What's next in AI?

Record date: 10/16/25
Air date: 12/3/25

Discover how artificial intelligence is redefining the future—faster, smarter, and more customer-focused than ever before. In this episode of the Future of Risk by Zurich North America, host Justin Hicks discuss the rapid evolution of AI, highlighting trends like agentic and generative AI, explainable AI, and hyper-personalized customer experiences with Amy Nelsen, Head of Underwriting Operations for Middle Market and Madhu Ramamurthy, Chief Information Officer. The conversation covers how Zurich is leveraging AI to automate routine tasks, enhance underwriting and claims with advanced tools, and deliver actionable insights—all while ensuring strong human oversight and regulatory compliance. Scaling AI for meaningful return on investment and navigating regulatory ambiguity are recognized as key industry challenges, but Zurich’s approach emphasizes starting with low-risk, high-impact use cases. The podcast underscores that AI is here to augment, not replace, human expertise, empowering underwriters and claims professionals to focus on what matters most: building relationships and delivering value to customers. Stay tuned to learn how Zurich is helping businesses meet tomorrow prepared through responsible, innovative AI adoption.

In this miniseries, other episodes include:

10/22/25: What is AI delivering so far
11/5/25: 5 ways everyone can benefit from AI today
11/19/25: Dark side of AI

Guests:

Amy Nelsen
Head of Underwriting Operations
Zurich North America
Connect on LinkedIn

Amy Nelsen has led numerous roles during her 25-year career with Zurich North America and currently heads Operations and Technology for Zurich North America's U.S. Middle Market business—where strategy meets innovation. With a proven track record in streamlining processes and driving digital transformation, Amy enables the organization to operate smarter and scale effectively. Known for blending operational excellence with cutting-edge technology, she brings practical insights and forward-thinking ideas to every conversation. When not optimizing workflows, Amy explores emerging tech trends and mentors future leaders.

Madhu Ramamurthy
Chief Information Officer
Zurich North America
Connect on LinkedIn

Madhu Ramamurthy is an accomplished technology executive with more than 25 years of experience leading large-scale business and IT transformations within the insurance and financial services industries. As the Senior Vice President and CIO of Zurich North America, he is responsible for overseeing the IT application needs of the organization’s business units and leads several key transformation initiatives. Madhu harnesses advanced technologies and artificial intelligence to modernize products, pricing, processes and the overall user experience.

Throughout his career, Madhu has progressed from software development roles to senior leadership positions, demonstrating a strong track record in driving underwriting, policy administration and claims transformations at leading insurers such as USAA and State Farm. He holds a master’s degree in computer science, the CPCU designation and is a PMI-certified Project Management Professional.

Outside of his professional role, Madhu is actively involved in community service in the Chicagoland area. He enjoys volunteering, traveling with his family and playing tennis.

Host:

Justin Hicks
Communications Business Partner
Zurich North America
Connect on LinkedIn

Justin Hicks is a Communications Business Partner at Zurich North America and supports enterprise communications efforts for the Direct Markets business and the Operations and Technology function. Before joining Zurich, Hicks was the first dedicated internal communications manager at Rivian's electric vehicle manufacturing plant in Normal, Ill. Earlier he served as public affairs communications specialist at State Farm, supporting claims executives and leaders.

(PLEASE NOTE: This is an edited podcast transcript, capturing speakers with natural speech patterns that may include incomplete sentences and/or asides, grammatical errors, verbal shorthand and some statements that may be less clear in print.)

EPISODE TRANSCRIPT:

MADHU RAMAMURTHY:

As AI becomes smarter, the challenge, in my opinion, is not to beat it, but to design a world where human capabilities are amplified, not replaced. There are so many places that only humans can go, where AI cannot ever go. AI will always outpace humans in pattern recognition, data analysis, speed—blah, blah, blah. But humans bring in judgment, empathy, ethics, moral reasoning. It's tough for AI to learn that. So, AI will dominate where logic rules; humans will dominate where meaning really matters. So, in my opinion, it will be a watershed moment for those who invest early, as they will be the ones who won't just adapt to the future but will shape it. See, AI can never replace humans, but humans with AI will replace humans without AI.

