WTW & Anthropic: Beyond insurance AI hype. A Talking Tech special

CHARLIE SAMOLCZYK: Welcome to another episode of Talking Technology. I'm your host, Charlie Samolczyk. And in today's episode, we are going to be talking about how artificial intelligence is no longer a theoretical conversation in insurance. It's here, it's moving fast, and it's forcing leaders to make real decisions under real constraints. We're going to go beyond the hype today. We're going to explore what AI actually looks like in practice inside, probably one of the world's most regulated industries, and what it will take to unlock value responsibly.

I am thrilled to be joined by André from Anthropic. André is going to share some of his insights and talk about Anthropic's mission, how they've evolved some of their models, and what responsible AI means when systems move from labs to production. And also pleased to be joined by Pardeep Bassi from our team, who leads our Data Science and AI proposition. And he's going to bring a practitioner's perspective on how AI is already reshaping insurance and look into things like underwriting and pricing claims, governance, and operating models. So, thank you both for joining.

ANDRÉ BALLEYGUIER: Thank you, Charlie.

CHARLIE SAMOLCZYK: André, maybe we could start with you. Again, as I said, thank you so much for spending the time with us. We're thrilled to have you here. I think it's probably worthwhile if you want to just take a couple of minutes and introduce yourself and maybe talk a little bit about Anthropic's mission and the areas that you're currently focused on.

ANDRÉ BALLEYGUIER: Yes, definitely. Thank you very much, Charlie, for hosting me. Yeah. So my role at Anthropic, I lead the applied AI team across the EMEA markets. So, my team is the technical experts who work with our customers end-to-end. Those are solution architects, applied AI engineers, as we call them. The idea is that from the first conversation up through to building a production agency systems, our team will be helping customers understand, based on their requirements, what they need to do with our technology to deploy these systems at scale.

So, my background is in machine learning and AI. So, I've worked in AI now for, I think, 13 years with one of the very first data scientists in the London market before I transitioned to work in AI GTM. So, at software vendors who sell AI technologies. So, I worked in two companies before Anthropic in this space. The first one, I spent around eight years there. I built the EMEA team of a customer-facing data scientist before taking other leadership roles. And then also, I spent two years at Snorkel AI, a company focused on AI data.

But throughout this journey, I actually had a lot of interactions with insurance companies and the insurance sector in particular. That's where I met Pardeep. I think that was almost 10 years ago now. And working with insurers, I've actually seen the journey of insurance companies on how they moved away or they moved from statistical models to then use more sophisticated techniques using AI and machine learning to then today use also LLMs to potentially modernize some of their customer interactions, as well as systems for claims and underwriting.

So, just a quick note about Anthropic. So I joined, I think now, three months ago, which in the Anthropic days seems to be an eternity because we have a lot of product releases and a lot of things happening every week. It's almost hard to keep up even when you work at Anthropic. But for the audience not fully familiar with Anthropic, we are a company focused on AI safety. So we built Claude. It's one of the frontier LLMs models.

The reason the company exists is, as the technology started to develop, it's proven that the pace of how fast the technology evolves could potentially make the technology potentially one of the most consequential in history. But a very powerful technology also means a lot of risk. And so it's very important to actually develop a safety-first approach on-- adopt a safety-first approach in developing this type of technology. So we are really thinking that we need to understand the risks of using and deploying this technology so that we can then put in place the right guardrails and help society prepare to the potential impacts of deploying AI.

CHARLIE SAMOLCZYK: Fabulous. And I think over the next half hour or so, we'll definitely get into the AI journey and probably get into the risks and understand how Anthropic is looking at the safety-first kind of approach.

ANDRÉ BALLEYGUIER: Yeah.

CHARLIE SAMOLCZYK: But before we do that, maybe, Pardeep, do you want to just briefly introduce yourself and maybe talk a little bit about your role and how AI fits into your work and what you're doing in insurance today?

PARDEEP BASSI: Great. Hi, everyone. I think, André, I first met you 10 years ago. I think it was at a conference, a very long time ago. So, I'm the Global Proposition Leader for Data Science at Willis Towers Watson within Insurance, Consulting and Technology. So, a large part of my role is working with insurers to really understand where and more importantly, how to apply AI to extract the maximum business value that it can do.

