Underwriting the Cost of AI

An interview with Will Ross by Sean Murphy

Will Ross did not set out to work in insurance. He was at Stanford, between an MS in climate and atmospheric modelling and an MBA, collaborating on a wildfire modelling project in a corner of northern California, when insurers started knocking. A better model, they told him, would help them solve some real problems. He took the meetings and what he found there redirected his career. "There's a lot of good people trying really hard," he says of the industry, but they were working inside systems that were holding them back. The way to help was not a better wildfire model. It was to rebuild the machinery of insurance around what would come to be called agentic AI.

That decision became Federato, where Ross is co-founder and chief executive, and where the mission is stated plainly as changing the way insurance work gets done. The company describes itself as AI native, a term that has been stretched to mean almost anything, which is why Ross is unusually precise about it. Cloud native once meant systems built to run elastically so that computing overhead could be stripped out; AI native, in his telling, is the same instinct applied to a new constraint.

The cost that matters now is tokens, the units by which large language models charge for the text they read and generate. What separates serious systems from superficial ones is how efficiently they serve the right information to the model at the right moment. Feed a model everything and costs balloon, accuracy suffers and hallucination creeps in. Serve it only what it needs and the opposite happens. That discipline cannot be added later, and it is the heart of his case against layering AI on top of legacy technology.

Will is willing to put his money where his mouth is on this argument. Federato caps what customers pay, so its price does not rise with their token usage. Insurers are not token efficient, Ross admits, and they are frightened of AI costs running away from them; Federato takes on the job of keeping them predictable, which is only possible if the underlying system is genuinely efficient. In an industry where he estimates around eighty per cent of technology spending goes to supporting legacy systems, the pitch is that AI should retire the old core rather than pad the budget sitting on top of it.

What that looks like in practice is a product called Control Tower. Insurance is a portfolio business, Ross explains, where no single risk matters in isolation and what counts is how risks correlate across the book. Historically, insurers analysed the portfolio quarterly or annually, then published a fifty page PDF of underwriting guidelines and hoped underwriters would apply it at the moment a deal crossed the desk. Control Tower inserts that same guidance directly into the point of decision, so portfolio steering reaches the individual deal, which is still where risk is transacted.

The results Federato reports, and these are the company's own figures, include 2.4 times the volume of quotes from the same underwriting team, and 5.5 times the share of bound business sitting in high appetite territory. Numbers like those, Ross suggests, are what persuaded Goldman Sachs to lead the company's 100 million dollar Series D, after speaking to customers representing roughly eighty per cent of its revenue. Senior executives are ready to move, he argues, because for the first time they can use the technology themselves. "They couldn't touch and feel the cloud."

The future he describes is a reshaping of standard operating procedures. The first line underwriter moves up a tier, commanding a small portfolio while the AI performs the initial motion and becomes, in effect, a junior employee. Humans check, apply judgment and hold the relationships; the one per cent of the job that is human contact still matters enormously, even as the rest transforms. Ross has no crystal ball on the timeline. What he will say is that AI can already handle the vast majority of an underwriter’s tasks, even when it comes to complex risk. The work that remains is not necessarily the least important: human judgment, relationship-building and the conversations that ultimately move a deal across the line can represent a small share of the overall tasks while carrying a disproportionate share of the value. Humans will continue interacting with humans, but the roles they perform and where they spend their time and expertise will likely change more than anyone is yet appreciating.