precision can generate a 30% to 50% uplift in underwriting results through actions such as portfolio pruning and recovering profitability from segments that appear healthy in aggregate.

Startup backed by more than $8 million

Soteris was part of Y Combinator's 2019 cohort and has spent more than five years working with insurers and MGAs before publicly launching its new platform.

The startup has raised more than $8 million in seed funding, led by Spider Capital with participation from Intact Private Capital, Amplify Partners, DCVC, Webb Investment Network and Overlook Ventures.

With its move out of stealth, the company is expanding its proposition from predicting risk to answering a question more directly tied to the business: how much money does each policy generate or lose?

"That's a tremendous amount of value already sitting inside insurers' books, just waiting for them to retrieve it," Shah said. "And once they do, they can invest those gains back into their own operations to both reduce prices and improve the overall experience for their customers."

For Soteris, the next chapter of analytics in insurance may not necessarily be about selling more policies, but about identifying with greater precision which of the policies insurers already hold actually create value.