How The Insurance Industry Is Using AI To Optimize Business

Forbes

This panel at Imagination In Action’s ‘Forging the Future of Business with AI’ Summit of Lisa Dolan, Marcin, Detyniecki, Christopher Paquette, Henriette Fleischmann and Henk van Biljon about how insurance companies are using AI now and how they plan to use it.  Subscribe to FORBES: https://www.youtube.com/user/Forbes?sub_confirmation=1  Fuel your success with Forbes. Gain unlimited access to premium journalism, including breaking news, groundbreaking in-depth reported stories, daily digests and more. Plus, members get a front-row seat at members-only events with leading thinkers and doers, access to premium video that can help you get ahead, an ad-light experience, early access to select products including NFT drops and more:  https://account.forbes.com/membership/?utm_source=youtube&utm_medium=display&utm_campaign=growth_non-sub_paid_subscribe_ytdescript  Stay Connected Forbes newsletters: https://newsletters.editorial.forbes.com Forbes on Facebook: http://fb.com/forbes Forbes Video on Twitter: http://www.twitter.com/forbes Forbes Video on Instagram: http://instagram.com/forbes More From Forbes:  http://forbes.com  Forbes covers the intersection of entrepreneurship, wealth, technology, business and lifestyle with a focus on people and success.

Transcript

00:00 >> Great.

00:02 My name is Lisa Dolan.

00:03 I'm a managing director with Link Ventures.

00:06 You probably have heard Dave Blunden speak.

00:08 We're a fund that's very much focused on AI.

00:11 We look at foundational models, horizontal models, and then also verticalized applications.

00:16 And so today we're going to be talking about the application of insurance.

00:21 The PNC market in the U.S. alone is just under a trillion dollars.

00:25 So this is in a camp of very high value user.

00:29 Even verticalized applications are really focused on founders that really understand

00:33 their user and are building proprietary data.

00:38 And so today we're talking to enterprises and startups alike, some of which we have

00:43 invested in, some of which I wish I had invested in.

00:46 And so to start, I'd like everybody to introduce themselves, and then we can go into it.

00:53 So starting with you, Marcin.

00:55 >> Yeah.

00:56 I'm Marcin Wietniewski, I'm the group chief data scientist of AXA, which is not that

01:00 known in the United States, but -- >> Oh, no, not that -- no.

01:03 Very well known.

01:04 >> One of the biggest insurers in the world.

01:06 And yeah, and I run a particular research team also.

01:10 And moreover, I'm one of the guys who started his PhD thesis during the winter of AI.

01:17 So I have been around on AI for a lot of years.

01:21 >> The winter of AI, can you give us like a rough date for that?

01:24 >> For my PhD?

01:25 >> Yes.

01:26 >> 2000.

01:27 >> 2000.

01:28 >> Okay.

01:29 >> Go ahead.

01:30 >> Hey, everybody.

01:31 My name is Christopher Paquette.

01:32 I'm the chief digital transformation officer at Allstate.

01:35 Hopefully a company you've heard of.

01:36 We spend a ton on marketing.

01:37 We're the good hands, right?

01:38 Go to Allstate.com, sign up today.

01:42 My role is as chief digital transformation officer, everything from process automation

01:47 and digitization with AI infused to creating better experiences for our customers and flows

01:53 and self-service and stuff like that with predictive service to longer term digital

01:59 incubation and thinking of the next disruptive product service models and routes to market

02:05 for the company.

02:06 So a lot of fun stuff that I get to play around with at Allstate.

02:10 >> Hi, everyone.

02:11 My name is Henriette Fleischmann.

02:12 I'm the co-founder and CEO of Posta AI.

02:15 We automate insurance claims, but we have underlying a couple of foundational vision

02:19 models that allow us to create spatial understanding of just 2D images.

02:25 Quite excited.

02:26 We're born out of MIT.

02:27 I'll hand it over to -- >> Thanks so much, Henriette.

02:31 My name is Henk van Leeuwen.

02:33 I'm an actuary.

02:34 I used to lead an actuarial data science team at PwC.

02:37 I'm currently actually finishing up at MIT.

02:39 But along with that, I'm also the CEO and co-founder of Fount AI.

02:43 Fount AI is a company that helps property and casualty insurers to optimize customer

02:47 acquisition through intelligently understanding risk profiles and combining that with marketing

02:52 technology and also the latest advancements in AI to help our customers acquire business

02:58 more effectively and efficiently.

03:00 >> Awesome.

03:02 So what's so exciting for me, for my seed, is seeing how quickly the space is developing.

03:08 We have LLMs as a clear application for such a heavy document, heavy language industry.

03:15 Intervision is a clear application for claim verification, as Henriette can speak to.

