CHARLIE SAMOLCZYK: In this episode of Talking Tech, I am thrilled to be joined by Laura Doddington and Lauren Finnis, both of whom work with WTW. And we are going to explore, I think the really timely topic of in this world of machine learning and AI. What's the role of the insurance expert and how is that being perceived and how is that changing. But before we jump into that, maybe actually I will ask both of you to introduce yourself and just tell us through a little bit what you do for the organization. Lauren, do you want to go first?
LAUREN FINNIS: So Lauren Finnis. I lead our commercial lines practice for insurance consulting and technology in North America.
CHARLIE SAMOLCZYK: Awesome.
LAURA DODDINGTON: And I'm Laura Doddington and lead our personal lines practice for insurance consulting and technology in North America.
CHARLIE SAMOLCZYK: Awesome. So we have all the bases covered. That's great. Let's jump in. So maybe we can tackle this head on. So, do you think machine learning, AI, can really replace the insurance expert and what are your thoughts on that?
LAURA DODDINGTON: So the real short answer is no, but the much longer answer is I think machine learning and AI are going to make an enormous difference to the industry. They already are making an enormous difference in the industry. And we see that whether that's through more advanced risk modeling and pricing and underwriting, whether it's through monitoring calls using voice recognition, whether it is through the automation that we're putting into claims practices, the digitization of our journeys.
Across all of that, machine learning and AI are making a huge difference. But I don't think that replaces the role of the expert. I think it is dangerous to just say, OK, great. Machine learning has got this. I don't need to oversee it. I don't need to view what is happening. I don't need to have transparency around it.
Whether that is around the quality of the data, for example, that the machine learning is using, whether that is around biased outcomes, which we know we've seen from some models, unexpected things that the model is doing that we didn't expect because of the level of complexity there. I think there's a lot of challenge around that, and I think it is absolutely imperative that the experts are there looking at what it's doing, making those models more transparent, opening up that black box so that we can truly understand what it is that we're doing in the marketplace.
LAUREN FINNIS: Yes. And I'd say definitely yes for commercial lines as well. The carriers we're talking to every day are really trying to figure out what is that optimal line between the human and AI and machines. And in that STP straight through processing place, where does the human touch add the most value is kind of the constant question looking in. And there are a lot of work streams out there to look at that and look, where does the human actually beat the machine? Where did the human make a different decision?
A lot of the places we see in the commercial space because corporations are less homogeneous, you see that there are just some places where there are too many variables for us to map and so we still do need a human. But I think in the world of AI, and it's not a new concept, people still need to understand what AI is doing and how those applications are working. I mean, and I say it's not new.
We have a care client and they had a young person come on board a team and then everybody else in the team left, and everybody had to build back from scratch. And the young person, first couple of months in the insurance field, broke across all the-- broke apart all the raters, figured out what underpinned them, and now he is one of the leading people at the table of how that business goes forward. That's not AI, but that's always applied, and that applies to AI concepts as well.
CHARLIE SAMOLCZYK: Yeah. I mean, you both use the term 'human'. So we've been talking about experts, and I don't know if it differs between PL and CL, but what's your definition-- I assume there's multiple, actually, but who are the experts and the humans that we're talking about that are interfacing with AI?
LAURA DODDINGTON: Yeah. So if I think about it in the context of personal lines, and particularly if you think about some of the retail products that we have like home and auto, for example, I do think most of the customer interaction, if you like, in terms of going and getting a quote, for example, should be done straight through processing, and therefore it is using AI and models behind the scenes to allow that straight through processing to happen. So where do you need the expert, then?
First of all, to determine what models and what AI you want to have. So what should that straight through processing journey look like? AI can't tell you that. It can inform it and can help you make good decisions, but you also need an expert who is part of those decisions. And then all the way through to if I need advice, for example, there may be elements of advice that AI can give, but I may also need advice from a human around some of that. I may need advice at the point of claim, for example.
And so I think there's going to be that balance around, yes, I need to digitize things. I need to straight through process things. I need to manage my expenses very carefully. And actually, customers want a lot of that as well. As a customer, I don't necessarily want to spend two hours talking to somebody to get a quote for my auto insurance. I want to be able to just do that myself, quickly, digitally, and simply. But that needs to be balanced with, and there are areas where we do need experts in building out the AI and building out the models that we're going to use and also then in interacting with the customers for certain interactions.
CHARLIE SAMOLCZYK: Do you think that there is maybe some, I don't know, the expectations are too high. As in I think AI and ML get a lot of buzz, but they are just techniques, right? They're tools, and you still need to figure out how to-- I think that's what you're saying. You still need to figure out how to apply those tools and where you're going to use them and where you're going to get value and where they don't-- where they shouldn't be applied because either they're not mature enough or you need a different kind of judgment applied that maybe only the human or the expert can provide.
