Scout InsurTech Interview with Will Ross

Will Ross is the Co-Founder and CEO at Federato,  an AI-native insurance platform modernizing and automating the full policy lifecycle, from submission triage all the way through to billing and policyholder management. Will was interviewed by Chris Luiz Co-Founder and CEO at Scout InsurTech.

Will, you've said that carrier executives are telling you in private to bring on the robots. What has shifted?

“This is the first time in two decades that we've seen a new technology introduced to the marketplace that we can all experience as consumers, and therefore intuitively understand what it can do to transform jobs within an enterprise. What we're hearing executives talk about more and more behind closed doors, and increasingly in public, is that they're looking at these technologies and saying, "I can see how this is going to change the way work gets done in our industry, and I'm not inherently looking to be first or to take outsized risks, but this stuff doesn't feel all that risky to me." They're all trying to figure out how they make the right-sized move at the right time for their unique business circumstances. That doesn't mean everything's going to change tomorrow, but the majority of people are actually looking to evolve their organizations. I don't think that's been true in over 20 years in insurance.”

If AI can replicate 94% of underwriting decisions identically, what does the underwriter's job actually become?

“When we talk about replicating a meaningful percentage of underwriting decisions, we're really talking about the inputs to deterministic systems like raters and quote and policy template engines. It's important to remember that in insurance, a deterministic outcome is ultimately what we're after. What we've found interesting is that by pointing machines at the first pass of analysis, we're often able to shift the underwriter into more of a position of command and control. Rather than being a desk clerk assembling information to feed into a rater, they're reviewing how AI populated those inputs, understanding the logical flows, and identifying what I would almost characterize as “cuteness” in the behavior of the model, or inserting nuance where there isn’t any.

Humans have an uncanny ability to sense what is realistically going to help a deal get done. Say a broker comes to me and says they're looking for a certain deductible on a risk. If AI can do a thorough job of identifying why we can't use credits and debits to get there, I can increase the likelihood that the underwriter will quickly reach the conclusion that they need to call the broker and say, "If you want to be at this price, you're going to have to work with me on one of these other variables." Without that AI-driven research, it's a lot easier for an underwriter to turn a blind eye and take what seems like the most convenient path.

AI is more consistent and more thorough in the overall evaluation of a risk. Where it tends to miss is on the nuance at the final intersection. And I think it's important that we say that clearly rather than falling into the tautology everyone repeats, which is just ‘underwriter pairs with machine.’ To be direct: machines are better than the underwriter at the majority of the process. I believe there will be far fewer jobs that look like the underwriter title does today. It doesn't mean there won't be people with that title, but there will be far fewer people doing what underwriters do today, because the vast majority of what underwriters do is not that uncanny human intelligence. Anyone who disagrees hasn't spent much time underwriting. The senior executives know this. Everyone knows the vast majority of underwriter time is not spent that way. It's not a secret.”

A lot of carriers are trying to move on AI but struggling to get past the pilot stage across the full policy life cycle. What's going wrong?

“What we've observed more than anything is that there are some very simple AI use cases people have been able to pick up and run with, things like basic submission triage. What inevitably happens after that initial success is you look to do one of two things: scale it across many lines of business, or you listen to users who say, ‘If AI can help me do A, how can it help me do B and C?’ As you start to broaden the scope, you move out of the bounds of those lightweight solutions that add value on their own. And as you do, you encounter a lot more interdependency on existing systems.

The problem is those existing systems cause an inheritance upward through the AI. Take latency. If an AI agent hits an API endpoint in a legacy core system and that endpoint takes 40 seconds to respond, AI cannot speed up that underlying compute. Factor in that the AI actually needs to hit that endpoint 40 times to do a single rating, and you've literally seen outcomes where servers crash from the latency and memory build up on limited compute.

You also inherit poor tooling. I was recently sent a spec for a production implementation that had 87 API endpoints to integrate to a single production system. That's just bad architecture, regardless of AI. And then there's data hygiene. A lot of these systems have very poor data schemas, and it turns out that just dumping information to AI and letting it figure things out doesn't work. There's significant work going into engineering the context that gets presented to these models so they can make real sense of it, and legacy core systems are not well-positioned for that.

In all three of these cases, as carriers look to expand and make their AI efforts more meaningful, they hit these basic first-principles limitations of underlying systems. AI can do some uncannily impressive things on top of them, but the second you try and do something real, you will be subject to the limitations of those systems.”

ACORD data shows 47% of insurers are now destroying value in underwriting, up from 9% in 2021. What's driving that?

“There are two sides to this, a more benign one and a more troubling one. On one side, interest rates are higher. If you're thinking about an insurer broadly accumulating capital, they can afford to take more risk on the underwriting side and rely more on investment income when rates are high. That explains some of the drift. But the dramatic nature of it, and the fact that these numbers haven't been this high in decades of studying this data, points to something else: a trend toward a bimodal distribution of combined ratios.

What that means is you have two humps rather than one. Think of a normal bell curve, but with a second bump. A bimodal distribution typically represents an advantage being gained by one group. A classic example is students cheating on a test: there's the regular average, and then a separate cluster of very high scores from the people with the test bank. It's a common way professors spot cheating.

When you apply that thinking to the insurance market, what we believe is that AI is playing a role, but not primarily through how carriers are adopting it. It has more to do with how brokers have adopted AI. Certain brokers have made massive investments in figuring out where to market difficult risks. Some carriers have built strong defenses against that adverse selection, and some haven't. The ones who haven't are losing, and you see it in the divergence: a lot of companies with combined ratios in the high 80s, and a lot with combined ratios well into the 100s, with fewer than historically hovering in that mid-to-high 90s range that used to be the norm.”

You've argued that brokers have invested far more in AI than carriers, and that they're using that edge to route bad risk toward carriers who can't tell the difference. How real is that problem?

“Extraordinarily real, especially when it comes to difficult risks. And the broker doesn't need to run fancy machine learning in real time to do this. All they need is to identify, given a sector of risks, where they've had the most success placing them. If they've had consistent success placing a poor class of risk with a particular carrier, that carrier is going to receive every single submission that fits into that class code.

Where we think this is heading is a lot like cybersecurity. In cybersecurity, there is no such thing as a perfectly secure system. You're constantly making your system more secure while someone else is constantly trying to penetrate it. It's an arms race. That's exactly what's happening here. The stronger the quality of what you're doing on the insurer side, checking appetite, dynamically turning on and off pockets of risk you want to select against, the better guard you'll have against someone else's system that's trying to identify whether you're a market for a known weak class of risk.

There's also more consolidation on the brokerage side than the carrier side right now, which means the collective investing power on the broker side is somewhat stronger. But we are seeing an insurer strike-back moment. And again, this furthers the winners and losers dynamic. For the carriers who aren't maintaining discipline, it's not a matter of letting a few bad risks in. Floods of risk can come onto your books very quickly in the current environment.”

Scout InsurTech Thanks Our Partners