Using AI for Reinsurance | TRP #166

Hey there. Welcome to the Reinsurance Podcast, the place where we dive into all things reinsurance, the coolest part of insurance. We're your hosts Jared and Ben, a couple of ex-practitioners who love the industry so much we founded Super Seed to tackle some of its biggest headaches. And we're here to share our insights and stories with expert guests as we uncover what's really going on in the industry. Welcome back to the Reinsurance Podcast where we bring you another upbeat, not another episode on AI, but it's going to be a good one because >> we're two humans. >> Oh, that's true. But also because this one is about the thing we love most on the title of this show also >> podcast. No reinsurance. Damn, I knew I was I was going to get there. I >> got there at the end. But >> I'm excited because we are going to be talking about specifically AI in reinsurance and you know what was it for? Should you bother? Yeah. >> Do you actually need AI? Is it a threat to the very existence of the people who make this industry great? >> Yeah. All those topics and more >> and more >> will be coming. >> This is good. Um I'm I'm going to be doing a a session with an a young underwriters group in a couple in a few weeks >> um specifically on AI and reinsurance. >> This is the warm-up. Yeah. >> Or you're just going to not show up and give them this recording. >> Here's a podcast. Listen to it. Um no. So I won't I won't jump into everything that because that's like a a 40minute presentation that I'll I'll be giving. Um, >> you you you'll be prepared then, I'm guessing. >> Wow, I'm only starting to prepare. I wish I would have done I wish I would have finished my 40 minutes. >> Better. >> Yeah. >> Tune in next week. >> This is the other way around. This is like my preparation for the prep. Okay. Well, >> an episode brought to you by deeply under research then as usual, but hopefully some random opinions that you can take with you to the workplace and pretend off hack. Well, I think if we if we sort of zoom back out, we talked um we'll kind of give some context around like where I is where AI is really effective and where it sort of is has historically been less effective at times. Um and at this at the very highest level, it's worth remembering or reminding the audience that um AI is an exceptionally good guessing machine. Mhm. >> And when prompted with a request or a task, it will go about guessing every word in a sequence. >> Did you see by by the way, did you see the um paper that came out? I think it was I can't remember which university it was. One of the like top Ivy League universities. They did a study that proved that the reason why AIs hallucinate is because of the way that we've set up their I'm testing like scoring environments. >> They they worked out or the models basically found out that they achieved higher scores if they guessed. >> So the hallucination is them trying to achieve a higher score. Yeah. >> So so the reason they just make up stuff or guess stuff is because there's a reasonable probability that they get it right. So when you benchmark and test how well AI is doing, you get a better score by guessing. >> Y and there are um >> peer-reviewed research. >> Yeah, there are there are new like model companies emerging that are designing themselves to essentially try to rebuild the way models work with starting with like more context first and then working itself back to this element. But at the end of the day, there's always going to be some element of what it is doing is is predicting the next most likely set of of words and things. But you're right, like some of these problems stem from the way the data was originally presented, which is why it's very good at making it thinks it make it making it look like it knows a thing without understanding it. Um, so it still makes it really, really nice for things that don't have universally true elements to it. So things like writing, things like preparing a list of topics to discuss on a podcast, for example. >> The producers not listening to this. >> Um, but it's it's brilliant at that because it's you can look at that and go, "That one's not quite as relevant. I can skip it." Or you can look at, you know, note summaries and go, that's actually not how you would spell that. And if you know, but if you're sort of um transcribing an a podcast or something, it's close enough for the person to to get the value out of it. Um, so it's brilliant at those things, but when it gets more into like it needs to have like exact exact correct things, it it struggles a bit more. Um, so when I think about where it will fit in reinsurance, I think the starting point is understanding where its limitations are and what it's good at. Um, and businesses will look across what they're doing and build out use cases around various things. Um and that kind of leads me to the second major thing that I think will needs to happen is for an a reinsurance institution whether it's a carrier a seedant broker or a reinsurer um those firms will need to be like very explicit I don't know what they're trying to achieve and set those targets to be quite small like we want to have it where all of our notes do this and teams centralize that and we can get a summary of like something where it's like very much pointed at a specific problem and not just this like AI for everything everywhere all the time thing because that will result in more noise and more chaos and harder quantify quant to quantify value in these types of things. >> I've got another angle as well. Let's go for it. >> Have you seen Chitty Jitty Bang Bang? >> Oh, when I was like seven. There's a character in that called the child catcher and he like hands out. It's a bit bit dark really actually. So >> that sounds dark. >> Yeah, I know. Exactly. It's