JUSTIN HICKS:

Welcome to Future of Risk presented by Zurich North America. We explore the changing risk and resilience landscape and share insights on the challenges that businesses face to help you meet tomorrow prepared.

HICKS:

It's hard to predict where AI is going because of the speed at which things are changing, but today we'll dive into emerging trends and we'll take a peek at what's around the corner in the world of artificial intelligence. I'm your host, Justin Hicks and today I am speaking with Amy Nelsen, Head of Underwriting Operations for U.S. Middle Market, and Madhu Ramamurthy, Chief Information Officer – App Delivery, both at Zurich North America. Amy and Madhu, welcome to the podcast.

AMY NELSEN:

Thanks for having us.

RAMAMURTHY:

Hello. Thanks for having us.

Explainable AI: The next big trend

HICKS:

Okay. So, for our first question today—you know, speed and efficiency sound like they're so last summer, —with AI. I actually want to start with a little bit of a game, and it's going to be a fill-in-the-blank. Madhu, I'll start with you. All right: the buzzword that will take over AI in the next one to two years will be?

RAMAMURTHY:

Actually, I can think of multiple buzzwords, <laugh>.

HICKS:

What's the best one? What's the best one that you have?

RAMAMURTHY:

I'm thinking of generative AI, agentic AI, explainable AI. I would say it is the explainable AI—or responsible AI—that's the buzzword I can also, if you let me add, I can say generative AI. Hyper-personalized customer service is another thing that I think of. And then, real-time assessment and pricing.

Things are evolving at a lightning speed in the AI space, so I wouldn't say speed and efficiency are so last summer yet. Though, I would say that efficiency is already becoming table stakes nowadays.

Zurich is already having tools to automate routine tasks, summarize documents, and support customer service chatbots, personal assistance. Our own generative AI used in Zuri Chat are some best examples. But things are quickly shifting beyond just efficiency, Justin.

Take agent AI, for example. This is an advanced AI that can plan, make decisions, adapt to new information, pursue goals with a fair degree of autonomy—which is so different from what we are generally used to with the traditional bots that we have had in our landscape for decades.

While the traditional bots are just software programs designed to follow scripted instructions to automate simple tasks, the agentic AI can reason, learn from experience, break down complex tasks, and find creative solutions—almost like having a digital assistant that thinks for itself.

Within Zurich, we already are running multiple pilot programs in underwriting and claims areas, of course with the very clear human oversight and governance. And then, the other two things that I mentioned: one is on the real-time, data-driven risk, and then hyper-personalized customer experience. So those are, again, our big buzzwords nowadays. So, the data-driven risk assessment—real-time—that's like, for example, take a trucking fleet segment of our insurance business. In the commercial motor insurance, trucking fleets are generally considered as high risk and are always priced high.

But an agentic AI can look at the real-time data of a particular trucking customer who may have telematics devices and driver safety programs, and the algorithm will come and recommend preferential terms and pricing for the particular customer. So, it's a nuance that agentic AI can bring, which we may not have otherwise. So, it may not be a one-size-fits-all. A human underwriter still makes the decision, but the agent offers nuanced insights so that we can price the risk more accurately. And then, on the hyper-personalized customer experience, I feel like AI won't just answer questions—it'll anticipate and proactively advise customers’ needs before even the customers know that they have those needs.

Like, I can remember Henry Ford, who apparently said, "If I had gone and asked what my customers wanted, they would’ve asked for faster horses." But he instead went and built a car. That’s what I see with AI.

HICKS:

So, the game was fill in the blank. I was looking for a one-word answer. Madhu, you gave me many words. <Laugh> Many, many, many words. That's okay. No, I'm just needling you a little bit, man. It's all good. That was a very detailed answer.

Amy, I'm going to ask you the same question, and you can choose to answer like Madhu did, or you can take a different approach. The buzzword that will take over AI in the next one to two years will be.