LAURA DODDINGTON: Yeah. So a model is not a decision.
CHARLIE SAMOLCZYK: Yeah.
LAURA DODDINGTON: A model can inform a decision. It can help you make a better decision, but you still have to make a decision. And so I think where we sometimes go wrong is basically like, great, I've got a model. Awesome. Put it out into the market. And you have implicitly made a decision that model is good. We need to be much more explicit about, is that the model I want to put in? What will the impacts be? What will that mean for my business? How will that impact my distribution team, for example, or my call center or my claims handlers? And make those decisions explicitly rather than just assuming, AI's got this.
CHARLIE SAMOLCZYK: I feel like you could make a t-shirt, a model is not a decision.
LAURA DODDINGTON: A model is not a decision. Exactly.
CHARLIE SAMOLCZYK: And maybe on that theme. So some of the pitfalls maybe that you've seen as people are contemplating where they use these, how they use them. Are there examples where maybe things potentially went wrong with using models?
LAUREN FINNIS: Pitfalls, and maybe Laura can talk more about specific examples, but I sit in the US marketplace and it's a highly regulated market. You need to have admitted capabilities. And so we have a lot of capabilities to model risks and to apply new techniques. But if we can't break them down and explain them to a regulator, you can't be deploying them in the US market. And so model interpretation is a big challenge in the US market in particular.
And I think also the talent scarcity is a pitfall. There are only so many people that really understand these new data science applications. Am I machine learning, large language models, et cetera. But then in that population, really understanding the core concepts of insurance, it's not a simple and straightforward industry. So you have to have people who have both, and that's a small pool.
LAURA DODDINGTON: Agree. And maybe to give a couple of examples in the personal line side of things... So let's imagine you have a house on a street where there's been a lot of burglaries, for example, recently. And so you decide, OK, the responsible thing for me to do is to install a burglar alarm. That is going to reduce my level of risk. If I build a machine learning model of risk, for example, a GBM or something, it will potentially say that houses with burglar alarms are higher risk.
Why is that? Because people who install burglar alarms are more likely to live in areas where there is a high risk of burglary. So now my model is saying, OK, higher risk because you have a burglar alarm, therefore I should charge more for this risk. And if you just put that into the market without thinking about it and without an expert advising it-- it logically makes sense. There is a reason why that is what's in our data, so I put it into my pricing.
Now imagine you're the customer who's been with this insurer for 10 years and has said, hey, actually I've decided I'm going to install a burglar alarm because that is the sensible, prudent thing to do. I'm going to phone my insurer to tell them I've done that because my risk has changed and it's very important. I know I should phone my insurer and tell them. Phone my insurer. They put it through the system. AI says your price is going to go up by 15%. What?
I just did something to reduce my risk. That makes no sense. And there's a logical reason for how you got to that point, but it makes no sense in terms of the real risk. That person's risk has reduced, and you just put their price up. So maybe you might be able to manage that through your model management and your data, maybe finding some new interactions and things. But if you don't spot it, it's a real problem in the real world, and those sorts of examples happen.
CHARLIE SAMOLCZYK: I think that comes back, a little bit to your point, Lauren, about transparency. I mean, you talked about the regulator, right? So being able to really understand, I guess Laura, to your point, what's going on in the model and why things are changing, be able to explain them and apply some, I guess, some business rationale to say, is this the outcome that we really want?
LAUREN FINNIS: Yeah.
CHARLIE SAMOLCZYK: This is the Talking Tech podcast so maybe we want to shift gears a little bit into the technology side. I think sometimes it's easy to just say, oh, ML and AI. Those are two terms that you can throw out. But I guess a question to both of you, what's the tech that underpins these tools, these techniques?
LAURA DODDINGTON: So when I think about machine learning in particular, as an industry we've been building GLMs, generalized linear models, for a long time. And as we move into the world of machine learning-- and that's actually been happening for quite a long time now. This is not a brand new 2023 technique.
But then we start looking at much more automated ways of modeling, for example, and of fitting and fitting decision trees, for example, using techniques like GBMs. And so often when we're talking about things like machine learning in the context of insurance and building models, it's a lot around the modeling techniques that we are using to deploy.
LAUREN FINNIS: In the commercial line space I'd say nobody yet is really far beyond piloting. I don't see anybody in full deployment across the functions. But we see carriers using large language models on the intake process. The commercial line space is still a place where huge amount of transactions are happening via email and documents. The first entry isn't keying into a system. And so large language models are being used to translate those documents and remove the keying from the carrier process today, and then there's, similarly being used on the claims side.
So WTW has been supporting carriers for years in helping to build predictive models in the claims side to claims triage, which claims need early intervention, but they were doing that with traditional data elements. Now with large language models, you can introduce unstructured data into that. So those models can be looking through and saying, hey, which terms that are coming up in briefs are indicating that this is the claim that needs early intervention. So just introducing a whole new body of data into the work because the large language model is able to scan and pull out phrases that we weren't able to use before.