So, everything from individual point solutions to large, AI transformation-type projects. But I also spend a good amount of my time enhancing our technology. So that's making our technology usable by AI, but also embedding insurance-specific expertise into the design of AI agents specifically within our Radar Suite. So, yeah, my role is quite varied, and I think AI is going to have a huge, transformative impact. And I'm very keen to have this discussion with André to cover both from an insurance specific and a cross-industry perspective of how we think things will progress and the value that it should extract.

CHARLIE SAMOLCZYK: Yeah, very good. Maybe, and hopefully, an easy question to kick us off. If you think back with both of your journeys into AI, was there a pivotal moment for you when you realized, wow, AI is going to fundamentally change, well, insurance or just the industry that you're working in?

ANDRÉ BALLEYGUIER: Yeah. I'd say there are definitely a few moments. I'll describe one in the insurance sector in particular. I mean, the reality is that the technology keeps surprising me as it evolves. So, I actually take an example before the LLM era, so that was a pre-2022, but I think it was seven or eight years ago working with one of the largest UK insurers when I-- a P&C insurance company.

So, I was working with a team of pricing actuaries. Their main goal was to actually develop models that were more sophisticated than the current models of the firm to capture weaker signals to understand, what would be the risk that the insurance would take underwriting some of that? In that case, we were primarily focused on motor insurance to start with.

This was, I think, a team of 7 to 10 people, quite a sophisticated, I'd say, statisticians, data scientists at that time. And they were using a statistical software such as R and Python to develop those models. And using automation, we were able to actually, very quickly, get to more accurate models in a way where they were able to iterate much faster towards the results and release by delegating more to the machine.

In that case, it was about testing more hypothesis, testing more models, and understanding, what would be the hypothesis and/or the signals that would be most relevant for the portfolio. So, this trial and error to find the best model was really something where I could see that there could be automation in this space. And automation, as it gets more and deeper and deeper, broader and broader potentially could make the work of an insurance company, in particular, when you think about underwriting or reserving, where actually a lot of it is about manipulating data and using it in the best way, much more optimized and efficient.

Because the reality is that finding the signals in the data could potentially mean being able to take less risk or better measure the risk for customers, an potentially ensure that you have a more accurate price. And in that case, for an insurance company, it's literally a competitive advantage. And so ever since, now it's with agents. There's the same realization, but just orders of magnitude bigger, right?

CHARLIE SAMOLCZYK: Yeah.

ANDRÉ BALLEYGUIER: It goes beyond just statistical models for underwriting. It's really using automation across all the lifecycle of the insurance. And this is much more pervasive than it used to be. Yeah.

CHARLIE SAMOLCZYK: Yeah. Very good - chasing the signal, Pardeep.

ANDRÉ BALLEYGUIER: Yeah.

PARDEEP BASSI: So, my experience was actually quite similar. I didn't expect André to say that. So I built a GBM model for trying to predict AD frequency in the pricing space. And it wasn't with André and his team. It was independently. And I saw a massive improvement in terms of the predictive power with significantly less data. So, that was my first more of a machine learning aha-type moment. But where I really saw Gen AI add significant value was when we were developing Radar Vision, which is our model monitoring and business outcome monitoring tool, where we apply Gen AI as an enrichment over the top of the deterministic SaaS tooling that we've got.

And I expected it to be really good at summarizing data, drawing insights across multiple time periods, more of an automation acceleration of insight. We'd already understood. But what I found was, it was providing insight that I didn't have before. So, I was able to draw conclusions across multiple models, consider multiple things at the same time. And the one particular example which really brought this to life was, we were looking at the performance of a particular model, again, claims frequency from a pricing setting, and the AI was able to pick out a seasonality effect.

So, it was able to determine that electric vehicles, which accelerate quicker than ICE vehicles, in winter months in certain parts of the country where it rained more, was able to pick that effect up and explain that and gave that extra insight that was potentially missing or more difficult to understand as a practitioner. So, genuine value add in a specific insurance and pricing setting was my aha moment with Gen AI.