NELSEN:

So, I agree on the agentic side, and I think that's something that’s being talked about a lot. I think when we think about one to two years, which is a bit farther in the horizon, the thing that's a little bit difficult is the acceleration that we've seen even over the last year, I think, is a bit greater than any of us expected. And so, you know, just thinking forward one to two years, it's a bit like—you know—we might not be talking about agentic in a year or two. We could be exploring a whole different universe.

But what I'm hearing, both from industry peers and others even within our organization, is a lot of discussion about scaling, right? And how are you identifying those solutions that are going to scale within your organization?

For me, obviously specifically focused on underwriting, and scaling obviously leads to your ROI and return on investment. And I think back to the early days of the internet and when everybody wanted to build a website and e-commerce, you know, was hot, and things are sort of that new now. But I think what we've learned from past mistakes is to really make sure we do have an eye on how things will scale and be adopted and ultimately return the value to our investment.

And so, I think it's not as sort of sexy as a buzzword of agentic or some of the other evolving technology speak, but really, at its core—for us and probably other large carriers or other large organizations—scaling to get that ROI is going to be, I think, continued to be talked about in the industry.

Why regulatory clarity is crucial for AI success

HICKS:

I wouldn't disagree with that. I mean, if businesses aren't able to see the ROI on it, then they're probably going to be less likely to invest in it—which I think we can safely assume might be a bit of a misstep—but I can understand where they may be coming from in that regard.

So, as we continue to put on our prognostication hats, Amy, I'm going to come back to you for one more fill-in-the-blank: In a word, the biggest challenge over the next one to two years will be?

NELSEN:

Yeah, and I would say maybe rephrase as "could be." The regulatory environment is still evolving, right? And different in the U.S. than in Europe, to a degree. And with global companies really having to kind of navigate. At Zurich, we have an AI governance framework that is really trying to keep their finger on the pulse of what is evolving in the regulatory space. And we're trying to really be on our front foot in that regard. So, regulatory is one word—but just to expand on that a little bit: I've--

HICKS:

Sure.

NELSEN:

-- had opportunities recently to hear from a few of our state regulators at a conference I was at, one specifically from New York and the other in New Hampshire, Department of Insurance. And what I heard from them is pretty positive, in that they want to support us—you know, the carriers—in innovation. And they are really thinking about the balance of holding true to the consumer protection and all of the things that they really excel at, while making sure that at the same time you're supportive of innovation.

Which, you know, it's a bit of a balance there too, especially with AI changing as rapidly as it is.

HICKS:

Madhu, I'm going to come back to you with the same question. The biggest challenge over the next one to two years will be?

RAMAMURTHY:

I totally agree with what Amy mentioned from a regulatory standpoint, but I'll add a slight nuance to it. It's the ambiguity. So, in my opinion, ambiguity is the one that kills innovation and AI and not rules themselves. So, we should work closely with regulators to bring more clarity on AI usage.

What we have done is—from a regulation and governance standpoint—we took the framework that was developed with humans in mind and human-developed processes in mind, and we awkwardly stretch that framework to work for AI. So, that's what is causing that ambiguity or lack of clarity. So, it's important that we bring in the clarity so that we are not only aligned and compliant from a relations standpoint—that clarity is required even for our customers and stakeholders.

So, we need to be in a position to explain with clarity on where we are using AI, how we are using AI, what decisions are AI making, and also to our own employees and underwriters and claims professionals on how are they going to use AI. And to emphasize that there will still be human oversight, which is absolutely important. So, there should not be any ambiguity around when the human oversight comes in and when the AI comes in.

So, in short, ambiguity is what is going to be the biggest challenge, and it's important that we provide more clarity around that.

Moving from proof of concept to operational AI solutions

HICKS:

The human oversight piece, I feel like, has been one of the consistent themes that we've experienced throughout this podcast miniseries. So, I'm glad you pointed that out. And I love the phrase ambiguity kills innovation—I feel like that's really well stated. All this stuff, though, to me—Madhu and Amy—sounds very kind of pie in the sky still, in a lot of respects. And I know that a lot of times we can easily be stuck in this pilot phase of AI innovation as we try to bring this on to our various businesses and industries.