LAURA DODDINGTON: To that point around models being used in areas outside of traditional actuarial, if you like, within insurance companies, I think is really important because if you go back 10, 15 years, analytics was things that actuaries did in maybe pricing, reserving, capital modeling. But it wasn't necessarily being thought of across the entire organization. I think that's a key trend, just seeing that change across the entire organization.
So whether that is around managing my call center team, for example, and using natural language processing to assess which of my agents are struggling on the phone, for example, and when we should help and give some training to by-- in the past, do that maybe by auditing three of your calls each week. That's not an effective and efficient way to do it. And so actually there's applications, I think, across the entire organization as opposed to purely in the areas that we think of traditionally.
LAUREN FINNIS: Yeah. I'd also say in that space, one of the things that we do a lot of is because so much is evolving so fast is we do a lot of capabilities benchmarking with our carrier clients because they want to understand what is best in class right now, what are their peers doing, and how to build their roadmaps to get there. Because we see a broad swath across the industry, we're able to help carriers position, you know, where their current tech landscape is relative to peers, relative to best practice, and chart out how to get there.
CHARLIE SAMOLCZYK: Yeah, that's what I mean. It must be a real challenge, actually, for some insurers to figure out how much they want to invest, where are they-- are they falling behind? I think the other thing, Lauren, that you mentioned that was interesting was you talked about the data. Is the data that the insurers carry today, is that enough to feed where we're going with these models, or does the data landscape need to change now?
LAURA DODDINGTON: I think there's a massive amount of external data sources out there. If you're on the carrier side, people are trying to sell you data every single day. And so there is not a lack of data out there. The challenge is working out what's the data that's going to add value versus the data which is going to cost you a lot of money and not add value because it's so correlated with data you already have. For example, the data quality is not good enough.
And so I think that's more what carriers are struggling with is which data should I use and how and when and how can I optimize the value of that data. And so carriers are starting to think about, OK, what will my approach be to quickly evaluate each of these pieces of data in a methodical, repeatable way so that I can make good decisions?
LAUREN FINNIS: And I think also weighing quality versus quantity of data. Carriers have been around for many, many years and many, many acquisitions have large quantities of data that are sometimes hard to interpret because they've been layered through so many different systems. The newer carriers coming in, they're mapping more and they're realizing upfront that they want to capture everything, including client data, upfront.
That real core CRM client relationship management data, and that enables a lot. And so the war between these carriers that have big historical huge quantities of data and then the carriers that new data maybe can add third party data to augment it. But it's always been a problem to figure out the balance because it takes a lot of investment of time and money and speed of your system to bring in more data.
LAURA DODDINGTON: I think the piece that you mentioned, Charlie, around where to invest is really important. I think a lot of carriers have invested around people, which is good. So I need more data scientists, I need more actuaries great. A number of carriers have invested around data. I know I'm going to need more data to feed the engine. I don't think enough carriers have thought about how to invest in new processes, new governance, new technology that's really going to support end to end how to scale this up. So still feels sometimes like an R&D project here as opposed to something that is scalable and can be used across the entire organization.
LAUREN FINNIS: Absolutely.
CHARLIE SAMOLCZYK: And we were talking actually in a previous podcast just about the explosion of large language models and gen AI and that in that actually one of the big challenges is around the governance, the process aspects, and how are we going to use it securely, safely, with the transparency so that we I guess meet the regulatory and compliance needs and can explain what we're doing.
LAURA DODDINGTON: Well, and even if you think about a large carrier who's been around for a long time, for example, they will have a well established second and third line of defense in their risk functions, right? And those second and third line defense will understand pricing very well, probably. They'll probably understand underwriting very well. They'll understand claims very well. They are challenging each of those areas appropriately because that's their job to do to challenge to be able to understand the risk and say how are we mitigating those risks.
Often those teams may never have been in a data science department, for example, and they've had very limited exposure to AI, machine learning, and so may not be well equipped to actually even ask the right questions to understand the level of risk to understand if we are addressing that risk appropriately. I think that's a challenge in organizations because I think we're taking on a lot of risk without necessarily understanding the risk that we're taking on.
LAUREN FINNIS: I'd say that understanding across the company too. Everybody's not going to be a data scientist. Everybody's not going to know how to actually code anything. But, you need your core team that is facing the market and facing your functions to understand what the data science applications can do. And understand if there's a problem in front of me, this is something that could be deployed, and then go get the person that does understand it.
And I think that that education is happening in pieces, but I'm not seeing it happening broadly across. And at the same time, those people need to trust the data science applications. They need to trust what's happening. So again, the understanding is really key. If they have that skepticism, you're not going to get the holistic environment that you need to support really deploying that across your business.
