🆕
U.S. property and casualty (P&C) insurers had their best underwriting year in decades in 2025, nearly tripling the industry’s net underwriting gain from the previous year. And the momentum has held; on September 2, 2026, Verisk and the American Property Casualty Insurance Association reported a $31.7 billion USD net underwriting gain for the first half of 2026, one of the strongest half-year results in recent history. Regardless, the same report warned performance varied sharply by both line of business and geography, with commercial auto, umbrella liability, and other casualty lines still deteriorating. Even in a record year, then, many insurers may be leaving millions of profit on the table because of how they judge their own books. The problem is one many startups will recognize: decisions are based on averages. And averages hide individual losers. Why Insurers Think in Segments The binary nature of insurance claims outcomes – a claim either occurs or it doesn’t – means that insurers typically evaluate their model’s viability by aggregating policies into “segments: that share characteristics relevant to risk or expected claims. Think personal auto, homeowners, or commercial lines, for example. As a result, insurers risk unwittingly retaining individual negative-profit policies hidden amongst healthy segment averages. A segment can look profitable, while a share of the policies within it quietly lose money. Per McKinsey, improvements in policy-level precision can lead to a 30-50% uplift in underwriting results, which is a potential that has drawn a wave of AI companies trying to push insurance analysis below the segment level. Boston-based Earnix, for one, offers tools designed to segment customers more precisely for health, life, and P&C insurers. Newer entrants, like Soteris, a machine learning startup that recently came out of stealth with more than $8 million USD in seed funding, focus on flagging unprofitable policies one by one. How Big the Blind Spot Can Be Sunit Shah, Soteris’ founder and CEO, illustrated the problem with auto insurance particularly. “A typical insurer might think of ‘married couples’ or ‘single-driver policies’ as segments and think of all the policies within those groups the same way,” he told The Startup Magazine. “The insurers already collect enough information in the application process to [individualise analysis], they just didn’t have the tools … before now,” he added, smiling. Shah claims that some carriers his company has worked with were holding on to profit-reducing policies that in some cases amounted to as much as 30% of their book. The company says it spotted that pattern through its first product, which predicts a policy’s expected loss ratio and has been used by carriers since 2020, processing more than 100 million submissions covering over $180 billion USD in premium. Its newer tool goes a step further, flagging individual negative-profit policies directly and estimates their impact on profit and EBITDA. In theory, an insurer could then drop the policies that were never going to pay off while leaving the rest of the segment unchanged. According to the company, this can be done without changing rates, policy forms or regulatory filings.The CEO also noted early proofs of concept showed 70% to 125% gains in bottom-line profit from moving analysis to the policy level.Beyond press releases or business success, these figures point to how much can hide beneath an average that looks healthy. The Same Trap, At Startup Scale Few startups hold millions of insurance policies, but many run their businesses on segment-level thinking, including a pricing tier, customer cohort, or sales channel that can look profitable in aggregate. Meanwhile, a meaningful share of accounts cost more to serve than it brings in. Blended customer acquisition costs can hide a channel that never pays back; an average gross margin can hide a handful of heavily customized enterprise contracts that erode it. Shah’s point about insurers applies here, too. Most startups…