How Agentic AI is reshaping insurers’ core technical functions

So today we're gonna go beyond the hype. We're gonna focus on how AI is impacting the insurers technical functions, things like pricing, reserving, portfolio management. We'll explore where agentic AI can play a meaningful role, not just in automating tasks, but in changing control cycles and decision making and the way that technical expertise is applied. And why ultimately the hardest part of this shift isn't necessarily the technology, but it's the people and the processes and the operating models that need to be in place to leverage it. Today I'm joined by Duncan Anderson who leads WTW's Insurance Technology Practice. Duncan, thank you very much for joining me today.

Duncan Anderson: Pleasure.

Charlie Samolczyk: Before we get into agentic AI, I think I want to ground perhaps this topic of AI in your journey, because you've probably seen more change in this space than most. Can you talk us a bit through your background, what you've seen over the last 35 years in the P &C industry, and how big a disruptor do you really think agentic AI is?

Duncan Anderson: Yeah, sure. Now, my current role, and actually has been for seven or eight years, is to lead our insurance technology business. But I'm not actually a technology professional. I'm not a software engineer. I'm an actuary by profession. And a large amount of my career has been focused around personal lines, pricing, and related analytics. So that's my background. I've been with the organization in one form or another for just over 35 years now. So seeing some changes in that time. mean, it's hard to believe it, but when I started in the building I'm sitting in right now, we didn't have a PC on each desk. So we did a lot of work by paper. That's how far back my career goes. So I've seen quite a few changes over the time, like the introduction of PCs, like the internet, for example. But in the world of pricing and personalized pricing, the introduction of statistical modeling in the 90s, the focus on demand. modeling and optimization techniques in the 2000s, the introduction of machine learning, which is technically a form of AI, maybe around the 2010s, a lot of the automation that then followed. So lots of changes over that time, big changes in distributions from direct writing to aggregators and the like. But I do think that the changes from generative AI and in particular, agentic AI, will be the biggest changes I've seen in my career for.

I think the insurance industry, but for the technical functions in particular. So think this is a really, really fundamental, wide-reaching change that I think probably will be bigger than anything I've seen in my career.

Charlie Samolczyk: We've, we've, I mean, obviously we've had a lot of discussion around about AI in insurance more broadly. And there's a lot of discussion around kind of efficiency and productivity and coding support and the impact on the talent pool you know, a lot of that conversation focuses kind of on faster development or lower costs or, you know, doing the same work with maybe fewer people. But I think today we want to go a level deeper into kind of the technical function in insurance. So talk about pricing and reserving and portfolio management and the mechanisms that actually determine performance, capital outcomes or your competitiveness in the market. So maybe the first question, how do you think GEN.AI and guess agentic.ai will affect the way that insurers actually run their technical functions?

Duncan Anderson: I think there's probably two questions there, one about generative, one about agentic, and I think there's probably different answers there. But I guess the first thing to say is the technical function is quite a lot of different functions, quite a lot of different disciplines, quite a lot of different control cycles. But I think one overarching observation is that over time, over the last decades, these sort of cycles have existed in one form another, but what has happened is that the analytics that's undertaken in each of them got more sophisticated. There's been regulatory change, which has changed the nature of and increased the burden on what has to be done in each of the steps. There have been technology changes, which have enabled the different techniques, different approaches, more automation. But fundamentally, over all those years, the people-based structures have pretty much stayed the same. And I think one of the things that will change with agentic AI is that those people-based structures, what people do, will maybe change a bit more than has happened in the past. Probably better to sort of illustrate it with some examples. So there lots of technical functions as pricing, portfolio management, reserving capital, operational claims. Given my background on personalized pricing, let me start with pricing and portfolio management perhaps, and maybe just to start with portfolio management as an example. So, know, first thing to say is that AI has been used for years in the form of predictive modeling and machine learning. So that's been around for a decade or two. So AI has been used in that sense to create predictive models, to predict things. I think what is changing with generative AI is that we're now getting into a richer sort of functionality where generative AI actually suggests things, it recommends things, it makes technical analytics much more accessible to a wider audience.