03:22 And then multimodal models are a clear application for fraud detection, which is a big issue.

03:28 And so for my seed, it's really exciting just how quickly it's moving.

03:32 So from your seed here, let's start with you, Marcin.

03:35 How are you at AXA?

03:37 Can you tell us a little bit about some concrete examples, concrete use cases that you guys

03:42 are using in-house, AI use cases?

03:45 >> Okay.

03:46 So I think I will not talk about the traditional AI because this is -- everybody wants to hear

03:52 about the LLM use cases.

03:54 So we're thinking about LLM in two ways.

03:58 The first way we're thinking is -- and actually we deploy to all of our employees, I mean

04:03 collaborators, which is around 150,000 people in 55 countries.

04:08 We kind of produce or put in place a secure access to open AI backend LLM.

04:18 And the use case behind -- so the thing is, it needs to be secure.

04:21 So that's the thing.

04:22 We don't want our data to go there and so on.

04:23 So we have a kind of a quite complex system so that we make sure that our data remains

04:28 with us.

04:30 And the use case behind that is that we believe that these LLM copilots are not smart enough

04:35 at this stage to kind of take full tasks.

04:38 But there are a lot of little tasks all over the place which are going to bring a lot of

04:42 value.

04:43 Okay?

04:44 And this is one of our use cases is putting a -- and what is very interesting is that

04:47 since we're handling ourself, our platform, we are not using just the intelligence of

04:53 the LLMs.

04:54 We're using the intelligence of collective intelligence of the users.

04:58 Because now one of the challenges about this is that people don't know what LLMs are.

05:03 And the training is not, okay, you have a transformer and this is a decoder and a decoder.

05:07 It's not what people want.

05:09 People what they want is to understand what the machine can do for them.

05:13 And so how we crack that one is we have actually prompt types.

05:18 And actually people can propose prompt types and people can say I like them or I don't

05:21 like them.

05:23 And this is also organized by vertical.

05:25 So this allows us not just to have intelligence coming from the LLM itself but actually collective

05:31 intelligence because people come up with prompts which are very useful for our distribution

05:36 channels, for underwriting and so on and so forth.

05:38 So that's one kind of big case we have.

05:42 Then we have the classical ones that maybe some of our guests are going to talk, like

05:46 call centers.

05:47 The other one I wanted to talk about is let's say our big bet.

05:50 And so our big bet is that if you think about insurance, what do we sell is a promise.

05:55 And the promise is encoded in a contract.

05:58 And contract is wording.

06:00 Okay?

06:01 So the thing is we're saying, okay, that's great.

06:04 Now we don't believe or at least we don't trust LLMs yet enough to say, okay, LLM, just

06:09 take care of the contract and do whatever you want.

06:12 What we do is we translate.

06:13 We use the capacity of this contract of these large language models to translate.

06:18 And we translate the contracts into some specifications for our systems.

06:23 And so we have for our call centers and things, we have systems of rules and we call this

06:28 a computable contract approach.

06:30 We use these LLMs not to answer directly the customer, but actually to translate the contracts

06:37 into this.

06:38 And if you see the economics, it's very interesting because, you know, I query an API to open

06:43 AI 4.0.

06:44 It's quite expensive at the end of the day.

06:46 By doing this, you query only once per contract and not once per claim request, which is very

06:52 interesting.

06:53 So that there are two things.

06:54 We have, like, the easy one that I think it's underestimated.

06:58 And then we have this kind of more bigger bet, which is quite transformative because

07:03 then once your contracts are in a computable form, you can do a lot of things like patching,

07:07 like compiling, like comparing them, like simplifying your portfolio.

07:11 So that's how we see at AXA in LLMs transforming our business.

07:15 Okay, so this is a question for all of you guys, maybe starting now with you, Christopher.

07:21 How do you see AI transforming pricing and underwriting and claims processing in one

07:27 year today and then in five years?

07:30 And this really hits on the point of how quickly we're moving.

07:34 That's point one of the question.

07:35 Point two is then how do we see the balance of risk shifting and then also where human

07:41 capital should be living?

07:44 So let's start with Christopher and then we'll circle back to you.

07:48 That's like six questions that I just got asked all at the same time.

07:51 A lot of fun.

07:52 I'm happy to get your perspective first.

07:53 Let me start maybe in claims in one year and some of the types of things that we're thinking

07:58 about that I know other major insurers are as well.

08:01 And they all, for me and us, like they got to come back to like what's going to create

08:05 value for the business, right?

08:06 And like there's a few easy examples that I always go to to illustrate where when you're

08:11 processing a claim, you got to get a few things right.

08:13 So one, you got to get all the contact information up front because it's not just the people

08:17 that were in the car that were hurt.

08:18 It's the people in the other car or people in the corner.

08:21 So there's like information collection.