ANDRÉ BALLEYGUIER: And I would say, today, as beyond the insurance sector as well, some of our latest models that we've ruled out, finding those patterns, being able to almost correlate those different patterns together is a task that, when you think about it, machines will always be faster and better than humans, as long as you can reliably uncover those patterns in the first place. You've probably heard of our project called Project Glasswing where we work actually with a select number of companies to roll out the latest AI models more safely.

And because they are able to identify those patterns faster. And also, we could potentially even mix and match different patterns, that creates almost a cybersecurity risk of using this technology. Unless the cybersecurity companies are making sure the software are safe using the same technology in the first place. And that will, I believe, come more and more across all of the industries. Cybersecurity is probably one that's a bit more controlled, a bit more narrow in terms of what is the art of the possible, and the patterns are more repeatable. In insurance, it's probably more complex or maybe not. Who am I to judge that? But I do believe that those patterns will be found faster as well, and better, and more and more autonomously by the machine as AI technology improves. Yeah.

PARDEEP BASSI: Was that your mythos model where you were able to--

ANDRÉ BALLEYGUIER: Yeah, correct, yeah.

PARDEEP BASSI: --link separate vulnerabilities and potentially identify bigger risks?

ANDRÉ BALLEYGUIER: Yeah, that's correct. Yeah, that's the model mythos that we only used as a preview to some of-- a select number of cybersecurity companies for them to ensure that the software that is right now fueling the world's technology is safe, because those models are now able to identify and uncover those patterns that a cybersecurity expert could potentially identify too using more a mix of rule-based and AI approach with, let's say, less powerful models. But would take them potentially years and years just to uncover one or two. In this case, within a few days, you can already find much more. Yeah.

PARDEEP BASSI: I think the speed of development is quite frightening. Because when we first started developing that AI enhancement to Radar Vision, we had a less powerful model. When the newer models came out, we were shocked as a team in terms of the insight that it was giving that wasn't available before - so that's only going to increase. I think building these solutions with the future development pace in mind is equally as important as trying to understand where the value is right now.

ANDRÉ BALLEYGUIER: Exactly. And that's what still surprises me today. When I compare 10 years ago, where it was the early days of using some of the most sophisticated machine learning techniques, where you delegate some of the pattern recognition to the machine, things that were typically done by an actuary or statistician more manually, that's really-- this ability of detecting patterns is something that the machine is more and more reliably and more and more autonomously able to do at scale. And that's the big difference. And that's what I think matters a lot for the insurance sector, which is all about understanding patterns and the data in a way that makes a measurable, more precise. Yeah.

CHARLIE SAMOLCZYK: Well, let's dig into that. So, I mean, you've been talking about actually real world applications and how it's being used, not just the hype, because I think the applications are significant and material. Given insurance is one of the most regulated industries in the world, maybe a discussion a little bit around how AI is really being deployed today.

PARDEEP BASSI: I think it's fair to say that there's a lot of hype, and there's a lot of insurers who are stuck in pilots. So actual production-ready solutions are relatively limited at this stage, but I think they're increasing quickly. So I think one of the biggest things to think about is, where is the tangible business value rather than a nice marketing message in terms of, we've got a Gen AI solution and we can go to our board and say, look, we are AI ready. We're implementing AI. So I think a really in-depth understanding, specific to insurance, where AI adds genuine value. And we've got a framework in WTW, which I'll take you through very quickly, which we've been using to help distill the problem.

So we think that there's this four key stages in terms of AI adding genuine value. The first stage is all around automation. So, we think that automation will become easier, better, and faster with the use of AI. So that's existing actuarial workflows, which can be enhanced with the use of AI. And that allows you to use extra data, go deeper in your analysis. And the second step out of the four is the predict stage, where you're still leveraging deterministic machine learning models as tools for your AI. But you're able to explain and interpret what's going on better and quicker. So that allows you to understand why your models are making the predictions that they are.

And then the third stage is what we call the suggest stage. So, this is where we're seeing the AI actually genuinely give you new recommendations, give you suggestions in terms of pricing actions to take based off of the information it has. So, it moves from a tool that helps you enhance the predictability, give you a deeper understanding, and help automate tasks to actually give you suggestions before moving on to actually taking actions.