So, you that's happening now. So just as example, one of our software offerings we introduced a year or two ago is a portfolio management monitoring tool. It automatically ingests emerging data on a daily, weekly basis. It automatically runs a barrage of tests to see if there are segments of the business which are performing differently to that which is expected. So by monitoring sort of model.

Duncan Anderson: Productivity or looking at underwriting factors to see if there are new segments that are deviating in a way we didn't want to see if we're writing more business and the like. So we produced an automated tool that surface these and if you'd like in a quote traditional way using machine learning techniques, what we've been able to do more recently is put a generative AI layer on that, which enables a broader audience essentially to converse with the tool and ask it plain English questions like you know, tell me if I've written more business in particular segments, tell me if there's a segment where I've written more business and the loss ratio seems to be deteriorating, what are the five segments that I should be most worried about, which rating factors seem to be being less predictive than I thought given my claims model. So these English language questions can now be asked in an English language response and provided back, which makes it much more accessible, but also increases some of the sophistication that is possible because some of these large language models are very, very good at spotting multi-dimensional patterns that maybe humans would be less good at doing. So the traditional stuff where we had highlighted where the top five things going wrong might be, but generative AI enables you to look at where exposure has changed as well as how profitability is doing and spot connections between the two. So think generative AI adding a degree of sophistication and suggestion to that already

And I think generative AI can be used in quite a few places in a portfolio management control cycle. as well as sort of noting what's happening with the merging experience, might suggest underwriting actions you might want to take. It might suggest changes to the rating structure itself. If you decide to do something, generative AI can help with some of the testing of a newly deployed rating algorithm sort of it's akin to software engineering, it's quite good at supporting regression testing. Generative AI can be used for data enrichment, unearthing new predictive factors, new rating factors, new underwriting factors that you might want to use. So there's all sorts of areas where generative AI can help out in the portfolio management cycle. Even on the traditional predictive modeling sort of cycle, the pricing modeling cycle, you can get generative AI or you can get automated AI traditional stuff.

Duncan Anderson: running models more automatically and then generative AI spotting potential suggestions that you might want to make to the core model. I think lots of areas where generative AI can help, which is of a material assistance, but I think it's when you move to agentic systems that the game changes. So it's agentic AI, which is the force multiplier. So that is when all these useful tools that suggest things actually link up, interconnect and still with an appropriate human in the loop, and we'll maybe come to that later, will actually act and actually do things around the whole control cycle. So the whole control cycle actually becomes a much more automated thing where, subject to appropriate governance and checks, things are actually done. And that has huge consequences for the speed of the control cycles. So certainly the sophistication is very good at spotting things that maybe humans would not.

The key thing is it means everything happens much more rapidly. The control cycles are much more rapid. And I think it moves to something much closer to a real time control cycle in all the different things across pricing, portfolio management, reserving, and everything. And I think that's one material change. And I think there's another interesting consequence of this as well, which is that for many, years before AI was a thing, we've been feeling as a firm that ensure technical functions should probably converge over time. The need for faster, more granular analytics, greater agility, more automated insight pushed insurers towards having technical functions communicate more with the same metrics, with the same data. And we see the sort of slow change in that direction as well. I think when you have agentic systems doing this, I think that becomes much, much easier. In part, some of the slowness to have that change toward the combined technical functions a little bit down to human resistance to change perhaps. AI agents may have no such concerns over sharing information and data in that way. So I think that's a potential change as well. that's a very long answer, but there's quite a lot of changes I think we will see from this. I don't know what the time scale is, but likely to be faster than we think, I think.

Charlie Samolczyk: There's a lot to unpack there, certainly. I like the bit. You were talking about generative AI, almost changing the paradigm of how practitioners interact with the tool and generating insight, but importantly, it being able to now suggest future action. And then you use the words, the force multiplier, which is agentic AI, which think is powerful.

And then, know, lastly, I think that alignment, because I know that we've been working on that for a long time, the, know, agentic AI as an accelerator to align the technical functions. You did also happen to mention the kind of the human factor in there as well. And, you know, I think sometimes this is where people maybe naturally get uncomfortable, but I think unnecessarily in a lot of cases, you know, in your opinion.