We had Barry Perkins join our most recent podcast within this miniseries, and he referenced Death by a Thousand Pilots as a phrase that I thought was kind of clever. You know, we don't know how to figure out or discern which pilots are the ones that we need to take and which course is what we need to take, and therefore we can kind of find ourselves getting into trouble.

NELSEN:

Yeah, and you know, like I said before, it's definitely a real thing. I have opportunities to speak with folks at other large carriers, and as soon as they hear me start to talk about broad-scale implementations that we've had—I'll touch on Sixfold a little bit—their ears sort of perk up, and everybody's really, like, sort of hungry for insight into those use cases that were super successful.

So, we started our first really foray in the U.S. Middle Market space into Gen AI specifically, was with a vendor called Sixfold. And when you think about—for those of us really familiar with a Middle Market P&C underwriter and what they're tasked with—the volume of information that they have to absorb, digest, and read in order to get an understanding of the risk that we receive from a broker, and that they're working towards potentially trying to get a quote out, they've got to understand a customer from multiple points of view, whether it's looking at their property or workers' compensation or auto or general liability.

In addition to that, they're also thinking about: what industry is this insured in? Are they a manufacturer? Are they real estate or retail hospitality? There's just a whole, humongous data set that folks on the underwriting side are really having to contend with. And so, part of the work that we ask our underwriters to do is document their risk. You know, like what are the different aspects of the risk that are important to us? And to be able to kind of do that documentation, you obviously have to understand the account in its entirety and be able to sort of redescribe that.

And so, with Sixfold, it's kind of taking all of the information in that we received from the broker. And in addition to that, anything—you know, a customer has a website, a summary about what it says the customer does on its website. And so, we've basically taken all of that information and provided the underwriters with what I would call, like, a Reader's Digest version of all of that information. And that's been huge.

And it's actually a piece of work that we also looked to transition to other resources to give underwriters time in the past, and that wasn't super successful. So, I would say our folks were really excited.

And Justin, to your point about what I would call, like, human-in-the-loop, is we're not necessarily asking or using Gen AI to do underwriting for us, right? We're actually using Gen AI to summarize a lot of information that's difficult to consume. And so, we've had really, really super positive feedback.

A little bit of a different flavor of AI that we've also launched across country is through Near Map. And that's more imaging AI, right? For those of us kind of paying attention to the space, we've heard a lot about maybe medical imaging helping diagnose different medical issues. Well, if you think about that in terms of property, Near Map has a super extensive library, and they actually own planes that they fly all around the country.

HICKS:

Is that all? They just own planes. <laugh>

NELSEN:

Yeah, yeah. And getting sort of imaging—aerial imaging—of properties, which helps us to help our brokers and our insureds understand what we're seeing from a risk perspective, whether it's equipment that's on a roof, or ponding, or things in a property that could potentially lend themselves to loss.

Really utilizing that in our underwriting flow in an automated way—where maybe we would've had to send risk engineers out to get that kind of initial assessment. So, in the imaging space, that's been amazingly positive as well. And then something that I would say is less, like, less of a vendor solution but more of, like, a desktop solution is looking at how we use things like Microsoft Copilot with Outlook. I mean, we're a big Microsoft shop here at Zurich, and diving into just even some of the more—what seems like—basic capabilities to help our teams really lean into AI.

I would say in those few cases, moving from pilot to production, it wasn't super difficult, but it takes definitely some patience and some really good project execution. But we've had lots of success. And then we've started on a bunch of new use cases that I'm really, really pumped about—that the future of underwriting is going to be amazing here at Zurich.

HICKS:

I can remember it was eight to ten years ago when drones were, like, the big thing for adjusters going out and looking at roofs and getting the aerial imagery and all that stuff. And it's amazing how far we've come just in a relatively short amount of time.

And that aside, the idea that we can just have condensed that much information down—like you were talking about with what Sixfold is doing—and making it easy for our underwriters to just be able to get the information they need and access it and read through it quickly.

I mean, I would love to have that at my job. I mean, I do a lot of reading for my work too, and if I can get that condensed, it's great. Let's save some time that way. If you could've told me that this technology would've existed back when I was young and coming up, I would've been a lawyer. You know, I don't have to go—I can go to law school and I can get all the condensed down into one—like, sign me up.

