Insurers Are Losing Money on Policies They Can’t Identify. AI Aims to Change That

Even in a strong year for the industry, one startup says as much as 20% to 30% of an insurer’s policies may be unprofitable. It has raised $8 million to help find them.

US property and casualty insurers have just posted one of their best underwriting years in decades, according to a report from Soteris, a Y Combinator-backed startup that sells software to the industry. 

However, strong overall results can hide weak spots. Soteris says many insurers are quietly losing money on a meaningful share of their policies and have no reliable way to tell which ones.

Soteris is coming out of stealth with more than $8 million in seed funding and a new product aimed at that problem. Its CEO, Sunit Shah, spoke about what he sees as the industry’s blind spot and what fixing it could mean for insurers and their customers.

Why Insurers Can’t See the Problem

Insurers set prices and judge performance by grouping similar customers, such as drivers of a certain age in a certain region. Because any single policy either produces a claim or doesn’t, actuaries have long treated differences inside a group as random noise. A group can look healthy while some policies inside it lose money, and the true cost of a policy only shows up when claims arrive, sometimes years after the sale.

“Every insurer knows they’re writing policies that will lose them money,” Shah said in a statement. He argues they lack the tools to find those policies. Working with insurers, he says, he found that even those with strong books were carrying policies that cut into profit, in some cases 20% to 30% of the total.

An Industry Moving Toward the Individual Policy

The size of the prize is why the idea is getting attention. McKinsey has estimated that insurers could lift their underwriting results, the profit made from writing policies, by 30% to 50% through coordinated moves such as trimming unprofitable business and recovering profit in groups that look healthy on average. The consultancy’s 2025 insurance report points to finer segmentation and personalization down to the individual as where the industry is heading.

Soteris is one startup working on that shift. Its approach uses machine learning, a form of AI that learns patterns from historical data, to analyze an insurer’s past policies in millions or even billions of ways. The company says a team working in spreadsheets might run dozens or hundreds. By combining those analyses, it scores each policy on its own instead of as part of a group.

The company says insurers can get a score in under a quarter of a second, when a quote is requested or at any point afterward, and that setup takes under 90 days.

From Claims to Profit

Soteris’s first product, in use with insurers and the agencies that sell on their behalf since 2020, predicts how much of each policy’s premium will be paid out in claims. The company says it has scored more than 100 million policy applications worth over $180 billion in premiums, and that customers improved their claims payout ratio by 5 to 15 percentage points within a year.

The new product goes a step further, to profit. The money from one policy is often split among the company that sells it, the one that holds the state license, and the one that provides the capital, each taking a different share. Soteris says a policy’s claims cost doesn’t reveal what it earns for the company writing it. The new tool estimates that figure and reports it in EBITDA, a common measure of operating profit.

The company says several pilot projects showed insurers’ EBITDA rising between 70% and 125%. These were pilots rather than full rollouts, and the results are Soteris’s own.

What It Could Mean for Customers

For someone buying car or home insurance, Shah says the main effect is on the insurer’s side, since the tool helps ensure each policy is profitable. He argues the benefit to customers comes indirectly.

“Insurers’ main goal is to grow their policy count, so the secondary effect is that this enables them to lower prices to grab more market share,” Shah said. “So for most people, we’d expect their rates to be lower than they would be if the insurer were not using our product.”

That is an expectation, not a guarantee. Shah’s answer focused on prices and did not address whether customers flagged as unprofitable could be denied coverage or not renewed. The company’s description of its own product, which has insurers stop writing policies that were never going to pay off, suggests that is a real question.

The Fairness Question

Insurers’ use of AI has raised concerns that people could be priced unfairly or turned down by an algorithm. “It’s a fair concern and it’s one we take seriously,” Shah said. He points to existing rules against discrimination and requirements for bias testing, and says regulation is a reason to build AI in-house instead of relying on outside chatbots. Soteris says its models let it show how much each factor contributed to any result, which he says matters to insurers because getting regulation right is essential to succeeding in the industry.

Why the Wait, and Who Is Backing It

Soteris has worked with insurers since 2020 but is only now going public. Shah says the problem took time to solve. Building the right team was difficult, and he found that experts from outside insurance weren’t automatically a fit. “Domain expertise mattered a lot,” he said. His own path started with a side business building pricing models for life insurers while in graduate school. After a PhD in economics, he traded at hedge fund Pine River Capital and helped lead a project there to build a property and casualty reinsurer. “That’s really where I learned how genuinely hard insurance is,” he said.

The seed round was led by Spider Capital, with Intact Private Capital, Amplify Partners, Foundation Capital, Webb Investment Network and Overlook Ventures also participating. Soteris would not name any customers. “At this stage, we cannot publicly share this information due to contractual obligations,” Shah said.

What to Watch

The industry’s direction is clear: more insurers want to price and judge risk one policy at a time. The open questions are how much of the potential gain is real outside pilot projects, and where it ends up. Insurers could pass savings on to customers through lower prices, as Shah expects, or keep them as higher margins while writing fewer policies. Regulators are likely to watch closely which customers end up on the wrong side of the scores.