The Impact of AI on the Future of Insurance | Episode 4
Throughout today’s insurance ecosystem, AI is the topic of the moment. How it will affect digital transformation, disrupt solution workflows, and potentially supplant human labor are just some of the more urgent questions that carriers, agents, and advisors are asking. Listen as host Jamie Yoder, President and General Manager of Sapiens North America, discusses the impact of AI on the future of insurance with Anand Rao, Distinguished Professor of Practice in Applied Data Science and AI at Carnegie Mellon University.
Jamie Yoder: Hello and welcome to the Sapiens Insurance 360 podcast. I’m your host, Jamie Yoder, Sapiens’ North America President and general manager. So glad you’re tuned in today. This is where we discuss the latest trends, news, and issues from across the insurance landscape, our solutions and technology. And today we’re going to be discussing, I think the hottest topic out there, the impact of artificial intelligence on the future of insurance. And I’m particularly thrilled to have an exceptional guest joining us today for this discussion. Dr. Anand Rao is currently the distinguished professor of Applied Data Science and Artificial Intelligence at the Carnegie Mellon, and most recently was the global AI lead at PWC. Honored to say he is also a long-time colleague and my colleague, co-creator, and lead for the future of insurance research.
And just as a little highlight, you know, Anand, who has over 35 years of industry and consulting experience advising C-level execs on AI and not to go into all of it, but is recognized as one of the top 50 data and analytics professionals in the U.S. and Canada, as one of the top 50 professionals in insurtech or the top 25 tech leaders in consulting, and has won the number of awards for his academic and business papers.
So Anand, I mean clearly, you’ve earned the right to talk to everyone on what’s really happened with AI? And I think the most intriguing thing is you don’t just talk about it. You’ve, you’ve implemented this in many organizations and in many areas. And a big part of your remit at Carnegie Mellon now is really to teach the operationalizing [of] the AI in the institution. So welcome and it’s such a great privilege to have you here.
Anand Rao: Thank you very much for having me. And it’s been a pleasure joining you again after I did a six, seven years at PWC, joining you now. Looking forward to it.
Jamie Yoder: Yeah. Excellent. Well, let’s jump in. I think we got a lot of ground to cover. I think this is something we clearly could spend and have spent years on. And so we’re going to just hit the highlights. But, you know, a topic that’s clearly at the forefront of every industry, but particularly insurance. Such an information-intensive business is, AI and, you know, I think largely when I look more broadly, it’s really about how you unlock the potential of digital and what that really means to insurance organizations and how that’s fundamentally changing. You know, all facets of, you know, across the value chain. You know, to start off and talk a little bit about some of the five waves of digital and what they really mean. I’d love to get some of your take on what those changes really are and how we think about more broadly.
Anand Rao: Yeah, this is a great question, Jamie. As you know, I mean, we have been involved in the digital transformation for 20, 25 years at least, and digital transformation means different things to different people. And also, I think to be fair, it has also evolved, right? So what was digital in the late 90s and the killer apps that the dot-com era was this then the Internet, smartphones, and the digital has been changing as well. So more recently I’ve been thinking about this all-digital transformation, literally successive waves of digital. So I think the first wave to me is very much the digitization. So that’s the first wave where all kinds of processes, whether they are physical processes or decision processes in insurance, in every other industry sector has been digitized, so converted into zeros and ones so that we can start reasoning with them, start, start analyzing them. So that’s the first wave and this has been going on for quite some time. I think most organizations are still doing digitization, but I would say most of that is behind what they’re doing. So they’re still at the manufacturing level and so on. There’s probably still digitization happening with Iot, industrial Iot, and so on. Right. So with the digitization comes the era of big data. So lots and lots of data, volume of data, velocity of data, variety of data, all of those things that people talk about. So that’s the sort of the second wave.
Then we started seeing with more data, I think could take one off two parts. When you have lots and lots of data, what you start doing is start standardizing the processes, start simplifying them, and you start to standardize and simplify. You automatically go and automate them. And we saw a number of businesses doing that. So really impacting the bottom line. So cost savings, productivity improvements, that’s what the automation wave gave you. And just because you have lots of data doesn’t mean that you need to go to the lowest common denominator and simplify and automate. You can go the other direction, as we have seen a number of companies do, [to] start personalizing the experience, whether it is your agent experience, customer experience, claims experience. Any of those experience can be personalized. And then there is also a parallel process that happens in analytics, which Kelly, Kevin Kelly from Wired calls it “cognification.” So as you get that, so you start cognifying the domain expertise of people and that’s the analytics wave that we have seen improving experience, better revenue, better stickiness, and so on, obviously leading to the top line growth of revenues and margins.