Madhu, what do you think, though, about this topic and how are you seeing some of these things—maybe once seemed theoretical—and what type of results we're getting, and then also what we're going to be able to see in the future?

RAMAMURTHY:

Sure. And then Justin, if you need to use Sixfold, you need to get your CPCU (Chartered Property Casualty Underwriter) first.

HICKS:

Uh, well, fair enough, fair enough.

RAMAMURTHY:

So yeah, I'm proud to say that we are partnering with Amy and the rest of the middle market team to implement several underwriting AI capabilities that Amy just mentioned. They're all highly impactful to the business, so I'm so glad that the team is using that. So beyond underwriting, there are multiple other AI initiatives that we are either in the POC state—proof of concept stage—or that are already operationalized at scale. I can think of multiple initiatives within claims that reduce data entry and analysis time for claims professionals and staff legal by using multiple AI technologies and also optical character recognition. So, there are initiatives like smart document intake, document intelligence for staff legal, claim compass, and then we have planned something for the AI foundation claim assistant.

So, there is multiple initiatives that are in the pipeline, both on the claim side and underwriting side within Zurich. In addition to that, we are also using AI to automate and bring more efficiencies, even within our IT department, to improve the overall quality. In our software development life cycle, we are using tools to write better requirements, automate our software coding, automate the overall testing, and then get more coverage, more quality deliverable. So that is bringing in a lot of efficiency, and that is like near-term. So, for 2026, we have already baked in efficiencies, so that’s coming. To summarize, we definitely are graduating from this POC purgatory or this whole pilot jail, I would call it, right?

To operationalizing this at scale with many of the capabilities in, I would say, relatively low-risk areas. As Amy mentioned, we are not having pricing and core underwriting, so we are more focusing on low-risk areas. But low-risk area doesn't mean low impact—so it's low risk, high impact. That's where we are focusing on. And we are very cautious about testing some of the POCs in a more controlled environment with really careful human oversight, especially when it comes to the area of risk assessment or underwriting or pricing. So, there are multiple irons in the fire—some are already operationalized at scale, and some we are still testing.

Reducing time to market with AI-powered solutions

HICKS:

Yeah, yeah. No, I love what you said. You know, everything that is digital is kind of becoming AI-enabled, and then we can communicate with those things using natural language conversations. It feels like that's a direction that we're headed. I want to ask—we've had some of our other guests in this miniseries talk about speed, and you mentioned efficiency just a moment ago. I think it's—maybe it shouldn't all be about speed—but speed is obviously kind of at the center of a lot of our conversations when it comes to AI and what it can do for business. What's one way you're seeing acceleration from AI, and what do you think this means for the future, Madhu?

RAMAMURTHY:

Great question. So, from my vantage point, I see that that'll be a drastic reduction in the time taken to launch new products. End-to-end lifecycle of a product development will tremendously shrink, leading to better speed-to-market. Not just the product work—it's also IT work. Everything. The timeline will shrink. In fact, at some point in the near future, that will be real-time product development. The products will evolve and adjust real-time based on the customer's needs, and advice will be hyper-personalized based on the context, behavioral, and real-time data. And also, much of AI's impact will also be behind the scenes—automating the middle and back-office processes, reducing manual reviews, reconciling data. It won't be flashy, but it's where billions are tied up in friction today. So, I see that both from a customer-facing standpoint, product offering standpoint, as well as behind the scenes in the mid-office, back-office areas.

AI-powered efficiency for brokers, agents and customers

HICKS:

You keep coming back to this point of personalized and having a personalized experience. And I think that that's remarkable, frankly. And I've even heard that we're going to get to the point we have like a personalized internet experience, or we'll have chat-to-work capabilities on demand in the workplace and things like that. And I don't know how far away we are from that, but I find all that to be fascinating. Would you agree, Amy?