What does this mean? Does AI replace actuarial and pricing expertise? Does it augment it? What's your view on

Duncan Anderson: Don't think it replaces it. I think it increases the value of deep expertise and it enhances it. There might be questions about whether you need so many experts, but in terms of expertise in general, I think that will absolutely be necessary. So think what AI will do, what agentic AI will do, it'll take away the drudgery. It'll do the legwork. It will do the handle turning. It will do the easy stuff. I think it won't be doing some of the harder stuff where judgment is required or whether are ethical judgments to be made, nor can it replace human accountability, I suspect, in a highly regulated industry. So we work in a regulated industry, I think there will be humans accountable for decisions and the actions of an insurance company for some time to come. So I think there will need to be humans in the loop and they will need to make appropriate decisions interrogating the AI systems, understanding what is proposed, being able to explain what is proposed and essentially approving particular things. So I think it doesn't replace experts, I think it enhances it. It doesn't replace human accountability, but I think that accountability will change in its nature.

Charlie Samolczyk: Just to step out of it there, I was gonna ask, does accountability shift in a world where agents handle routine work and escalate exceptions? But I think you've answered that.

Duncan Anderson: Yeah, you want to say exactly how, what do mean by accountability will change? That's your next question.

Charlie Samolczyk: Okay.

Charlie Samolczyk: So, know, experience or expertise, sorry, not experience, expertise gets enhanced, sorry, that again. So expertise gets enhanced, but I think the accountability piece is really important. So what do you mean by, you know, I guess how will accountability change?

Duncan Anderson: Well, I think ultimately the leaders in an insurance company will remain responsible and accountable for the ultimate outcomes. And I think that means they need to take responsibility and accountability for two things. Firstly, they're going to have to design the agentic system. They're going to have to work out what they delegate, if you like, to the agents and when they wish to review exceptions. there's a design thing that they have to do.

And then once designed, actually have to operate them and then review the exceptions. oversight sort of evolves to being a policy driven thing with an exceptions based activity. So leaders need to define the guardrails, the thresholds for review, what decisions can be devolved and what can't be. And then they've got to think about what will need to be surfaced so that they're appropriately and complete judgment can be made in each case. think that's, know, leaders will design and operate the systems to help them with that, to help with governance. I can see a world in which you actually train up a number of agents that are actually governance agents, so policemen agents, so sort of design them to be particularly fastidious things, looking for issues to question and, you you pick one set of agents against another almost to... to check that what's been done is okay, but ultimately that will not replace bottom line accountability. So I think that's what will happen there. I think also this, the idea of big lumbering cycles where you look at data and things in a particular set time period, I think that will change. Whether this will be a fun change, I don't know. I'm reminded a little bit in the 1990s when we used to get paper memos coming around every morning and every afternoon. And then when one day someone put an email, PC on the desk and all of a sudden you can get emails all the time, which was great for a day or two. And then that changed the way we worked a little bit. So I think there may be similar changes in terms of the real time governance that will happen with the agentic systems when they're up and running. But I think that's the nature of it. think we're a regulated industry. I think that will define what has to be surfaced and where the ultimate accountability sits. And I think, you know.

Duncan Anderson: humans will not be replaced for the ethical side and the ultimate financial accountability for any time soon, I feel.

Charlie Samolczyk: You mentioned that we are a heavily regulated industry and I guess in that context, certainly everyone needs to live up to a certain standard and the tools do as well.

Off the shelf AI though, I guess my question is that sufficient or appropriate or do we need something I guess more domain specific and explainable?

Duncan Anderson: I think it very much has to be domain specific. think they very much need training with relevant insurance background. I that goes in a number of different directions. So at the heart of a lot of what we do in pricing, reserve, and capital modeling, there's a very specific hardcore calculation. Regulators are very unhappy if that changes for no apparent reason. So I think at the heart of a lot of what we do, we will need a deterministic algorithmic core.