Now both these waves that lead on to the fifth wave, which is basically the AI wave that’s been coming and as you said, it probably is a tsunami. So just to recap, the five waves are digitization, big data, automation, analytics and AI.
Jamie Yoder: Yeah, no, absolutely right. It’s something that, you and I spoke a lot about and written about in the past. It’s this whole notion and, you know, it’s almost the bionic organization, how these things collectively operate. And how does an organization adapt and evolve to pull all of those pieces together, the automation and the augmentation? So a lot of what’s happened with the AI around…and Gen AI really have been around…how it sort of partners with the individuals to get more and more out of, you know, get more out of the people.
Anand Rao: Yeah that’s right, yeah that. That’s why I think the analogy of waves makes a lot of sense. So just as one wave is receding, the other is coming and then the other one gets wrapped up with the earlier wave and becomes bigger. So there’s this sort of constant flow, ebb and flow of the wave.
Jamie Yoder: So how do you see these affecting broadly, you know, really the future of insurance?
Anand Rao: Again, as we have worked through, Jamie, over the past decade or more, tend to look at the impact of all of this digital on insurance and sort of at least two big buckets. One, the sort of more I would say, transformative changes, and the other one is more disruptive changes. So when you look at transformative changes, whether it is P and C or life and annuity or even reinsurance, you’re seeing every aspect of each of these sectors, the value chain change and AI analytics, all of these five waves that we talked about impacting them. So let’s just take some very specific examples. If you look at agent experience or an advisor experience, it belies an anybody side. So now you have ways in which you can quote unquote cognify, or make the systems, the AI systems, capture the domain expertise of people. So we have people in the industry, what, 20, 30 years [of] experience, advisors have been doing this for their entire lives and capturing that experience and making that better for some of the junior people coming into the, the workforce, the agent force, or the advisor folks.
So that’s just one way in which these technologies are, as you said earlier, augmenting what an agent could be doing with a better experience that also provides consistency. So again, regulators are happy because you’re, you’re essentially bringing a consistent level of service, well documented, and it gets the expertise of the people. And we have seen a number of early views around how Generative AI is essentially pulling up the average for all of the entry-level people.
Jamie Yoder: How do I actually now apply it? What do I do, right? So from an underwriter level or at an agent, you’ve talked in the past about, you know, this really being like an answer engine, right? So how does this really sort of help you? One of the things we did at our customer summit in Europe recently was [to] look at Gen AI all around you, just to add, to look at coverage gaps, and to what am I covered for, what am I not covered for, and then trigger that into some operational process. And same thing on the client side, summarize claims files to really pull out all those insights across all that unstructured data.
Anand Rao: So exactly. Customer experience, exact experience, I mean, everything sort of changes.
Jamie Yoder: Yeah, and you mentioned a second piece to that, the second bucket, and that’s to disrupt it. I’d love to hear more [about that].
Anand Rao: Yeah. On the disruptive side, again, let’s just take the P and C side, so I know at least a few companies that promised that there’ll be autonomous vehicles on the road by 2022, last time I checked it’s 2023 now, and still not truly fully autonomous vehicles. But I think there’s still got a long way in terms of autopilot technology, driver-assisted vehicles, and so on.
So what I think is happening and it’ll be a much more of a slow adoption, is this whole notion of autonomous vehicles or driver-assisted vehicles. And these things I think will start changing the way we look at insurance. And I know we have speculated in the past about how the auto insurance might become, not that you don’t need auto insurance, it’s just that the auto insurance might become more of a business insurance as opposed to a consumer insurance if more and more of the vehicles are just as they have, 50,000 mile warranty or 100,000 mile warranty, they might start including no accidents for the 100,000 miles, 200,000 miles. So they’ll be covering the accidents as the number of incidents fall. There is some truth to that and some studies that are being shown that the number of accidents reduce. So the frequency of those accidents reduce. So we could see a future where more and more of the risk gets taken out and that the risk gets packaged by the manufacturers. So you might have the large auto manufacturers doing deals with the large insurers and insuring all of the autonomous, semi-autonomous, vehicles. So that changes, if you like, a way in which insurance would work on the auto side. Again, we are seeing the same thing on the trucking side or on the commercial auto side. I think it’s still early days, but that could be very disruptive, the way auto insurance works.
Another example, I know we have worked on it and [it’s] sort of still early days, I would say is on the commercial side, especially on the industrial. You always look, is that a catastrophic event? How do we prevent that event or how do we ensure for that event, and how do we cover the losses? And I think we are moving from that to much more of a preventative or a risk mitigation approach.