NELSEN:

Yeah, I do. And customer experience is really more—you may be on the claims side first in commercial. There are tons of applicability when you dive into like the small business, almost like direct or independent agent-to-consumer, specifically and personalized policy writing experience. But you know, I think about acceleration in addition to all of the things that Madhu shared. What I've even personally seen over the past year is the speed-to-market of building a model is exponentially accelerating, right? So, where in the past it could take quite a bit to build an LLM (Large Language Model) and construct your ground truth, and you need really specialized skill sets, we're seeing LLMs and some of the things that we're testing built in like days and trained in days.

And so, speed in terms of the technology evolving is certainly something that we're absolutely seeing. But you know, truthfully, I wouldn't say it's necessarily about speed per se, but a bit more maybe in terms of when we think about efficiencies, right? And sometimes you sort of have to take a step back and look at, in underwriting across your value chain, what are the things that we have—whether they're underwriters or UAs—having to do that are low-value activities, so that we can really help pivot and drive those folks towards what we would consider more high-value activity. Like really assessing the risk, really helping our clients, customers, brokers get to the best deal possible. That's going to be a win-win—a win for Zurich, a win for the customer. It's definitely about taking away the things that are consuming time that would be better spent on other activities. So, there's still going to be a little bit of the kind of like speed component from a business perspective. But I would say, when I think about the speed, I think really about the advancements in the technology. The other thing—I think Madhu touched on this a little bit—is conceptually we think about our traditional, like predictive analytics space, where you're doing portfolio analytics and you're somewhat limited in terms of the insights that you know exist. You know, thinking about AI as a potential to identify and help us understand things and trends in our portfolio that maybe we're not even really thinking about.

That's super powerful and maybe also, to a degree, speaks a little bit to Madhu's point around product development. I haven't seen a ton of advancement yet in that space, because again, I think a lot of us are really focused on what is our process and what are parts of the process where we can interject—whether it's Gen AI or other imaging or other types of AI—to help streamline. Those are sort of the easier and less complex areas. But I can only imagine, based on how these models are really getting a) super-fast, but b) able to answer more complexity in questions, that a year from now, we're probably going to be having a different conversation about where we are.

Boosting time-to-market with AI-powered language models

HICKS:

I wanted to step back briefly, Amy, and because you mentioned the speed to market and how we take these models and accelerate them. You said LLMs, which is large language models. Can you explain that a little bit, and maybe provide an example for someone that may be new to this and may not understand what that means exactly?

NELSEN:

Yeah, I have actually, like, a really recent example of something that we've been testing, and not that I want to share true secrets, but I don't think this one is super complex. As most large carriers do, we, in the Middle Market P&C space, get loss runs as part of a new business submission. And those loss runs come from maybe there's a hundred sort of different carriers that give us the loss experience for the client that we're looking to provide a quote for, and part of what we do internally is get the description of what the loss was, and then we associate that with what we call a coverage type. So, just taking the simplest example: is it auto? Was it physical damage collision? Really, to help us understand kind of categorically where some of the loss experience is coming from. And it's actually a manual exercise. And so, we were thinking about it, and we said, well, we have all the text-based descriptions. We've got tens of thousands of rows of data. Let's look at these descriptions and try to build a large language model that says, based on a description, this is the type of loss that we would categorize it in. And so again, that LLM—couple days to build, right—because simply not a tremendous amount of data points, really very focused use case. But what we were able to identify there—it's not actually a success story, at least not yet. So, we had the model built, and then we ran through a bunch of test cases.

And what we were seeing was somewhere in the range of 60% accuracy. And so, we were able to get to that point in a couple of weeks. I could probably spend the next 25 minutes talking about, like, so what do you do when you kind of get there? Because there's a lot to be said for what does your underlying data look like. And then, is it a nuanced decision? And things like that, which is kind of the next step in investigation. But we were able to get there, like, incredibly fast.

HICKS:

Yeah.

NELSEN:

And now, able to understand, like, okay, what else about this paradigm that we're trying to solve is not maybe obvious to what we had originally thought. So, given how quickly we're able to now get to these answers, it really opens the door for a lot of exploratory work and a lot more things that we're able to introduce in the space.

Preparing for a future shaped by human-AI collaboration

HICKS:

Gotcha. Gotcha. And so, what we're seeing now is AI getting smarter and smarter, learning by the day, the hour, the minute, the second. And Madhu, I'm thinking about human capability being able to either outshine AI or at least use AI to augment their own performance in the years ahead. What do you think about that?