So, the end state is, with the right controls, you'll have multiple AI agents as part of your analytics team working together, communicating, sharing insight, and helping you make better decisions. But the human in the loop is essential, especially in insurance, where accountability will always remain with the human. You may have less humans and more AI agents in the future, but the accountability will need to remain with an individual, an individual human, who then needs to have the confidence that the system is acting in the right way, with the right safety, and the right controls that they themselves can be accountable for the decisions and actions are being made.

CHARLIE SAMOLCZYK: And, Pardeep, maybe just to your opening bit there saying that actually, there's still a lot of hype, and things are getting stuck in pilots. Where in that four-stage model are things getting stuck?

PARDEEP BASSI: So, I think automation is relatively easy and simple. So, extracting, translating. I think on the prediction side, I think there's a slight misunderstanding that machine learning is less applicable, less useful than it's not. So, I think the machine learning phase, as part of the fit predict component, is misunderstood in terms of how the Gen AI, the nondeterministic nature of that works with your machine learning models to give you that really strong, deterministic, repeatable element that you require.

So I think it's the predict stage where people start to become slightly unstuck because they want to move to Gen AI and forget about machine learning when you need that really strong foundation of machine learning and underlying data strategy as a business, and then the suggest stage is where it's very much a chatbot interface type setting where insurers are beginning to understand and realize that that will be a supplementary component to a human making decisions.

So, I think the human plus AI model in the suggest phase is something which needs to be developed, both in terms of understanding where the AI adds value, where the human applies judgment, and where you have the right control. So, human plus AI operating model is where people are getting stuck in the suggest stage. I don't think many are in the act stage at the moment, but I think the wider business transformation required to embed a new operating model to effectively blend together human and AI agent capabilities is rather difficult.

ANDRÉ BALLEYGUIER: From our perspective, we look at value delivered to our customers by insurance and elsewhere across three main pillars. One, is make employees smarter. And for that, we provide platforms and tools for personal productivity of employees. One we've been quite successful at is the software engineer and developer productivity, replacing a lot of the manual tasks in the software engineering development life cycle to more AI-generated that indeed, it's very important, throughout the rollout of the technology, that we keep it safe and we keep the human in the loop.

So that's why we have strong enforcement of permission systems, as well as the user inputs in the parts that make the most sense in the workflow. And it can be controlled from a governance standpoint, but also at the individual level. And so we do that both for software engineers, but also more broadly, particularly with tools such as Cowork, across all the knowledge workers who could think potentially pricing analysts and/or underwriters to then use-- because a lot of them, their work is actually manipulating data and making decisions on this data.

So, there are a lot of parts of the workflow that are, let's say, a bit more mundane, a bit more repeatable. And so those ones that were-- that's where you can trust the machine, as long as there's some level of oversight. But then there are parts where there are more complex decisions where actually to make the decision, the AI can really help generate useful summaries and/or insights that they can then move to the next stage of their reasoning. And so for that, we definitely-- it's about making employees smarter. That's the first pillar.

Second one is more optimizing processes. So, taking for example, an underwriting or claims triage process, modernizing it with AI. And so there are parts of the workflow that can potentially be based on additional inputs, or data that was not accessible before, or that can be-- decisions that can be optimized as well, automated. So, we do see initial adoption of those, but it's still in the early days, I would say, in the industries we have.

Even how companies need to deliver this is still to be determined in terms of, there are various patterns. We see where companies are better. For example, embedding quite expert team within the business lines to do it and forward deployed way instead of trying to do it in a centralized way. That's something we've seen quite well, because there is such an importance of involving the business and compliance teams as early as possible in the workflow and as deeply as possible in the development cycle.

And the third pillar is really making products better using AI. So, embedding AI systems, in that case, it would be Claude APIs and Claude models within products, that are sometimes often customer facing. There are a number of tech companies that already do this today, and we can start seeing some titles as well in the insurance sector. So, this is the third pillar for us. Yeah.

I'd say the third pillar is very common in tech companies that are basing their entire technology on AI. First pillar is very common in large organizations. And second pillar is still in its early days, but it's something we are seeing more appetite for across our insurance customers, across underwriting, claims, compliance, and others.

CHARLIE SAMOLCZYK: You touched on talent and roles there. What's your experience right now around how AI, I guess, is affecting those roles? It sounds like it's making them more powerful.

ANDRÉ BALLEYGUIER: Yeah.