LAURA DODDINGTON: There can't be a culture of we do data science over here, and then in this whole other part we run the business. They have to be together.
CHARLIE SAMOLCZYK: You guys are-- I mean, you're both, in your respective areas, kind of at the coalface working with clients and helping them try and tackle these problems. So maybe just talk about that a little bit. So, to be very direct, how can people work with WTW? How can we help clients? And what are we doing today to help support in this space?
LAURA DODDINGTON: As WTW, we have tools and software that can really help support in this space. And so one of the tools that Lauren and I both work a lot with is Radar, for example, which is an end to end analytics capability. So going from building machine learning models through to using that to make transparent decisions, opening up the black box so that we can use it as a business to scenario test and see what do we actually want to deploy, to deployment.
So getting those models out into the market quickly and easily and with agility. So I've made my decision - I want to deploy it today. I don't want to wait x months for me to be able to get it into my policy admin system, whatever that might look like. All the way through to then monitoring the output of that all the time so that I can see when I need to change those models and what those models are doing in real time in the actual marketplace.
So, Radar enables all of that end to end. And that's really important. But I don't think that is everything, because you can buy a software tool and give it to a few users and then walk away and think, great job done. It's really not. If you're going to transform your organization through analytics, yes, you need tools and technology, and we have that. We work with insurers day to day around that. But you also need to think about the processes that are going to go around that.
What are the skill sets we need now versus the skill sets we needed three years ago? What's the governance that we're going to have? How are we going to embrace this with a culture that's going to use analytics end to end? And so all of that holistically needs to be considered. And so we do a lot of work with carriers not just around here's some technology, but also actually let's address how you're going to transform your organization through analytics.
CHARLIE SAMOLCZYK: Yeah, so strategy work. And Lauren, you also mentioned benchmarking?
LAUREN FINNIS: Benchmarking, definitely. I think everything that we're doing is kind of about this-- if you're here at ITC, you feel it. There's just this constant frenzy about speed and keeping up, and everything we've got going with our carrier clients tends to be about speed to market through both technology and consulting.
I think with Laura's point, our people work in the technology that we sell, so we also find that we're able to help carriers think of new applications. I was talking to a customer yesterday about distribution analytics, and that's not something somebody traditionally thinks about with their core rating engine. But our software does a lot more than that. And then we also have that consulting and change management and the whole piece of it because we have a holistic practice that can support the technology implementation.
CHARLIE SAMOLCZYK: A couple themes that's pulled together there. So we've got new use cases that people are thinking about. We've got if you talk to people a year ago, some of the topics that I think-- some of the frenzy that's on the trade show floor today wouldn't have been there with the large language models. Investment is flying into this space. I think maturity will come as people get their hands on it and figure out what they're doing with it. Just maybe quickly from both of you, what do you think the future has in store in this space?
LAURA DODDINGTON: I don't think the world's going to slow down anytime soon. I was reading something yesterday that said this is the quickest pace we've ever seen in the insurance industry, and it's also the slowest it will ever be... ever again. Right?
CHARLIE SAMOLCZYK: Yeah.
LAURA DODDINGTON: And so pace of change is not going to reduce, and so therefore, I think as an industry, we have to keep getting faster and keep challenging ourselves to be faster. And that cannot be done through brute force of, great, let's just get more and more people doing it. That can only be achieved through saying, yes, we are going to use automation.
We are going to use AI. We are going to use machine learning - but we're going to use it with our experts. And so I think the future is around-- I think it will be more and more analytics across the entire organization. I think the companies who are able to make and deploy really good decisions really fast are the ones that will win in this environment. And so I think we'll see more and more of that use of analytics throughout.
LAUREN FINNIS: Yeah. I think everything Laura said is applicable to the commercial lines space. I think the other thing in the commercial lines space is we see a lot of things that are already very prominent in personal lines moving more prominently into commercial lines. I don't just say it because WTW is part of it, but we do see digital trading going up market into the mid-market and large commercial space. And as that happens-- we're working on that with our Neuron offering, but we also see that that opens a whole new bucket of analytics and a whole new amount-- there were things that were never tracked before because submissions didn't come in digitally, and there's going to be a ton of new data that can be deployed across the business.
CHARLIE SAMOLCZYK: I am glad that we have experts like you guys. Why don't we end it there? Thank you so much for taking the time. I'll let you get back to the conference, but it was a pleasure speaking with you both.
LAURA DODDINGTON: Thanks for having me.
LAUREN FINNIS: Thank you.
SPEAKER: Thank you for joining us for this WTW podcast featuring the latest perspectives on the intersection of people, capital, and risk. For more information, visit the Insights section of wtwco.com. This podcast is for general discussion and/or information only. It is not intended to be relied upon, and action based on or in connection with anything contained herein should not be taken without first obtaining specific advice from a suitably qualified professional.
