Why is AI splitting insurers into two groups? Federato, Anthropic, Goldman Sachs and Tokio Marine at the Insurance Leadership Summit 2026 in New York

Why did Federato say half of top insurers are destroying value?

Host Miguel Edwards of FiveM Consulting opened the afternoon with a show of hands. Almost everyone in the room uses AI. Far fewer use it to change how insurance work is actually done.

Will Ross, CEO and Co-Founder of Federato, told the room of more than 100 senior leaders that this would not be a sales pitch. He then went straight to the numbers. He pointed to ACORD’s value creation study, which found that nearly half of the top 100 U.S. property and casualty insurers destroyed value over the period studied. In the 2021 edition, he said, that figure was nine percent.

For the first time, Ross said, results are not clustering around the same average. They are splitting into two groups. In statistics, he explained, a two-hump curve usually means someone has an advantage. His view is that the advantage comes from how firms use data and AI.

He warned against what he calls the “Rusty Bolt Trap”: taking a new tool and bolting it onto an old process. In a software company, about 80 percent of costs are people. In insurance, he said, it is closer to eight percent. So the real leak is in how risk and capital are matched, through layers of distribution and reinsurance. He pointed to Michael Hammer’s 1990 Harvard Business Review article, “Reengineering Work: Don’t Automate, Obliterate,” and summed up his talk in one line.

“You cannot draw a new future on a broken canvas.”

Will Ross, CEO and Co-Founder, Federato

He also had a warning for younger firms planning a core replacement. Some 20-to-30-year-old companies, he said, “have not yet realized that they are the new legacy.”

What did Cassie Kozyrkov say leaders get wrong about AI?

Cassie Kozyrkov, Google’s first Chief Decision Scientist and now CEO of Kozyr, started with a crowd game: shout “cat” or “not cat.” When a tiger appeared, the room split. Her point was that the system failed because the leader never said what it was for. Purpose, she said, is “the alpha and omega.”

She then compared AI to a genie. The genie (the model) is already very good. What firms forget is the lamp, meaning governance, data and security, and the wisher, meaning the leaders who must be clear about what they want.

“A genie without a lamp is just a poltergeist with better branding.”

Cassie Kozyrkov, CEO, Kozyr

She cited a Cisco study in which 83 percent of leaders plan to deploy AI agents, but only about 13 percent say they are ready. Copilots that only give advice are one thing, she said. Agents that can act across a company are another, and they demand redesigned teams and workflows. Her closing line was a reminder of what insurers really sell: “You are in the business of decision architecture.”

How are carriers moving AI from pilots to production?

Varada Bhat of the Financial Times asked Mike Ram, Head of Insurance at Anthropic, what the carriers moving ahead are doing right. He named three signals. Executives are “leaned in, not bought in.” The business is not just at the table but driving. And change management is taken seriously.

Ram said he often has senior executives build something with Claude themselves, because the value only clicks once they use it. He described a harness, such as Claude Code, as the body that lets the model, the brain, take action. His advice: give business users strong models, a safe sandbox and training, and results come fast. In one three-hour session, he said, a small group of underwriting operations staff turned their written rating procedures into Claude skills. It was not a finished agent, but it came close to matching the accuracy of the current process.

On ROI, his rule of thumb was simple: “It’s easier to measure a process than a person.” On regulation, he said a human should still review an AI-drafted email to a broker about a risk, even when accuracy is high. And he pointed to new connectors from data providers such as Verisk as a sign the plumbing is getting ready for enterprise use.

What did Goldman Sachs’ insurance survey show at mid-year?

Jared Klyman, Global Head of Insurance Asset Management at Goldman Sachs, gave a mid-year update on the firm’s global insurance survey, now in its fifteenth year. The survey was filled out in February, before energy prices moved. Inflation, he noted, had risen since then, and rates had not fallen as much as many respondents expected.

He was candid about forecasts. Most insurance investors say the next recession is one to three years away, and that has been the top answer in every year of the survey. In the meantime, he said, insurers still have to run their business, which is why many are taking on more illiquid assets to earn income. For many private assets, he added, there is a public asset with similar risk.

On AI, Klyman read the survey as bullish. Only a small share of insurers said they do not need AI. Cost cutting is the easiest place to start, he said, but likely the least useful in the long run. The bigger prize is helping underwriters find better risks faster.

What can insurers learn from banking?

James Thaler of The Insurer from Reuters asked Kelly Galanis, a Partner at Goldman Sachs who has spent more than 20 years in investment banking, what banks got right. She said the old story of insurtech disruption has not played out. The full-stack carriers are a David and Goliath story, and the large carriers “will never go away.”

Galanis said partnering is the best first step, and often the best way to decide what to buy later. Building in-house now comes down to AI talent. She expects one or two more renewal cycles of testing, then a sharp sorting of tools. Being a little behind right now, she said, is not a bad thing, because the tools keep changing and getting cheaper.

She also shared what she hears from clients. Some large banks and insurers are telling vendors that sell bolt-on features they will cut them and move that budget into their own AI plans. Her prediction: the winners will be firms that are true infrastructure, and many small point solutions will merge.

Why do specialized MGAs move faster?

William Steenbergen, CTO and Co-Founder of Federato, hosted two repeat founders. Chuck Wallace, a former Air Force pilot, helped start Esurance and later co-founded HDVI, a commercial auto insurer for heavy trucks built around telematics data. Bobby Touran started the digital surplus lines broker Pathpoint, then Rainbow, a portfolio of industry-specific commercial programs focused today on restaurants.

Both said the modern MGA owns the whole loop: product, distribution, systems and people. Feedback from the market flows back into the product, and that loop gets harder to compete with over time. Touran said keeping all data structured from day one has made Rainbow more agile with AI, though he admitted that even a San Francisco tech company finds it hard to use AI well. Rainbow uses public data on restaurants to classify risks and watch for changes during the policy term.

Wallace said AI tools make building faster but do not change the build or buy decision. They do raise the bar for vendors, whose systems must be easy to connect. On speed, he drew a clear line: move very fast on user experience and distribution tests, and slow down on systems of record. Touran’s story of an engineer who asked “What’s a renewal?” after two renewal cycles drew a laugh, and made his point that business and tech teams must keep teaching each other.

What does AI look like at the scale of Tokio Marine?

Will Ross closed the program with Robert Pick of Tokio Marine Group. Pick said most core systems in the U.S. market are now “modern enough.” The work is less about ripping out big systems and more about enriching what is there, for example pulling rating or workflow out of a policy system. His test for “modern enough”: systems must connect through APIs, run on a tech stack that is still supported, and have people who know how to support them.

He said shared services must run as a cost center, not a profit center, or trust is never built. Tokio Marine set up a global AI hub in the last year and a half to help each group company adopt AI in its own way, not to centralize it.

His advice to vendors was clear. Lead with how well you integrate, and understand you are “a moon or a planet, not the sun” in the carrier’s system. Telling a carrier it must be agentic by next Thursday is a fast way out. Asked what worried him most, he did not hesitate.

“Governance and safety is a year behind AI capabilities right now.”

Robert Pick, Group Deputy CIO, Tokio Marine Group

Topics