CHARLIE SAMOLCZYK: I don't know if, going back to the hype bit, I don't know if there's anxiety around how our roles change and what roles we need in the future. But in practice, what are you seeing there around how it's maybe affecting roles and companies, but also how it's affecting, I guess, the talent in the market overall?

ANDRÉ BALLEYGUIER: That really is an economic index related to this. And the latest is that today, almost 60% of the use of AI is more for augmentation rather than necessarily automation or replacement. There's a large increase of the number of occupations that can be impacted by AI. So I think we saw, within one quarter, a 30% increase of the number of occupations that can make use of AI as part of their workflows, either for augmentation, in that case, for automation. And also for each occupation that can be impacted or affected by AI, there is also a large increase in the number of tasks that can be delegated to the AI for, again, augmentation automation. And that's because the model capabilities are increasing.

We see every seven months or eight months, there is a curve that tracks that that has been quite accurate to predict how this technology is going to improve over the years. But every seven months, the duration of the task that can be delegated to the AI system, where the AI system would be better performing than the top human experts, every seven months, this duration is doubling. So, if today we have-- if you can delegate a task of potentially a few hours reliably to the AI system, it means that in a year, well, this duration of a few hours, may be multiplied by 3 or 4, if we follow the same curve.

PARDEEP BASSI: Yeah. And just to add to what André said, so my view is that deep industry domain expertise becomes more important longer term as you're able to pass tasks over to the AI to automate and perform, but also the ability to effectively embed the institutional knowledge in the real experience, say, for example, underwriters into the AI, which will then help the more junior members who start to effectively have that 20 years of experience at their fingertips, guiding, suggesting, and improving their output.

So, I think New entrants to the workforce will be enhanced in terms of their impact on the business, because they've got access to AI, which has been trained based off of all the institutional knowledge. But the manual, repetitive, generic type roles will become less and less commonplace, because I can perform those better, faster and at lower cost. Whereas the really deep insurance expertise will always remain because you need to design the AI agent and then monitor it to ensure that it's doing what you want it to and effectively set the guardrails and controls and apply judgment. So I think application of judgment on top of AI will become incredibly important, and it may be that you need significantly less people to do that. But the importance of that becomes larger over time.

CHARLIE SAMOLCZYK: There's a governance kind of-- well, I was going to say management, but actually maybe a leadership aspect to this. So I suppose with it moving so quickly, there is a challenge to organizations to, how do leaders make sure that this lands in the organization? And I guess that we cater for the people that are the practitioners, but I guess-- and we'll get on maybe the next topic, André, we can talk about is responsibility and what Anthropic's doing there. But just on the leadership bit, are there things that you're seeing that companies are-- either pitfalls or things that they're doing well from a leadership perspective around AI?

ANDRÉ BALLEYGUIER: Yeah. And I second Pardeep on indeed the fact that I would say what AI will not replace is knowing what high quality or good looks like. In order to use AI in the best way, you need to have this sense of knowing what good looks like. So, you need to have accountability. For example, you need to have taste. You need to have customer relationship, depending on your sector.

So, being able to, let's say, judging the output is something that will be always critical in the AI space if you want to use the technology in the right way. Otherwise, you won't be competitive, and that's where indeed the shift towards deeper experts being critical rather than just having generalized only is definitely something I'm seeing. But they'll definitely be a squeeze on maybe entry level jobs that were-- where people don't have this sense yet.

So, I think it's important to look at, for the leader perspective, a practitioner should know when to trust it, should when to check the model, should know what evaluations to build before you actually roll out the AI. So, the evaluation is the system used to control indeed the output of AI system and/or some of the decision makers. And so that's from a practitioner perspective.

From a leader perspective, I think it's all about having the right teams in place for redesigning an entire process around the model. So here, the idea is not to necessarily redesign the model around the process, but it's really having an approach where you think about the AI system as a system that needs to be deployed in production from the early days. And so it's about bringing all teams, such as the compliance teams first, the business owners into the room from day one. Because if you forget one of those teams, potentially, you won't be able to go to production. You can't have the system in production, in particular in a regulated industry, without putting compliance first.