08:24 There's understanding the severity of the accident that you got to get right as well.

08:27 There is using some of that data that's there, including then also telematics that gets thrown

08:33 in and traffic and weather conditions and like all this stuff that comes around.

08:39 So like problem number one is how do you get all of it in?

08:41 And there's an AI opportunity there, of course, whether it's telematics or computer vision

08:46 or whatever it is to get the data in.

08:49 So that's step one.

08:50 Step two then is like, okay, how do you create value from that?

08:52 Well, there's a few things that just like anyone who's worked in claims and tell you

08:56 you got to get right.

08:57 So you have to identify as soon as possible those claims that have not just property damage,

09:02 but way more importantly, bodily injury.

09:04 Why?

09:05 Because somebody's hurt and because you got to take care of that person.

09:08 And if you don't, it's bad for them, of course, it's also more expensive for the insurer and

09:12 everyone involved.

09:13 So identifying BI, you want to reduce attorney representation rate.

09:17 Why?

09:18 Because that's just external third party spend that can't go to the person that's hurt to

09:22 help them get better.

09:24 And so things like, there's a number of these like, you got to do this, you got to do this,

09:28 you got to do this.

09:29 And so what we're using is that data collection to build models to do all of these things

09:35 more effectively.

09:36 And they're not new models, but man, they're maturing a whole lot more rapidly than they

09:40 used to.

09:41 So that's kind of where I see it claims in a year.

09:44 Maybe I'll let somebody else jump in on the five year part or the other parts.

09:47 Okay.

09:48 So you jump in on the five year part and then also how you see risk changing and where human

09:54 resources should be living.

09:56 Yeah.

09:57 Okay.

09:58 So, okay.

09:59 So on the five year and on pricing, I really will answer on the pricing part.

10:02 So I think on the pricing and five years from now, I hate actually to make prediction because

10:08 I'm a research scientist.

10:11 But if I have to do this, I would say, and I agree with you, I think one of the thing

10:15 is the pricing will be, we will be able, and I don't know if market will go this way, but

10:21 to price more dynamically.

10:23 Okay.

10:24 Because then you say mentioned telematics, for example, anything which is a flow, today

10:28 we do everything this in a statistical way.

10:30 We have a lot of testing.

10:32 If you think machine learning, kind of if you project yourself, I mean, you know, GLM

10:35 is like a, let's say it's a one layer neural net.

10:38 So you can quickly imagine you have a bigger net.

10:41 So you could then transition this telematic and price directly on the behavior.

10:45 So you could, maybe telematic is not the only application, but you could do more dynamic

10:50 things.

10:51 That's one thing.

10:52 The other thing, and I think you would react on that, it's not necessarily the claim, but

10:55 it's AI is also able to understand better the world.

10:58 Okay.

10:59 It's kind of like, because you have images.

11:00 So for example, you understand better your risk.

11:03 And so it's going to allow you to price much better.

11:05 And these at scale, because usually one of the problems of these great ideas of aiming

11:09 data is that you don't have data.

11:11 Now, if you have satellite data, you have satellite everywhere, so you can look everywhere,

11:15 but now you need to have labels.

11:16 And I think here AI, maybe that's not the five year far away, but I think this is where

11:21 I see changes on the thing.

11:23 And maybe you want to react and then I come back to the talent.

11:27 I can speak a little bit for the property side.

11:30 So if I look at underwriting risk assessment and then claims, I think in five years from

11:34 now, you probably have areas that are not insurable anymore at all.

11:37 I mean, you already see people like Liberty Mutual, for example, just opted out of Marblehead,

11:41 which is an area here north of Boston.

11:43 Right.

11:44 And then to Massachusetts, because the risks have been becoming too high for flood insurance,

11:48 et cetera.

11:49 Flood usually is not covered at all, but the rest of the houses are affected anyway.

11:53 So how do you price that into your models?

11:58 How do you ensure that areas in Europe, America, et cetera, are still insurable?

12:04 I don't have a good answer to that because there is climate change to that.

12:07 Probably also a government aspect that needs to cop into it.

12:11 I agree on the risk assessment, something that is done today.

12:15 I think it can also be used in catastrophes to set a claims faster and allow people then

12:21 to be triggered very specifically to take additional pictures of their houses, homes,

12:28 et cetera, based on the catastrophe impact to then get them settled immediately.

12:33 That has two impacts, by the way.

12:35 It makes it cheaper to settle that claim because the repair can happen faster because the contract,

12:41 contractors can be engaged quicker.

12:43 I think on the risk assessment side, it's also interesting to collect more and more

12:47 data around appliances, around different interiors, et cetera, to get the depreciation right.

12:53 And a content detection is one of the hardest pieces there because the differentiation is

12:58 so critical in terms of understanding value.

