Insurers Are Betting on AI to Find Hidden Profit. Can the Models Hold Up Under Pressure?
U.S. property and casualty insurers just posted their best underwriting year in decades, recording a $60.9 billion net profit in 2025. Even so, insurers are increasingly under pressure to improve profitability at a time where the cost of insuring people and assets is harder to predict.
A combination of inflation, geopolitical shocks and climate volatility has made insurance pricing highly unpredictable, according to Arthur J. Gallagher & Co., a prominent insurance brokerage. As risk becomes more complex, segment-based pricing can hide individual policy risks. Two policies that look similar on paper and sit within the same risk segment can produce very different financial outcomes for an insurer.
This reality is pushing more insurers toward granular underwriting approaches, where AI and machine learning models evaluate individual policies on a much broader set of variables than relying strictly on segment averages or Generalized Linear Models (GLMs). The potential payoff is substantial. McKinsey has estimated that this kind of policy-level precision could deliver a 30% to 50% boost in underwriting results.
From Lemonade to Soteris, several insurtech companies are already building around this shift. Soteris, for instance, deploys machine learning software to score the expected profitability and risk of individual policies. This data helps insurers spot policies that are likely to lose money before they writing or renewing them.
For the company’s CEO and founder, Sunit Shah, the economic opportunity extends far beyond any individual insurer.
“There’s over $5 trillion of insurance written worldwide each year,” Shah told International Business Times, arguing that even a relatively small improvement in how insurers identify profitable and unprofitable policies could have significant implications for the industry.
The Risks Behind AI-Driven Underwriting
Regulators and insurance leaders are not entirely certain that the insurance industry is ready for the AI it is deploying. According to Grant Thornton’s 2026 AI Impact Survey, 44% of insurance executives say governance or compliance challenges have already contributed to AI underperformance. Meanwhile, only 24% of insurance leaders are very confident they could pass an independent AI governance review within 90 days.
This fear is not unfounded. The same policy-level precision that could improve insurers’ margins also widens the surface area for error. And when machine learning models repeatedly surface wrong predictions, they can limit access to essential coverage, especially for marginalized groups. When this pattern emerges, it is cause for a regulatory review.
In light of this risk, regulators are intensifying oversight towards insurers relying on AI-driven risk prediction. The New York Department of Financial Services (NYDFS) now monitors insurers in line with its Circular Letter No. 7, to ensure AI systems do not inadvertently discriminate against protected groups.
The National Association of Insurance Commissioners (NAIC) has also launched a pilot of its AI Systems Evaluation tool across 12 states from March to September 2026. The tool is designed to verify that high-risk AI models aren’t hiking premiums unfairly, are free of bias and remain under human oversight. “We don’t want to stand in the way of innovation that generally serves consumers,” NAIC president and Virginia’s insurance commissioner, Scott White, said in a keynote at the NAIC’s 2026 Spring National Meeting in San Diego. “But we do want to make sure that it is used transparently, fairly, and in ways that hold up to scrutiny.”
Shah does not dispute the concern. “Insurers need to be smart with how they use analytics on policies”, he said, to make sure their decisions aren’t “unfair or discriminatory.” Lemonade points to a potential solution, suggesting that ML models can in fact help curtail biases in premium pricing. In a company blogpost, CEO Daniel Schreiber wrote, “machines can overcome the biases that contaminate their training data if they continuously calibrate their algorithms against unbiased data.”
Beyond the regulatory question lies another concern: that sudden market shocks may throw off the very predictions that AI models are built to make. A spike in vehicle repair costs, or a major climate event could distort the historical data ML models rely on — making a policy look temporarily unprofitable for reasons that have nothing to do with its long-term risk.
From Shah’s perspective, modern models are built to avoid “overfitting” or reacting too heavily to edge cases. Because Soteris’s systems ingest an extensive array of data points collected by carriers, he notes that they are naturally diversified against single macroeconomic shocks.
Furthermore, backtesting on historical data helps prove the technology can handle volatility before it ever goes live. As Shah points out, “Insurers tend to be cautious entities by nature; they are not going to move forward unless they are confident that the results of a historical test are representative of what will occur in the future.”
Policy-level risk prediction is not the insurance sector’s first contact with artificial intelligence.
The technology has been widely adopted for back-office tasks, from data extraction to claims processing. But this shift toward predictive underwriting is shaping up to be its most high-stakes application yet.
Soteris’s recent $8 million seed round is a signal. Investors are betting that the market for policy-level risk prediction will grow significantly in the coming years. As adoption scales, the market will soon find out how closely the promise of AI-driven underwriting matches its reality.