I've seen way too many pilots, and that's one of the reasons why some of the pilots don't go to production, where after six months, they had not been or very limited engagement from the compliance teams. But those teams need to be at the forefront and need to be working hand in hand with the business and alongside some of the experts, engineers who are very aware of the latest and greatest technologies in the space, knowing that the technology evolves a lot.

So, bringing people in the room, and then having almost an iteration cycle because this is almost-- some of it can maybe R&D. These problems that have never been solved before. And stuff like putting on a pedestal some of those demos that are nice-looking demos, but that will never make it to production, really make it more concrete with the right stakeholders in the room and also measure adoption and cycle time of development. All of those things are things that need to be rethought in the era of AI and analytics. And that's some of, let's say, guidance I would bring to companies or my teams and other teams who work with customers and deploying at scale how to, let's say, put in place to make sure that these models get deployed. Yeah.

PARDEEP BASSI: Yeah. And there's a lot of commonalities, André, with the machine learning productionizing phase that large organizations went through where the effort wasn't in terms of understanding the technology or building a predictive model. It was more wider business transformation-type effort where you had to consider security, business operating models, business preferences, and then people in terms of ensuring it's adopted and used.

And I think with AI, it's a similar approach, but it's just quicker and even more disruptive. So I think the operating model that leaders need to think about is essential to effectively unlock that value. Because the operating models in the past have always been built around relatively siloed functions and people management. With AI, one of the really big things that we're going to see is, the people management structures become less important because you have agents between functions sharing information, breaking down silos, working towards common goals.

So,f think defining and operating model and then a path to slowly transition to that view, that end state, that vision, I think that's what AI leader will need to effectively deliver. And that will be the differentiation between those who talk about AI from a theoretical perspective and pilots and those who actually embed it into organizations, ensure it's used, and effectively ensure that the organization is significantly more competitive as a result with direct impact on business outcomes.

CHARLIE SAMOLCZYK: Can we talk a little bit about the responsibility of those that are using the tools? And I suppose, and I know this is near and dear to Anthropic's mission of looking at safety and societal impact, there's trust in there as well. And I know we talked about trust before. I think there's probably trust of the technology and the models itself, but also trust of the companies that are using it and deploying it. Maybe I'll just leave that as an open question.

ANDRÉ BALLEYGUIER: Yeah. I mean, trust is really like-- in its essence, it's one of the reasons why I joined Anthropic. And then as Pardeep mentioned, is, I saw, working so long for regulated industries, primarily banks and insurance companies, but also health care, how much trust could potentially-- was actually the main reason why companies would adopt at larger scale these type of technologies. Because as you delegate tasks and/or augment humans with AI technologies, you are potentially trusting a system to actually make decisions or take actions instead of yourself.

And so our view is that intelligence alone will not make an agent deployable in a regulated business. What will make it deployable is trust. And trust, in its essence, the reason why trust is even more important for LLMs, I'll come to this. But to solve the trust problem, I think it's mostly a governance and engineering problem rather than just a model intelligence problem.

So one characteristics of large language models is that by nature, their reasoning is a nondeterministic. So if you ask it something once and then you start to change just a little bit conditions, or even if you don't change some of the conditions input parameters, you ask it again, it may be a different answer. And so this nondeterminism is what, I would say, regulators and compliance teams are afraid of.

Because you may have repeated this output many, many times, but what tells me that it may not say something completely wrong and if something completely wrong could have potentially a very high risk? Even if it's a very small risk, if the risk and impact is very, very high, that needs to be taken into account. So model risk, because of this nondeterminism, is higher with LLMs than other systems.

But also because those agents in particular, as we get agents more and more pervasive and capable, we give them access to more tools like more systems. And each of those systems, each of those tools, they are permissions associated. There is some data that should not be accessed by some teams. And all of this becomes an engineering problem. So how do you make an agent using a system? They use it in the same way a human would.

The reality is that AI has become smarter also sometimes trying to find ways with other tools that a human would probably not naturally use, but they may actually try to use other tools to find their answer, even by maybe getting access to data that are not supposed to, if they can. And that's where this idea of creating secured system and secured agents, you also need to think probably more deeply around permissions and security and make sure it's enforced in the enterprise so that your agents don't start accessing data they're not supposed to make decisions.