13:02 And that requires you to have an understanding of any kind of furniture and label of furniture

13:07 and pricing of furniture or other things.

13:09 And it's a very complicated problem to solve.

13:12 So maybe in five years, definitely not right now.

13:15 Can I jump back in?

13:18 Yes.

13:19 There's kind of like when you got to understand what you're insuring the value of it in the

13:22 property, the pictures, the furnitures, everything.

13:24 And then there's also your bit around like, okay, how do you understand the risk and price

13:27 it much more accurately, not just for the person, but for the situation, the context

13:31 of that person is right there.

13:33 And so you could like let your mind run wild and get to a place where like, well, you have

13:37 very, very low premium auto insurance, but on the contracts that you were talking about

13:43 before, like you sign and you say, okay, I allow you to understand what I was doing in

13:47 the three minutes right before I got in a car wreck.

13:49 And if it was my fault, then my deductible is no longer 500 bucks.

13:52 It's $5,000 or 10,000, you know, and there's that risk-based contract that is available

13:57 and possible because of better data flow.

14:00 Same is true for properties, by the way, once you have identified risk, you can also do

14:04 risk preventive measures.

14:05 And there's more and more happening by insurance carriers around that, right?

14:08 Because insurance is becoming more and more commodities.

14:11 So how do you differentiate?

14:13 How do you become a partner by understanding leakages very quickly and have sensors that

14:18 are also becoming cheaper and cheaper and other things that you could implement there?

14:23 Yeah, absolutely.

14:24 So actually onto sort of the commoditization piece, I'd love to have Hank talk about the

14:30 acquisition piece, right?

14:32 So acquisition is Hank's focus and how you see the insurance acquisition business changing

14:39 from one year to five year and how AI is going to change.

14:43 Yeah, sure.

14:44 Thanks, Lisa.

14:45 I think one thing specifically relating to the one-year problem is the introduction of

14:50 agents.

14:51 I think Marcin put it really, really well is that small, simple task, there's massive gains

14:55 that we can make in that specific space in terms of introducing autonomous agents and

15:00 making that significantly more efficient.

15:03 I think specifically generative AI allows you to produce personalized and much more

15:08 individualized content in terms of reaching your actual target audiences to acquire them

15:13 through specific channels.

15:14 In other words, basically getting the content that you want in front of a specific audience

15:19 now is much easier and much faster than it has ever been, which allows for that personalized

15:24 outreach.

15:25 That also affects your risk profile because if you are able to do better risk targeting,

15:29 you're actually able to influence your underlying risk portfolio in a much more effective and

15:34 efficient manner at the end of the day.

15:36 The five-year journey there is, I think, one, it's a story of integration.

15:40 I think as we move from, I guess, these small, simple little tasks that Marcin mentioned

15:44 to, I guess, the big ticket items in terms of dynamic pricing, incorporating dynamic

15:49 pricing and including that within the acquisition side as well, integrating that part, understanding

15:56 risk profiles right at the top of the acquisition funnel and being able to link up pricing,

16:01 risk, and then also marketing as well as things like dynamic contracts, for example.

16:06 I think really there is the holy grail that you will be working towards the end of the

16:10 five-year period.

16:11 Yeah, absolutely.

16:12 Does anyone else want to talk about acquisition, perhaps at Allstate and how you're thinking

16:18 about how AI is going to change acquisition at Allstate?

16:21 Yeah, I think it's, I mean, similar themes like finding the right customer and getting

16:25 them to the right place.

16:26 We're one of the only insurers in the US that has all the channels.

16:30 We have, you can buy direct online or mobile.

16:33 You can call our own captive call center.

16:35 You can go to an exclusive agent who aren't Allstate employees.

16:38 Then we also have independent agents that offer, who are not employees and offer Allstate,

16:43 but also other insurance carriers as well.

16:46 Across those channels, we get a reach out like, where do they belong?

16:49 There's intelligence that has to go in there, of course.

16:53 You didn't speak too much about it, but also this idea of not just like, how do you attract

16:57 them and get the right message in front of them, but also how do you use those factors

17:01 to begin to rate at step one, right?

17:04 So that you understand what's the risk of this person and therefore what channel do

17:07 they belong or what products are they going to be exposed to?

17:10 Really keeping that transparent the whole way.

17:13 We use an internal, we call it Dell style pricing because Dell used to do this thing

17:17 of like, well, if you add a mouse and remove a keyboard, here's how much your computer,

17:21 like when you could build it online, if you remember back like 25 years ago, because I'm

17:26 old.

17:27 But like the same thing for insurance.

17:28 So like, well, if you want to raise your deductible, but lower your premium and how are you driving

17:32 and are you going to sign up for telematic?