they come from a country called Vulgaria. >> Yeah. >> Where children aren't allowed. And that's apparently why he exists. >> Oh, >> less sinister. I don't know. It's still quite Anyway, what he does is to trick people is he hands out lollipops and ice creams and trickle tarts and things to trick people into going with him to, you know, like the pi piper kind of story. >> Yeah. >> Um, and that is a thing that is happening in AI at the moment in the business world as much as in the consumer world. >> Yeah. >> And I'll explain what I mean here because there's just lots of raised eyebrows at the moment, yours included. AI costs quite a lot of money. >> Yes. >> Never more so than in a world with escalating energy prices and limited energy supply. The thing that we're seeing at the moment on a massive scale, much like with Uber back in the early days of the race for like Uber versus Lyft, etc., etc., is this enormous amount of subsidizing of AI to the point where actually in China at the moment there's a a battle happening between Alibaba and Bite Dance in particular where they're actually paying people to use AI. >> So you're actually getting money like as a prooning like suite trying to persuade you god just we'll pay you to actually give it a go. >> Yeah. I because they know that if they can get you hooked then eventually in about 7 years time you won't mind as I didn't on trying to get home somewhere the other day waiting 12 minutes for an Uber I that never came and then canceled on me and then paying just as much as I would have paid for a regular taxi. Yeah, >> I still love Uber. Sorry. Even though despite that um but I think we're seeing a similar thing in in the world of AI where at the moment >> it's basically free, right? If you're a user who's not even paying for a subscription, you get a ton of the a ton of service. You get a pretty high limit in the amount of credits that you have in the amount you're able to use it. If you're a business, it's quite similar. You pay a very low rate and you can see all the wonderful benefits and things it does. And there seems to be no downside. But all of that downside is currently being subsidized by the markets effectively venture capital, the street, people who want to be on on hand for the upside as and when it comes. Which means that I think probably at the moment we're all being handed lollipops and ice cream and being told to go with, you know, these AI suppliers without really considering if we build them into our processes. If we replace and lay off and get rid of, you know, the humans who know how to do the work and replace them all with pure AI without doing the >> cost benefit analysis, we might find ourselves in a position in a few years time where suddenly all the AI that was free is now actually quite prohibitively expensive. >> Yep. >> And there's no way back. >> Yep. >> Anyway, that was a bit of a long doom and gloom story. Back to the more cheerful AI reinsurance. No, it's it's it's important context though because I think anytime you're putting products into organizations like this and and SAS has done this in the past and so is like the computer revolution and all these things in the past. um you need to do this like ROI calc for enterprise spend. And if that ROI calc is pure headcount like that might be helpful in some instances in some departments but it might not be in in others. Um, but that ROI calculation changes drastically if instead of paying a £100 per seat for an engineer, it's like a,000 pounds a seat a month per like well now an organization is looking at this going, well, we were happy to spend four, five, 10 million pounds a year on these various subscriptions and models, but now this is like550 million pounds. Like it begins to go like that. We will not make that investment though. There's no amount of ROI it could deliver. Like it's five times what our whole employee base would have cost us. And and so those those things there will be a breaking point somewhere. And the question is sort of within the AI companies around like can they make money at the at those breaking points and there's research being done there around whether this is feasible and viable and how might they they generate income to >> in other ways. It's interesting right as well if I were to contradict my earlier >> story of warning and why >> I you might say that it is rescuable if competition plays out because we do have at least we haven't gone down the scenario where it was just open AI which I think would have been much >> scarier you know in 2022 I think when >> Sam Alman sort of kicked all that off and said here's what the thing we've got we're going to take it you know from here to the end of the world. Yeah, >> I the you also have Anthropic as well who are, you know, doing very well with with Claude and their stuff. You also then have X AI and I'll let other people have opinions on that. But you've got those sort of three in a race at the moment to do what will be, you know, one of the biggest IPOs, each of them individually. >> Yeah. That the world has ever seen. You know, you're seeing potential IPOs that will be larger than the 10, you know, largest IPOs in history just for one of them. uh in order to achieve the level of funding that they need to be able to get big enough to be able to have the infrastructure they thought >> and Google AI. >> So there might be four players and you'd think that there's four players a bit like in cloud computing. >> Yeah. Well, you're ignoring also Deep Seek and some of the the Chinese competitors as well. >> Um >> so I mean I'm curious around how the public markets will look at it. Um the effort for we work to go public, this is like four years ago now. Um was a shocking example of a company whose valuation was enormous based on what the revenue profile looked