The final point might be related to that, and that probably lacks a lot of other angles in the enterprise, is accountability. So when a decision is made, and if there is no human between the model, and let's say an agent that has made a decision to write the policy to a customer, who would be accountable if something goes wrong or if the policy is not adapted? So there needs to be a degree of accountability, and that's one of the reasons why having a human in the loop, in the parts of the workflow that makes the most sense for accountability, is still definitely required. In addition to making sure that the outputs are expected, it's about accountability.

And so there are many tools and techniques that you can use. And as I said, it's about governance. That's why compliance in the room from the get go is very important, but also about technology. And that's why it's important to have quite advanced technologists as part of the process. Even if some of those tools become easier and easier to use, if you want to have a pervasive and high-value agent, you will need to think about those considerations from the get go if you want to get trust.

And one thing I'll mention, just as a story, an anecdote probably, because that happened to me and to my team in the past. We had one project with an agency in US government where the model was predicting one prediction every day. This prediction was used to actually make budget decisions. We had almost a confidence interval around this prediction, so we were pretty sure when we're making predictions, it would fall within that interval.

And the reality is that one day, this number was really, really wrong for a reason where-- it was based on some, let's say, elements in the model that made the prediction be out of the norm. But because it was so out of the norm, it was noticed by someone, in that case, the governor within the US federal agency worked with. And even though we tried to save it, salvage, it's like two months later, this model that was actually quite high value and useful was decommissioned. Because that moment where this happened, where the model failed and didn't produce an output that was expected or was so different from the expectation, broke trust. It's much easier to lose trust than it is to recover it. So trust needs to be embedded from the initial design of the model if you want to go to production and if you want to scale.

CHARLIE SAMOLCZYK: So it certainly sounds like there's that kind of burden of creating trust. And you're right, André, it can be lost so quickly. Maybe, Pardeep, what are the implications of responsibly building technology and offering consultancy services knowing that we need to accommodate for that for this high trust level and this almost societal responsibility in what we're building?

PARDEEP BASSI: So I think André made two points that really resonated. The first was around the need for determinism and that balance between extracting value from the nondeterministic nature of LLMs, but that having that repeatable output that you require in a regulated industry. So what we're doing is, we are very quickly adapting our tools to be used by AI as well as humans. So that provides the deterministic call to the AI agents to ensure that repeatable, robust components are calculated as part of actuarial workflows.

But we're also changing the way our tools are used. So increased accessibility and usability, whether it's the free text prompt that you type into Radar Vision to understand why your model has drifted. So it's that human plus AI balance where you effectively leverage the nondeterministic nature of LLMs to ask it a broad question. It can go and find extra data, draw conclusions across multiple models, multiple assessments. But it's tied to that deterministic set of foundational calculations.

So as a human, you can understand exactly what the AI has used to form its view, suggestion. But you can actually, as a human, also go in and drill down manually into the exact deterministic foundational core which has been used. So blending that human plus AI capability together is a large part of what we're doing in terms of taking our software forward.

The third thing which we're doing is probably the most exciting thing, is we're building insurance specific AI agents into our software. So we're embedding our years of deep domain expertise, which is typically held within our consulting function, who often engage with insurers to set best practice underwriting, pricing, reserving processes, we're embedding that into the design of AI agents.

So in effect, when we deliver software out to market to insurers, not only is it easier to use, it's usable by AI, but it also has WTW expertise embedded into it to ensure that you're getting the best possible outcome. So you will have one of our experts feeding in their years of experience. So the software will be able to suggest and make recommendations based off of that context and that insight that it's been trained on.

CHARLIE SAMOLCZYK: We are quickly running out of time, and I'm sure we could talk about this for a lot longer. Maybe a bit of just looking towards the future a little bit. Where do you think the big opportunities are? And I guess married with that, where are the big risks and what should people be thinking about over the next couple of seven month cycles as they're thinking about their business?

ANDRÉ BALLEYGUIER: Yeah. I mean, in terms of areas, I think across the insurance lifecycle, underwriting, claims, product developments, understanding markets and market structure, all of those will be modernized and augmented with AI for sure. I can see your words, and it's hard to project that far off, as you said, 2030, because technology is evolving so fast. And right now, this curve I've mentioned is-- this exponential curve, we're even going a bit faster based on the latest model outputs.