17:34 That transparency of the rating factors that are underlying it have to start at step one.

17:39 And I think if I can add on to that, having that transparency right up front actually

17:43 helps in terms of prioritizing customer acquisition.

17:46 Because for example, if you know from a post sales data perspective and from a customer

17:50 value management perspective, if you know that your cross-sell opportunities, for example,

17:54 across multiple different products are that much higher, you can put in a lot more effort

17:59 and spend a lot more on a specific customer, maybe a specific customer demographic and

18:02 actually attracting them in the first place, which currently maybe tier one carriers like

18:07 the likes of Allstate are currently thinking about that as well.

18:10 But I know tier two, tier three carriers just aren't there yet.

18:13 And this is allowing us to basically take that insight and bring it right up to the

18:16 top of the acquisition funnel and actually implement that in terms of outreach and marketing.

18:22 So let's move to startups.

18:24 So I understand that both AXA has an incubation function and then Allstate also does investments

18:30 and then we have startups here.

18:31 So tell me how you at AXA think about generally your innovation strategy, when you decide

18:38 to incubate, when you decide to partner.

18:42 And then also, Christopher, I would love to understand how you determine when to partner

18:46 with startups.

18:47 So starting first with you, Marcel.

18:49 Yeah, so we have traditionally, we have one venture part, which is venture capital activities,

18:56 which kind of like what you do.

18:58 So I will not talk about that.

18:59 This is a classical kind of thing.

19:01 No, what is quite unique or what the new thing or new thinking that we have, and I'm talking

19:07 about internal innovation, what is quite unique is that, so before you used to build things

19:13 and believe that these things are going to be necessary for us.

19:17 So they're going to transform.

19:19 And one of the challenges we observe in the early, not early, but like after early stages,

19:25 not meet yet, is that, you know, the competition out there in terms of capital is so big that

19:31 we cannot fight.

19:33 So actually now we are thinking of spinning quite early.

19:36 So start to try, we try to incubate, I would say use cases, if I'm really honest, is like

19:42 matching use case to technology.

19:44 And then once this is kind of like already well connected to our business, to shift it

19:49 down so that they can be funded by capital.

19:52 And then this will scale will accelerate us.

19:55 So it's kind of a new way of looking at this venturing thing, because by keeping it side,

20:02 what we have seen in the past is that we build, we build, we build, and we're a little bit

20:07 earlier sometimes before the market, but then we get caught because there is a moment that,

20:13 you know, this becomes too big, it's too expensive and we cannot compete.

20:17 So that's kind of a new way of looking at these things because it's building to spin

20:21 off in some sense.

20:23 And so we try to make sure that this will be value when it's been off.

20:26 There's some capital challenges there, how many shares we keep and so on, because, you

20:31 know, if not, it's not so obvious to raise funds, but this is what we're doing.

20:35 And it's, I mean, early stages is working.

20:40 So I would say that we might have the opposite problem of building too late, too early.

20:47 And that's part of a legacy culture of, you know, if not built internally, then it doesn't

20:54 work as well.

20:55 And I think there's some truth to that, especially given the importance of the data and the models

20:59 that are underlying, and if you build it and you build it well, it can have a lot of value.

21:03 But the flip side is we build things too late.

21:06 And so we're changing that culture slowly, but asking questions like, you know, there

21:12 might be four big ones.

21:13 So like, can we do it faster?

21:14 Is it faster to do it externally?

21:16 Is it a commodity that eventually we'd be fine, like having the same thing as everyone

21:20 else or could it create a lot of value if it's proprietary?

21:24 Or something around, you know, just the cost to do it.

21:30 And then just is there like executive appetite and like, could you actually go do this thing?

21:35 And so like using those four questions, often you arrive at an answer of like, wow, this

21:40 could make sense at least to get a start with partnering and looking at buying and building

21:44 at least the first.

21:46 Maybe Christopher, I can ask you a question from the startup side, something that we look

21:49 at like, how do you approach a behemoth like an all-state, for example, like data is something

21:54 that the incumbent massive carriers have, which startups really want to get their hands

21:58 on and in terms of basically breaking that down into small little pieces that startups

22:02 can actually help yourselves out with and get on board with, for example, for startups,

22:07 how do you see that actually working and what does a successful approach look like in your

22:11 eyes?

22:12 Yes, I can't give you a direct answer on the data, obviously, without a whole lot of specifics,

22:16 but I'd say like the two critical factors for a startup working successfully with all-state,

22:20 and I'm just going to paraphrase your question a little bit, and it's like the Venn diagram

22:24 is like maybe 80%.

22:26 So one is like having just that really, really clear value-based use case, like it's one

22:31 thing that you do and you do it really, really well in one area and having an executive that's

22:36 willing to sponsor that.