So with the IoT, the CLIoT, and all the sensors coming in, you can start analyzing and doing predictive maintenance and then preventing or essentially extending the life of these systems. So essentially the risk managers, the insurer becomes a risk manager and a risk mitigator, as opposed to an underwriter. So that also would be a big impact.
Jamie Yoder: And I think what’s interesting, it’s like, you know, it’s like most innovations, it’s not that the models didn’t exist, the power and what’s the technology that’s possible allows you to amplify and alter the possibilities of it, right? So, you know, risk management, risk prevention has always been a fundamental discipline and rigor of, commercial carriers, insurers. Broadly, it’s the opportunity to do that in a much broader aperture, in a much deeper way, I think is where it really becomes powerful.
So it’s just as you said, I mean, I think, you know, we’ve talked to this before. If I think of, you know, the simple equation of the customers outcome, equals that those things they can control for, the deterministic plus the things that are out of their control. The probabilistic. And IoT, and all that codification in what you talked about, with the cognification, the ability to put more in the control, be more deterministic, allows you to, in probabilistic, is what insurance is covered for. It’s kind of like, I want to control your productivity, your outcome backstopped by capital for the role that insurance has always played. But in looking [at] that in tandem to increase the effectiveness of the customer in that total outcome. And I think that’s a real powerful, you know, essentially [a] formula you can think about across all the, all the different lines of business.
Anand Rao: And on this point, Jamie, so one of the, the more recent things over the past six months, what’s happened to your point around the technology has existed, but now I think it is being democratized, especially on the AI side is being democratized more. So again, with the Generative AI, large language models, it’s sort of starting everything from scratch. What the community or the AI community is now saying is, everyone doesn’t need to be a deep learning expert. We have done the work and it’s available now. You can build on top of it, right? So that level of modularity and democratization, I think, is really going to help insurers. You might be a claims expert or underwriting expert, but with some help from the idea that besides, you can build much more sophisticated systems than in the past.
So that’s, I think is a major change over the past six months, is this whole democratization process.
Jamie Yoder: And the pace of that change or this new thing, new, new models, new variations, new things are coming out. But, you know, along that line, getting to the heart of it. So that’s a couple of the sort of the uses, [where] you know, AI, where it fits. But even going more deeply, if you will, how do you see other sort of aspects of the full role of AI, both traditional and Gen AI, in that context?
Anand Rao: Yeah. So what’s happening over the past, I think a couple of years, it’s not really just number one, but these were released, but it’s been going on for at least five, six years. The whole notion of deep learning transformer models and the open data that has been available, all of that has essentially been gaining more momentum. So to answer your question, I would address Generative AI first. Now what we are seeing is, I mean, everyone is familiar with an ID stack that has all evolved over the past couple of decades with the storage and compute and the data layer on top of the application layer and then the channels and so on. So the ID stack has evolved. Now, the more recent evolution of that, I would say, it’s the convergence of software applications with the data and AI models, machine learning models. So that’s what’s sort of happening now, that all three of them are coming together. And now we are seeing literally a three-layered model architecture being sandwiched between your data and your application systems and what that is essentially doing at the base layer are what some people call as foundation models or large language models. So these are models that some of the big tech companies are already creating with externally available public level data. I’m seeing a number of those. So some of them are open, some of them are obviously tied to the technology platform. So that is the base layer, if you like, a foundation layer. Now I think there are already quite a lot of investment, as you said, is coming into the next layer, which is very sector specific.
So this is where I think for insurers, for example, on the legal world, people are essentially taking these large language models and building legal large language models sitting on top of it. And when you start having legal, then you can start for the claims and insurance. You can start building a claims model on top of it. You can do it by country, you can do it by corporate, and so on, right? So you start building these large language models that can then be used by many insurers. So there are still companies that are investing in building in these claims models, underwriting models. The same thing in the health care debate. So that’s a sector-specific model. Then you come to the third layer, which is very much a company-specific player, right? So at Sapiens, you want to use the claims model specifically with respect to your claims policies, your customers, and your products. And that’s the last layer that gets tweaked. So by the time you get the other two layers, it’s pretty much, I would say 70%, 80% done. And you have to spend that time on that last layer. And even there there are tools coming in so that, I think, is a very fundamental change. So most of it is going to be done by the tooling that you’re going to see coming through in the next couple of years and then companies will start focusing on their proprietary layer. I think that’s a big change from a Generative AI perspective, and I think that traditional AI is borrowing some of these as well. So now we are seeing the notion of knowledge graphs, of graph analysis. Insurers have used that for claims fraud, for example. All of that is now combining with the large language models, so that you have the explanations for the knowledge graphs and you have much better robustness and learning from the large language model. So you’re seeing the traditional AI also morph a little bit, and borrowing some of the old techniques that marry with the new techniques and those are some of the things that we are seeing, how to use large language models for reasoning, acting, just as we add rule-based underwriting, it’s just that you’ll have three of thought-based underwriting, but with the large language models, right?