RAMAMURTHY:

Great question, right? As AI becomes smarter, the challenge, in my opinion, is not to beat it, but to design a world where human capabilities are amplified, not replaced. There are so many places that only humans can go where AI cannot ever go. AI will always outpace humans in pattern recognition, data analysis, speed – blah, blah, blah. But humans bring in judgment, empathy, ethics, moral reasoning. It's tough for AI to learn that. So, AI will dominate where logic rules; humans will dominate where meaning really matters. So, in my opinion, it'll be a watershed moment for those who invest early, as they will be the ones who won't just adapt to the future but will shape it. See, AI can never replace humans, but humans with AI will replace humans without AI. So, adaptation is always key. AI is a tool. Humans decide how to use it, right? Skynet is coming. Skynet is coming, but we are well prepared to handle it.

HICKS:

And what is Skynet for those who may not be aware?

RAMAMURTHY:

<laugh>, it's—oh, sorry. I grew up watching Terminator and how the Skynet, the AI world, takes over the human world and the whole war between humans and AI. So, sorry about the terminology.

HICKS:

<laugh>. So, we're not saying that war is coming, we're not saying that exactly.

RAMAMURTHY:

<laugh> No, I was saying that we are prepared because it's not a war. It's not, it doesn't need to sound as bad as Skynet, but AI is here and now, right? So, we need to adapt, we need to prepare.

Why human connection remains vital in an AI-enabled world

HICKS:

And it really comes back to, I think, relationships, right? I mean, that's kind of what I know that our business in commercial insurance, and just in the insurance business in general—I mean, it's a relationship business, and people like to do business with people that they know and like and trust. Amy, what would you say then is how you see human capability working with AI into the future?

NELSEN:

I just absolutely recognize that what we have heard when we introduced Sixfold is from underwriters—it's a real concern: Is this going to ultimately replace my job? And the job of an underwriter? I think about it in kind of maybe a couple different parts, but first and foremost, it's really a relationship business. You know, our underwriters are out in the market meeting with brokers multiple times a week, right? And so, what we're really trying to think about is, like, that's a big part of the job. Assessing risk and understanding and being able to meet our customer's needs as well is something that we want them to spend more time on—less time researching a company on the internet, less time paging through 20 to 30 pages of a submission—and really giving them the opportunity to spend time on the activities that are actually a bit more fun, right? In terms of their job. And so, what we like to say within my team and what we aim to do is just really take the rocks out of the backpacks and let them focus on crafting the best deals. You know, being out in the market, being visible. We don't really view a future, at least in the Middle Market P&C space, where an underwriting role becomes obsolete. We really look at this as an opportunity to drive increased satisfaction in the ecosystem that our folks are having to operate in. And so, it's absolutely a real concern and I think we just need to keep reiterating for us, what the strategy is. And it certainly is not in our sites that AI ever really replaces humans. It's meant to, as Madhu said, augment your ability to do your job in a bit less painful way.

Breaking out of pilot mode: Practical AI adoption strategies

HICKS:

I love how you mentioned visibility in the marketplace and that AI will enable us to be more visible, because if it's truly about relationships—which I think we can all agree that it is—those relationships are only going to get enhanced through visibility. The visibility is enhanced by the technology, right? I want to conclude our discussion with this: the advice that you would give people who are maybe struggling to get out of pilot mode—maybe you're stuck in pilot mode and you can't figure out how to break through or how to find the right AI tool that's going to supercharge your business or the way that you work. What advice would you provide them? We'll start with Amy.

NELSEN:

First, I would say the earlier you start, the better. I think I was reading a study that McKinsey had published that just shows the exponential value that early adopters are achieving and realizing in this space, because you're going to learn—you learn a bit in pilot mode—but you're going to continue to learn as you progress to a broader implementation in your organization. And the way, especially with Gen AI, there's this concept of what we call a ground truth. And the more reps, the more at-bats you have with your large language model, the bigger ground truth you build. And so again, the earlier you start and the earlier you get to more broad adoption, the faster you're going to get to build that trust and build that quality.