22:38 I think if that's the case, then like we actually can move really, really fast to get the data

22:43 available to run the pilots, free pilots, invested pilots often don't hurt either.

22:49 So we do a lot of those, but it's that kind of stuff of like making it easy to engage

22:54 with that.

22:55 One thing I would add, which is maybe obvious, but it was not obvious for us for working

22:59 with startups is that you need dedicated teams internally for whom the success of the startup

23:06 is their success.

23:07 Because what happens is that most often the startups talk to the innovation guys and it's

23:12 not a success implementing a startup.

23:16 And so we have dedicated teams for this.

23:18 We call them, I don't remember, we call them startup clienting.

23:23 And so they do, well, they put a little bit pressure on the startup because then we do

23:28 panels, we do competitions, but then once one startup is selected, they are helped to

23:33 succeed.

23:34 And this is something which is missing.

23:36 At least it was missing in our case.

23:38 I agree with that.

23:39 Yeah, I think we see a lot of startups working with enterprises and ultimately it's great

23:43 on day one when they're talking about the partnership, but then ultimately when it gets

23:47 to execution, it gets lost.

23:49 There's just so many priorities that an Allstate may have or an Axe may have.

23:54 It's really hard to really prioritize a given startup.

23:56 So I think we discuss this and we discuss this in-house all the time, is the biggest

24:01 problem and opportunity that a startup has, if we talk to startups here, is data and is

24:06 access to data.

24:08 And so what we see is a lot of point solutions with insurance companies, a lot of point solutions,

24:12 a lot of MGAs that are focused on a given demographic.

24:17 And so how at your startups are you thinking about a mission that is both a point solution,

24:23 but understanding that in order to really build value for a carrier, you need to have

24:27 an ecosystem approach and you need to have a sort of a platform approach.

24:32 And so starting with the startups there, starting with Hank here, how do you at your startup

24:38 make that a mission?

24:39 And then perhaps we can talk about it and you guys can give some feedback.

24:44 So I think specifically in our instance, it's great that better risk assessment and moving

24:49 risk assessment to the top of the acquisition funnel and being risk away earlier is something

24:53 that everyone wins out of.

24:56 So I think that's really a mission for us that is very, very easy to sell to carriers

25:00 and get them to invest in that ecosystem because at the end of the day, it's something where

25:04 everyone wins.

25:06 I think our biggest challenge, as you mentioned, is just around actually carriers sharing their

25:11 data with us, which helps us to assess those risks.

25:14 Currently, how we're going about that, luckily they are industry wide data providers, which

25:18 we are leveraging, I guess, to get a broad sense of the data to get us started off with.

25:24 But from there still as a startup, I guess a big pain point that we have is we still

25:27 need that carrier specific data, which really represents their own risks to drive us further

25:33 in terms of that solution.

25:34 So I think the incentives are there and the base is there, but access to data, which is

25:40 I guess why I asked the question a little bit earlier, is still a really significant

25:44 pain point for us and I guess driving that overall mission.

25:47 I think it's the same as if you would build it internally, right?

25:51 You need to understand the user journey.

25:52 You need to understand where you're creating value and which systems and workflows do you

25:56 need to integrate with in order to make it easier for the person who's using your solution,

26:00 not harder.

26:01 So that means on the underwriting side, it might be enough to provide a PDF with a score

26:05 or a JSON file or whatever that can be integrated into existing data that is already there and

26:11 just edit to create a better business profile or different pricing model.

26:15 While on the claim side, when you drive estimates, in the United States, there are two big providers,

26:19 which are Verisk and CoreLogic that provide pricing information.

26:22 And it's clear that if you're not integrated with those, then you're not adding value because

26:26 then everything you're doing needs to be manual, got into it.

26:30 So you need to drive partnerships on those ends, right?

26:33 And make sure that the data is properly integrated.

26:36 And then there are other integrations that come later in order to make that process of

26:40 claims estimation completely lean, which is a challenge for a startup because not only

26:44 do you need to be sub-type 2 compliant, you also need to be integrated with all these

26:48 big players, right?

26:50 And you need sponsors usually, which is a chicken neck problem.

26:53 So it helps to find strategic partnerships with tier one or tier two carriers to drive

26:57 those integrations and make sure that the user journey is working perfectly.

27:01 Yeah.

27:04 So yeah, actually on the other side of the spectrum, so we kind of realized this, that

27:09 actually integrating startups is a big pain, in particular on data and so on.

27:13 And so what we did, and it doesn't work for all startups and all use cases, but we have

27:18 platforms.

27:19 So we have built two platforms and it doesn't work for all.

27:22 I mean, it doesn't work for somebody who's doing pricing, for example.

27:25 So we have one for health and one for commercial.