So you see the hybrid version of these sort of generative and traditional architectures coming together as well. So really exciting phase over the next decade or so, we are going to see more and more of this kind of stack emerge.
Jamie Yoder: Yeah, it’s fascinating, right? I mean, we always talked about this transformation, and you’ve covered it, right? It’s, you know, transformation is about change, right? And so it’s changing the way you engage. You talked about examples with agents. It’s obviously customers, right? So this answer engine mindset, changing the way the work is done. How underwriters, how they tap into claims handling and both the automation and the augmentation as you can get more and more sophisticated and more and more applicable to it. And it’s also about changing the way the work is done. So change the way you change. You know, these things also apply to how we’re building the technology and how we’re deploying it. And the pace of that change is just, you know, staggering. I’ve always, always felt that it is accelerating. And now, that’s tough to keep up with. But it’s a wonderful opportunity. So now, you know, it’s so incredibly interesting, right? In our practical level, insurers are obviously looking at this and obviously, one of the things that they, but what you’ve done and what you’re teaching on is around operationalizing AI. To sort of get that full return on a lot of these investments in this activity. What do insurers look at? How do they actually go about this?
Anand Rao: I think yes, you know, Jamie, it’s sort of relatively new. The notion of building machine learning models and using it, I would say at least until three or four years back, maybe five years back, the modus operandi for a large insurers and for any company has been, hey, let’s build a model, let’s collect the data, let’s build a model, let’s test it out and see whether it’s useful or not. So it’s very much, it’s like an artisan, right? So it’s much more of a craft rather than a really industrial process. Now, what people noticed is that’s all great, but you’re really not getting the adoption of some of these tools beyond the half a dozen, dozen people that are using the model. So you really need to scale this so that literally all your agents are using this, all customers are using your chat bot, right? So there is a level of scaling where you are going from a few tens of people to thousands, hundreds, hundreds and thousands, and even millions of people. So that’s where I think this operationalizing AI has essentially come to the fore, just as the software industry evolved to become more of an engineering discipline with software engineering, now people are talking about AI, evolving from a craftsman kind of a cottage industry to really a factory model. And that’s where the the engineering or AI engineering is coming into the fore, where we really need to look for practices, disciplines, tools, techniques that can do the scaling. And that’s where just as software engineering evolves, so we are looking at AI engineering evolving by combining some of the engineering practices so that dev operations or development and operations together, continuous integration, continuous delivery of software. So early days, in the nineties, you had to shut down the system. Do a patch, do an update, and then release it. So the system will be good and go down for maybe an hour, a couple of hours. And if things don’t go well, it may be down for a day. Now none of those things happened. So people are continuously updating their software and they’re doing that for the software.
So now we are talking about doing it for the models. So the models are also going to be not only continuously on, but continuously learning. And that’s where the power of the system comes in. So just as you are doing the underwriting or the customer experience and we all have that experience on the retail side, when you go into a retail electronics side so that the website gets updated, the recommendations are very much fresh based on what people are searching, what you have searched. So a similar kind of experience [is] coming into insurance and other areas. So that’s what operationalizing AI is, essentially automating everything from the start to not just building the model, but deploying it at hundreds and thousands of levels, monitoring the model, making sure that the model is performing within certain criteria, and even retiring the model. So that entire model lifecycle or AI lifecycle is what this engineering is addressing.
So the people part of it, the process part of it, and of course the technology part of it, all coming together. So that’s where this operationalizing AI is coming in. I know many of the leading insurers are already doing some of these things and deploying these models. Now a few of them have got hundreds of models deployed. And I think this is just going to continue, just as you have software, you’re going to have hundreds of models deployed doing very specific things.
Jamie Yoder: You know, it’s an excellent, that’s certainly a topic that deserves tons more time and in-depth discussion. And hopefully we, we can have a follow-up to go much more in-depth on that. But, you know, switching gears a little bit and sort of flip to the other side of it, so we’ve talked about all the great opportunities, right? But there’s probably just as much noise and buzz out there about the risks of AI and what that can mean. So what do you see as the potential risks of AI?