like and how sticky those those revenues were and the future value of what that company could achieve. I I'm cautious that someone like OpenAI who's generating somewhere in the realm of like 10 billion a year is worth $900 billion and most of their contracts are forwardlooking that are not sticky because Microsoft just canceled. Perhaps you're right, but perhaps there's a a world by which the the public markets go like we are nowhere near floating this company at $2 trillion like you guys want us to like and and so so that's one to to wait and see. Um but if we go back to the the reinsurance angle and this is I think >> we should >> I think this is um this is a thing that I think is really interesting is because of how AI actually works and because it doesn't have a great ability to like think net new there's a widely held belief that if you had introduced AI like in the early 1900s it could not um develop the theory of relativity and that is to say that it's great at reflecting what things look like and kind of making some connections and and there's some instances where it can do something new like in games right AI was very effective at creating net new ways to win go or something that humans had not thought of but that's fundamentally different than like net new concepts and ways to approach and solve problems within a a very complex like mathematical framework. And the reason I think that applies to reinsurance is I think the vast majority of innovation in our industry sits around how capital comes to market and where we source new pools of capital. How we structure products that are appealing to those pools of capital. how we build products that reflect the dynamism by which our clients exist in the world. So um products that have different structures based on numerous events like you look at a cascading structure a top and drop or something that has second, third and fourth event covers that like are all nuanced like >> lift and shift >> lift like that up. Um but you need people who have like profound genuine understanding of like principles of an industry for the sole purpose of solving problems that their clients face and an awareness of like the way those problems get attached to capital sources and to then come out and say oh we can do this and and if we take away the people for AI efficiencies what I think we will get is the equivalent of whilst leaner like AI slop reinsurance which is just oh it's the exact same all we do is sell a quota share or a two-layer tower and it's this and yep we can clear it at this you know price optimization thing and it it's more efficient in bolded air quotes but is it actually delivering on and and maybe some people in the world will look at that go that's good enough all reinsurance is is a commodity to make these into like but But I think practitioners in our industry actually look at it and go we are leaving profound value on the table in that future use case and and that' be more of my call to arms of like people still need to matter because otherwise we will we will absolutely kneecap a bunch of innovation that we have in front of us. >> This is a funny comment I from another angle I I I agree with what you said. It's interesting I to think about from the point of view of creativity for for sure and also in in an industry like reinsurance where so much of the deal making is not recorded publicly. >> Mhm. >> That the library the reference library I guess for any AI trading model would be quite low. >> Yeah. >> As well right. there's a relatively limited volume, a relative relatively limited number of people who actually understand how to do a good reinsurance deal. So >> yeah, >> if it's making recommendations on the basis of guessing what the most probable sequence of words is, >> yeah, >> based on like the three deals a year that get a press release about them. >> Y >> that's not really going to be good advice tailored to your client and what they ought to do, right? and and likewise they're not going to have access to the reams of data that you know is needed to make those decisions as well. >> Well, and you're right. I think that the fact that this industry the amount of data that lives in this industry is sort of like tucked away somewhere and people who are in it understand it. But even when when we were both consultants and we do things like global market sizing like reading through company annual reports and regulatory you know outputs to generate like what is the market size for these different countries and different products actually requires an enormous amount of like additional understanding of what is and isn't included and why and how these things come together and where things are getting double counted or um ignored entirely and and building a view of that up and because those things it's like oh then it's a report this was in one client's email that was done six years ago or every it doesn't live out in the open like coding does where it's easy to train AI on like the ways to to write software code for different products and different things and we are infinitely far away from like a a lovable sort of like build a website with AI for reinsurance contracts like build because as you just as you rightfully point out like it doesn't live in the open. Um and then you have this this idea of MCP model context protocols like giving the data sets to these models so they can learn how it all works but even then in reinsurance that's not super clear. You can give them every reinsurance contract you've ever written, but it doesn't one, the universe isn't enormous, and two, it's it's harder to understand like the structural value of those contracts because it's buried in legal ease. So, it's it's very difficult to like strip that back out. So what would be the data set that you'd pour into it that