But if I'm going to give somebody advice on where would we start? It's really about identifying a use case or a business scenario that's low risk, right? So, there's a lot of apprehension. If you have a model that has hallucinations and it's not a hundred percent accurate out of the gate, you can't trust that model to be making a decision that an underwriter would make. And so, the use cases that we've identified are low risk in that they're providing value, but the underwriter or the UA—they're still part of the process, right? There's still oversight into the output, and it's not automatically making any decisions, right? It's really about selecting those use cases where the adoption and the road to adoption is going to be less challenging—is immensely important.

If you pick a higher-risk use case, you're always going to get the feedback from folks of, well, it's not a hundred percent accurate. So, if I can't trust it—like, it's cool, but I'm still going to do what I've always done. Again, and we talked earlier about really leaning in to realize the value. It's starting with the use cases that are low risk, that have human oversight, that build your ground truth, build your accuracy, and then you have a model that's actually performing at the same potential level of accuracy as a human. And you can use that then for other parts of your business and lean into the more high-risk use cases. But you have to start with lower risk.

HICKS:

And as Madhu pointed out earlier, lower risk doesn't mean low impact, right? So, they can still be high-impact use cases, but that's still minimized risk. Madhu, how would you respond to that? If I'm looking to get out of pilot mode—if I'm stuck on pilot autopilot, trying to get out of that <laugh>—how would you advise me?

RAMAMURTHY:

So, yeah, first, to understand why we are in this pilot jail, we need to understand the reasons why we are in that, right? So, I would say there are three main reasons for that. First, legacy systems. You cannot ask a flip phone to run TikTok. So, try and sidestep and try and remove dependency with legacy systems when you're building AI capabilities, because legacy systems are—it's like we are building a future with the organizational antibodies that are designed to kill it. So, legacy systems not only resist AI, they murder AI. So, it is important that we sidestep and then either use modern platforms to build or find ways to avoid the dependency on legacy systems.

And then the second reason I would say is regulatory complexity. We already spoke about it. So, getting more clarity around AI usage is extremely important.

And then the third one, last but not least, is leadership buy-in. That's the key. There are several leaders who still are very skeptical and are in this perpetual wait for perfection, to Amy's point, right? It is not going to be the models that are going to be perfect to start with. It is important to realize that in the AI world, velocity compounds and perfection stagnates. Of course, speed cannot be at the expense of our responsibility to customers, regulators, and society.

So, to again emphasize Amy's point: take simple but impactful use cases, define the desired business outcome upfront that is clearly measurable. Engage your compliance team upfront—they should not be an afterthought. And then find the solution that is scalable and within the governance guardrails. Test them in a controlled environment, quickly learn, quickly fail, learn from that, and move on.

So, it's important that we ensure the AI solutions are not just advanced, but also trustworthy and safe, and that they are explainable to our customers and stakeholders.

HICKS:

Man, you said a lot there. When I say that, that's a compliment. That was an excellent answer—legacy systems, the regulatory pieces, and getting leadership buy-in. Jewels, man, you're dropping jewels on us right now and not letting perfection be the enemy of good. Let's keep this thing progressing and let's keep moving forward. I want to thank you guys for joining us on the podcast today. Madhu Ramamurthy, Amy Nelsen, thank you so much for joining us.

It was a fantastic conversation. We're recording this in October. And a special salute to you, Amy, for October is National Women in AI Month. So, thank you for helping push Zurich forward—you didn't even know that did you?

NELSEN:

I did not, I did not.

HICKS:

I appreciate what you're doing for pushing Zurich forward in the world of AI, so thank you for that as well. And thank you guys both for joining the podcast today.

NELSEN:

Awesome. Thanks for having me.

RAMAMURTHY:

Thank you very much for the opportunity.

HICKS:

Absolutely. And thank you all for listening to our deep dive miniseries on artificial intelligence. If you liked the show, please leave a comment or a review wherever you get your favorite podcast or drop us a note at media@zurichna.com. This has been Future of Risk presented by Zurich North America.