27:30 Yeah, it's a PNC, but commercial line.

27:35 And so these ones, these platforms actually integrate all data and try to coordinate all

27:39 the actors because one also of the challenges is that startups tend to solve a little problem

27:45 of the bigger picture.

27:47 And so we have built the platforms and so we can, for example, if you think very simply

27:51 risk, okay, and then you want to understand the risk from a corporate, okay, we want to

27:56 have understanding from, I don't know what, what is like, I mean, it's like natural catastrophes,

28:00 but we want to understand what in terms of fire, what you want to understand.

28:03 And so there are startups for each of these things.

28:05 And so what we built is we have this thing that connects to each other.

28:10 And then what is interesting is that now we're discovering is that initially we had one customer,

28:14 which was risk analysis, but then suddenly we discovered that there are other services

28:18 that are interested in particular for prevention because then, you know, when it's a big corporate

28:22 and then you tell them, Hey guys, you have a problem.

28:25 So this platform approach is interesting, but needs to be kind of focused that the danger

28:31 with platform is that then you say it does everything.

28:34 So that's why I said at the beginning, it works for some group of startups and these

28:39 you crack definitely because then you have the data, you have your APIs, everything,

28:43 and then it works.

28:44 And the integration is easy.

28:45 Yeah.

28:46 So sort of a final question for the enterprises here.

28:51 I was initially going to say, you know, what AI innovation are you most excited about?

28:55 But I'm actually going to change the question and ask if you did a startup, where would

29:00 you focus?

29:01 And I think that's a very interesting thought as you guys are sort of integrated into the

29:06 world, see where it can sell and what cannot sell with enterprises.

29:09 So just imagine, you know, you're young, you know, I mean, you're still young, but if you're

29:14 young, imagine we're not old.

29:16 Imagine we're young.

29:17 Okay.

29:18 You're just out of MIT.

29:19 Imagine the past.

29:20 You're just out of MIT.

29:25 Seeing what you've seen, where would you start?

29:29 You know, where would you start?

29:31 Aside from, of course, you know, where Henrietta and Hank are, but like, what would be, what

29:36 would be your startup?

29:38 Okay.

29:39 So, so if I was really young, I would like a moonshot.

29:43 No, actually it's not that of moonshot, but it's something I'm really interested in.

29:48 I would actually look at insuring AI.

29:52 It's like, how do we insure AI?

29:54 This is clearly something that I think is exciting and it's related to all my research

29:58 topics because I'm very concerned about, you know, risks on AI, transparency, fairness,

30:04 and so on.

30:05 But then you can see it the other way around.

30:06 Okay.

30:07 These risks, you understand them now.

30:08 Okay.

30:09 How do we insure them?

30:10 What do you do with it?

30:11 That's probably, I will probably make a startup, but I'm not good as an entrepreneur.

30:15 So maybe.

30:16 That's why you're old and work for an enterprise.

30:20 I have an MIT PhD who I think is someplace here.

30:22 I want to connect you to him working on that.

30:24 How about you?

30:26 So I would say, mine goes back to like insurance fund, it's been around forever.

30:30 Right.

30:31 And like, even at its best today is like, well, we hope nothing bad happens, but if

30:35 it does, we'll pay for it.

30:37 And you feel good.

30:38 And like, it's a piece of paper.

30:39 Right.

30:40 And that's the promise.

30:41 And I think there's nothing better than that.

30:42 And so with the data and AI that we have available to us, I think there's an opportunity for

30:47 all of insurance or a startup to pivot more into like, it's often written in the insurance

30:53 rags of we can predict and prevent risk instead of just covering it.

30:58 And so that's a process of like, well, I can analyze your behavior and what you're doing.

31:02 And what are the risks in your home and where the risks with how you drive and how you conduct

31:06 your daily behavior.

31:07 And like, that's hard enough to actually predict like, okay, what's the danger?

31:10 Like, that's kind of out there with telematics and other things already.

31:13 But then how do you prevent it?

31:15 How do you change my own behavior?

31:16 How do you influence?

31:17 And that's where large language models come in.

31:19 It's where influencing techniques come in.

31:20 But like, if someone could really crack that, like, none of the big insurers are really

31:24 doing that.

31:25 It's all just like, well, here's a check.

31:26 Sorry, something happened.

31:27 But like, if you could really get there, that'd be cool.

31:30 Yeah.

31:31 Awesome.

31:32 So I think we're almost out of time.

31:35 But if anybody has any questions in the peanut gallery that you guys like to ask the panelists.

31:44 You didn't cover the talent question.

31:47 You forgot the talent question.

31:48 Talk about talent.

31:49 The talent.

31:50 Yes.

31:51 Go for it.

31:52 Yes.

31:53 So I suppose that was one part of the question is where do you feel like talent will be going

31:56 in one year and five years?