Anand Rao: Now it’s, it’s interesting in the AI world, it’s always been not only opportunities on one side that is there, there’s a whole group of people talking about risks. And of course as insurers we like risk. So as insurers, risk is not necessarily a bad word. It’s actually good for us, right? So if there is a risk in the market, we look at, how do I reduce the risk, how do I mitigate the risk? So from an insurance perspective, the risk is also an opportunity. But of course, for the rest of the world, it’s something to mitigate and avoid. And that risk actually comes in different forms. I would say there are some concerns and I mean some, some leading technologists and scientists have talked about the existential threat to humanity with AI. I think that’s a little bit further away. There’s just nothing with a few people are looking at some of those issues and how do we prevent them. But we are concerned much more about what is happening today, here and now, and I would break it down into sort of three specific categories of risks which are very application-specific risks, right? So you build a chat bot, you build an underwriting system. What are some of the things that can go wrong when you build that? What are those risks? Then there’s an enterprise-wide risk that that comes in. So as you start using more and more of the AI, there’s maybe a reliance, more on the automation as opposed to you taking independent decisions, individual decisions, being able to stand behind.
So there’s an enterprise-wide risk that you need to be cognizant of. And then, then of course, there is, you can’t deny that there are societal [risks]. When there is more automation, there are going to be job losses of certain kinds, right? So not everyone is going to be replaced, but there’s going to be more augmentation. But definitely, there’s going to be some job losses in certain categories. So we look at these risks in sort of all these three categories, and especially at the application level, there could be risks around bias. So you’re training a model, let’s say, with your data in Midwest, let’s say just a flood insurance in the Midwest. That’s great. You’re using a model. It’s been very effective in the Midwest region. Now you can’t take the same model and take it to Florida and say, hey, I’ve done that, I’ve done everything. I’ve got to model Midwest on flood insurance. I’m going to use the same thing in Florida. It may or may not work, and more likely, it will not work given the the situation in Florida or anywhere in the South. So you need to get new data. You still have the original model, but you need to get the new data, train, retrain, so those are some of the things where your model might be biased with respect to specific datasets that we just built, the history. And when you got the data, all of that. So I think you need to be much more cautious about bias. Explainability, transparency, safety. There’s a whole host of these kinds of risks that people talk about and that I think you need to be as a insurer. If you are building some of these things, you need to be cognizant that you’re making sure that you’re addressing all of those things.
So we look at process and the global research topic around governance. So people are used to data governance. Now you start looking at model governance. So if we should be involved, we should be making the decisions. At what point in time, when they scope the project to build it. But they deployed all of those things, right? So just the basic hygiene and that’s evolving as well, right? So as insurers adopt them, then I think you can mitigate many of these risks. You can’t avoid all of them, but you can mitigate. And the last thing I wanted to mention, as I said earlier, risk is also an attractive proposition. There are a number of companies that are really worried about AI, AI risk, and now there are a number of organizations that are coming up with ways of quantifying this risk, whether it is bias, badness, explainability, and so on.
So now there’s that emerging opportunity for, can we underwrite AI risk and how can we do that and how different or similar is it to, for example, cyber risks or deepfakes is one of the things that AI has brought in but is also related to cyber. So there is an overlap between cyber risk and AI risk and is AI risk essentially AI risk insurance becoming a separate line on its own? Potentially it could happen, right? So there are opportunities for insurance as well as I think about AI risk.
Jamie Yoder: Yeah. No, I love it. It’s a topic near and dear to my heart. So I mean I love all the different specialty lines and all the, all the disruption that’s happening in every industry is obviously, we should use this as a way to really attract talent to insurance. It’s every disruption, to think change happening out in other industries is relevant to insurers and really being at the fore of those innovations because that risk can be underwritten and you can do some really exciting things and a subsequent podcast is going to be on the specialty space. How do you build distinctiveness at speed and scale, how do you facilitate products for things like that? So I love that little notion and I think it’s a great thing to end on, a fun note to end on. But Anand, thank you so much. You know, as I said earlier, we could spend days, months, [we] have spent years discussing AI and working on solutions in that space.
And there’s only great things to come for the insurance industry. And beyond that, you let off with about the waves of change and sort of the five waves. And I think the interesting notion I would add to that is there’s a gigantic boulder [that] has just been dropped into the water. So those waves are moving faster and higher than ever. So but all the, as you said, folding in each one of them. So exciting times for us. So [I] really, really appreciate you spending some valuable time and your efforts at Carnegie Mellon. And I have an opportunity to continue to work with you as you, I think, build the great minds for the future. So thank you.
And thanks again to all our listeners for joining us today on the segment. We’ve got more coming, so please be sure to tune in next time on Sapiens insurance 360.