So on the right thing, I can see every potential underwriting and decisions, maybe triaged, and enriched prepriced even before a human opens it. And it's happened already in some of, let's say, parts of insurance and parts in P&C, but that I can see it happening more pervasively across all insurance segments and also all sides of insurance companies as well.

In claims, there'll be more straight through processing of claims for sure. Adjustments and understanding where there's fraud potentially or when there's complexity will also be definitely triaged by agents in a way that's much more accurate than today. In products and launching new products in insurance, that is something that there's a lot of work across all functions, like risk, and marketing and design and products. All this product development, potentially, you can see the cycle of product development also on compliance obviously and regulation being much faster than before. Because that's all about getting process at the end of the day.

And some of the niche products that were not accessible in the past, potentially products that are more based on, let's say, seasonality, or more products that are very specific to niche markets, they may be accessible if they were not accessible in the past, just because of the lift required to launch a product that would be lower. And so, yeah, and then also across markets structure and understanding there's a lot that can be done.

But as you mentioned, there are risks. I can see, I think, the risk we mentioned some of them. But trust and failing to actually get trust is a real risk for the industry. So having human in the loop, audit trails when any of these agents are making decisions or are triaging case, will need to be actually implemented. Because there'll be cases for sure where the agents will not make the right decision that will have actual consequences for the consumer. And those consequences could potentially-- could be a missile policy, could be an execution that's hallucinated. That will set precedence.

And how these are handled, how the system was thought in the right way initially will potentially reduce the consequences of these adverse events. That will definitely happen. And the are other risks would be more societal. So I can see the workforce will be transitioning. So in particular, we talked about entry level jobs, where you have a lot of analysts working with maybe one or two senior people. Definitely, there'll be less need for those.

And so how do we make sure that these people who were educated, let's say, in actuarial science or in some of the functions that are entry-level jobs in the insurance sector are potentially reskilled in a way where we can potentially manage the effects, at least the short term effect of it. I'm confident in the ability of humans to adapt in the medium term, but there is always this transition phase where there'll be a squeeze on some of the markets and some of the skills that are currently or that used to be sought after and will be less sought after because AI will be more capable there.

And so that is definitely something we need to think about as a society. And universities need to work with AI companies quite deeply, I think, to make sure that programs get updated. Because there'll be more need for using and leveraging this technology and developing it. Definitely, less need for people who actually do those tasks day to day and are just learning about that.

And I think it's really a risk for us and for the industry to just stand still and do nothing and not consider the risks carefully. So it's also a risk for an insurance company not to use the technology, because the competition will be more and more fierce as some of those processes get augmented and/or automated. So it means more accurate pricing. It means potentially, if you are not adopting this technology, you will have the adverse effect and getting into your portfolio all the customers that you may have misplaced.

So that means that eventually your portfolio will be deteriorating, and then you could be losing competitiveness and/or losing margins. This is something where I do think not doing anything should not be an option for an insurance company that want to look at the future. And because this technology is so transformative that's not adopting it is a bigger risk than actually making the investment now. Yeah.

PARDEEP BASSI: Great. And then just three very quick closing remarks from myself to summarize, I think. We have a term in WTW called PPCU, Pricing, Product, Claims, and Underwriting. That's where we see insurers really differentiate themselves in terms of being able to compete and drive tangible growth and profitability. So application of AI in pricing, product, claims, and underwriting would be where I would focus.

The second thing is, I would think about the problem in two ways, both enhancing existing processes, making them better, quicker, faster, but also thinking about new things that you couldn't have done before. So AI is unlocking new capability. And if you're one of the first to understand and adopt those capabilities, that could be a huge differentiator.

And then the third and final point, none of this is material if you don't actually put it into production. So think about productionizing the capability and having the right operating model to ensure that you've got the right controls and set yourself up for short-term value extraction, but also an operating model that allows you to leverage the ever increasing advancements in the LLM frontier models.

CHARLIE SAMOLCZYK: Pretty clear message. Gentlemen, thank you. I think we can wrap it up there. André, really appreciate you taking the time to join us. Thank you very much.

ANDRÉ BALLEYGUIER: Thank you. Thank you for this great conversation. Thank you, Pardeep. Nice to see you again. And thank you, Charlie.

SPEAKER 1: 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, 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.