31:58 We didn't touch on that.

32:00 So I think like this one, I have a one year and a five year version.

32:03 So I think for the one year version, I think that talent is going to go for Facebook or

32:09 for Meta or for saying, and then five years in the future, they're going to come back.

32:15 And I think that and I think the reason is that actually insurance is quite interesting.

32:21 And it's compatible with kind of a mindset, first of all.

32:26 And second, I think this is being AI is being very much democratized.

32:31 It's getting much closer and closer and not only not only LLMs, of course, which are completely

32:35 accessible, but I think even AI, it's going to be I mean, deep learning and all these

32:39 things are going to become.

32:41 And then I think that the the actuaries and the data scientists is going to kind of fusion

32:47 into one role.

32:48 There are the actuaries and data scientists are going to be more or less the same.

32:53 Maybe not absolutely in everything because maybe the product, I would say it's the product

32:57 manager and the actuary, not necessarily the scientist.

33:00 Yeah, maybe.

33:01 Maybe.

33:02 Yeah, I agree.

33:03 Maybe because it will be so democratized.

33:05 But the thing is, then talent in general.

33:07 Well, yeah, I'm more of a data scientist.

33:09 I agree.

33:10 You're right.

33:11 We'll come back because then you will be interested into having impact in real life people.

33:16 I mean, this insurance, it really impacts a lot of people and for real.

33:20 And then I would also ask the question, which I don't think we touched on was where risk

33:24 will be moving to with AI?

33:28 Where risk will be moving to?

33:30 Yes.

33:31 Where risk will be held.

33:32 It's not, it will no longer be only in underwriting.

33:36 Where will it move to?

33:38 So on the risk part, sorry for, you want to say?

33:41 No.

33:42 On the risk part, I think there is something interesting on the risk part is that, so AI,

33:48 for example, if you think autonomous cars, right?

33:50 So now the thing is, this risk is going to be moving from random drivers having accidents

33:56 to now a systemic risk.

33:58 So much less accidents, but now suddenly you have a whole fleet of cars, which are all

34:03 autonomous and then let's say you have a bug and then you have this systematic bug and

34:08 then you have problems all over the place.

34:10 So risk is going to be transformed in some places because society will be transformed

34:14 because of AI.

34:16 And what would be very interesting is not the target, by the way.

34:19 I mean, this is an intellectual exercise.

34:20 It's in the middle, having autonomous and drivers at the same time.

34:24 So you have the systemic risk and on top of it, this thing.

34:27 So I think this is clearly a question that risk is going to be evolving and transformed

34:32 because of AI.

34:33 I would agree to that.

34:34 I think there is a recent lawsuit filed against State Farm who used a fraud model and was

34:42 heavily biased against certain parts of the population.

34:48 And so that is part of the systemic risk that wasn't exposed before, but it was trained

34:53 on historical data.

34:55 And as long as you do individual decision making, right, there cannot be claimed that

35:00 there is a systemic risk.

35:02 So that is going to be interesting for many insurance carriers or not just insurance carriers.

35:06 It's a general risk overall, right?

35:08 Because you have now a model that is applicable to many and it's one model.

35:14 The bias or whatever, whatever hallucination or whatever is going on in that model is a

35:20 risk you're exposing to a much broader population, even though the individual human might be

35:24 equally as biased or hallucinating.

35:26 Yeah.

35:27 And then Hank, what are your thoughts of where the risk and talent will be going?

35:31 Sure.

35:32 I think something else which is maybe under assessed is model risk in the organization

35:35 itself.

35:36 So model risk is an internal risk.

35:39 Insurers are made up of enormous amounts of models, some of the thousands of models that

35:42 you would be very well aware of.

35:45 And I guess with this proliferation of models that we're seeing and also the increasing,

35:49 as I mentioned earlier, integration of these models with each other, that's an enormously

35:53 complex ecosystem to actually manage with the introduction, for example, of agents,

35:58 autonomous agents now making decisions.

36:00 Suddenly model risk, even though the opportunity is massive, that is going to be a massive

36:04 challenge as well at the end of the day to quantify and manage that model risk within

36:08 the organization.

36:09 Yeah, absolutely.

36:10 Great.

36:11 Well, you know, you guys can come to us afterwards.

36:16 I'm investing in AI and insurance.

36:18 So yeah, thank you guys.

36:20 Thank you.

36:21 Thank you.

36:22 Thank you.

36:23 Thank you.

36:24 Thank you.

36:25 Thank you.

36:26 Thank you.

36:27 Thank you.

36:28 Thank you.

36:29 Thank you.

36:34 Thank you.

36:39 Thank you.

36:44 [BLANK_AUDIO]

Category

Transcript

Recommended