gives it like and and there would be versions of that that would live within um a Moody's system or a Verisk system that can say oh here's how you apply catastrophe models and structures of deals towards portfolios like there's something maybe there with enough data to help at least it guess at structures but without the human understanding of like what we're trying to solve for like it still makes it tough like but th those would be the firms I think best positioned for something there around net new structure creation um but even then it's going to be >> I don't know it misses the nuance there's a funny there's a funny thing around um the it was just at a conference and Todd from me to value made the comment >> I of he doesn't want to get li left just cleaning up the dog group. >> It was like this is an industry also quite focused on people and trust and long-term relationships and will reinsurance people vote to replace themselves on any near-term horizon >> especially when it is for many a kind of lifestyle business for one of a better word right yeah >> like I we have already a pretty lean industry if you think about the number of or the deal value per capita in our industry. >> Yeah. >> You know, there are people who are single-handedly responsible for trades worth tens of millions of dollars, hundreds sometimes. >> I And for them to say, wouldn't it be more efficient if I got a machine to do it? Well, why would you? >> You're smashing it as a human. >> Yeah. >> You know, like find things that make your job better. >> Yeah. >> And find things that make your impact better. Sure. But you're not you're not in the market to go and be like, I need to rapidly cut my costs or something like that. You're like, no, you need to make sure that you retain those relationships and keep that trust and still be relied upon, >> be able to show how good the work is that you're doing for your clients. >> Yeah. >> And be able to, you know, evidence it with good data and so on and so forth. You don't ever want to be in the position where it's like, "Guys, I know you love doing business with me and you pay me, you know, tens of millions in commission every year." But actually, I built a widget that's going to do it for me from now on. So, talk to the widget and and that widget is going to do everything. Instead, instead it should be and said, "No, >> I'm doing what I've always done for you. I'm the one you trust. >> Here's the evidence. Here's the data that shows what I've done for you." >> Yeah. >> It's a very different relationship with technology. I think people may miss that nuance, but it's important. >> Yeah, it is. It's super important. The when you're when you're going through that, the the thing it made me think of, and this is quite an old um innovation that I remember from when I was like in high school, but there was this belief that had come out that like in the future those sort of things. But there would be a case where you could just take a pill and it would be all your macros per each meal and this and people like this would be amazing. It would save me so Yeah, exactly. It saved me so much time having to cook and clean and do all this sort of stuff and every in the future everyone would take a pill at lunch, a pill at breakfast, and a pill at dinner and then they'll be off and set. >> But that's actually horrible, right? Because yes, it's annoying when you have to clean up and yes, it's annoying when you have to sort of cook and you make a mess or if you cut your fing like there's there's a lot of things that are unpleasant with the process of making and food, right? But the process of like eating or spending time with family over meals and like it's it's there are I think this is the same thing with reinsurance where trying just to maximize efficiency at all costs is kind of missing the point. >> Yeah. >> Like we should certainly look at it and go oh I no longer need to take notes when I go to meetings because my AI transcriber will be there and it will help me summarize it and make sure I don't miss any next steps. So I can just be like for sure because you spending an hour after every hour meeting writing the notes for it to circulate to your team is probably not the best use of someone's time and things will certainly get missed and you're like like so there are things that are efficiencies that are better for everybody. But it's not we should maximize efficiency at every single possible interaction without a recognition of like what if these things are just nice to do and it's it's important that we spend time and understand what you're trying to achieve and we build these relationships and bonds and I mean I think that's what separates us from a pure commodity in many ways is that whilst there's commoditization elements to our business right use of capital and those kinds of things it is long-term relationship driven and business around like what we unlock and achieve for people. So it's keeping those two things like at some level of of balance rather than use technology to run away from our humanity. We should use it to run towards it. >> A beautiful sentiment >> and one we have to end on I'm afraid because we have run out of time. But thank you for your thoughts. Clearly, there have to be a part two cuz I led us astray for most of the episode on all sorts of random uh sub uh narratives. But more to follow with real reinsurance AI use cases coming your way soon. >> Until next time. Thanks everyone. >> Thanks for tuning in to the Reinsurance Podcast. If you enjoyed our show, don't forget to subscribe wherever you get your podcasts and leave us a review to let us know how we're doing, preferably five stars. For more insights and updates, follow us on LinkedIn and visit our website at supersede.com/mpodcasts. You'll find the links in the show notes.