I guess by then I am including machine learning as honorary deterministic in that sense, I think. So I don't think we want non-deterministic LLMs coming up with numbers at the very heart of what we do. But I do think we will have agentic systems running these deterministic things in the middle. So I think one thing from an assurance perspective, we will need appropriately robust deterministic analytical cores with the agentic stuff built around it. But I think the other thing that we will need is that there is a lot of insurance specificity that needs to be built in for these things to be usable using off the shelf stuff. We've experienced this in developing some of our own software using off the shelf stuff just doesn't get the results that you need until you actually give it some background and some knowledge as to what claims are, what premiums are, what loss ratios are, how these things work. You also need to bake in strict governance everywhere and train things as well. A of these agents are a little bit like eager, keen to please interns. And sometimes you get the answers that look helpful until you scratch a little bit deeper. And they're, I'm so sorry. No, I didn't mean that. Yes, you're quite right. It's not that at all. So you need to make sure that they are well-trained to be reliable as much as possible and that they run checks on themselves as well. So I think just using off-the-shelf stuff without that appropriate training is very dangerous.

Charlie Samolczyk: And you kind of alluded to there that, you know, it's actually a bit of a different skill set, I think, for the people that are running these. You talked about, you know, making sure that the training is right or making sure that, you know, what they're doing is baked in or having the kind of the right oversight. So I guess, you know, what does that mean for the insurance professional? What skills, you know, become essential in an agent-driven technical function versus maybe how we've been doing it today?

No, I think there will be a slightly different skill set needed. think, you know, insurance expertise absolutely needed still. But on top of that, insurance experts will need to be appropriately AI literate. So I don't think they need to be deep, you know, software engineers creating the, know, you know, gubbins of this, but I think they need to be able to design agentic systems to design, interpret and challenge the AI agents. And probably if the prediction that there will be this sort of closer alignment of technical functions comes true, then I think there might be a little more need to understand expertise in more than one technical function so that you can converse across the boundaries a little bit more easily. I think some things will remain unchanged. The ability to innovate and to do R &D, I think that will mostly remain a human thing. Although we have seen already that some AI tools can be incredibly helpful accelerating some of that innovation and R &D. So you can set some modeling off with some vague ideas and the way it can be automated and test things out very quickly is quite interesting. And even when we've actually derived some new algorithms and we have some patented algorithms that we have, the actual execution of those in terms of coding them up.

Some of the coding AI tools are really quite amazing now as to what they can do and that can accelerate things as well. But in general, I think there will be a need to learn. I think this touches on the thing you said earlier about, I think the technology is largely there. What we're facing now as the biggest barrier is probably the human change management thing. So people do need to start learning now. I think companies that will succeed are those that are on the case with training already. I do know one organization, I wouldn't name them, but for the last year, they were very good at getting on top of this for the last year. They mandated 45-minute lunch and learn sessions on AI every day for all staff. So, some companies are really making a big, big effort.

Duncan Anderson: To drive familiarity with these tools. I don't think we're quite there, but we're not far behind. I think we can just see from the number of RFPs we're getting at the moment about AI transformation programs. Clearly a lot of insurers are concerned about what it means, but a lot of this is about getting people trained up, changing and thinking in the right way.

Charlie Samolczyk: Let me ask you a question, maybe putting on your other hat, if you will, because you're a practitioner, but you also run an insurance software company. So you kind of see it from that perspective as well. What do you think this means for insurance software platforms? And I guess how do they need to evolve to support a world where humans and agents are working together, I guess?

Duncan Anderson: Well, thankfully, yeah.

Well, thankfully, I think there is still very much a role for specialist insurance software, which I think is good news, but clearly we need to change a little bit what we are building, as well as how we're building on I'll touch on that one in a second. So, I mean, there's some obvious stuff. There's an increasing expectation that there are sort of supportive generative AI tools within the traditional sort of products that we have.

Charlie Samolczyk: Good, yep.

Duncan Anderson: A lot of our analytical products involve constructing models, be it a capital model or cash flow projection model in life insurance or a cash flow scenario testing thing for P &C pricing. People expect AI support in that model building, so we've done some of that. People now increasingly expect... tools to translate from one environment to another, so you know, from Excel into something so that, you know, there's clearly roles that generative AI can play in that. And as well as the interpretation of the thing I was talking about earlier about our software, which furnishes sort of insight in monitoring and merging experience. So an interpretive layer around that, you know, that's a nice use of generative AI. So that's one thing, but that's essentially, you know, using last year's stuff almost. I think the one that is we're keeping a very much an eye on is how our software and other software will play in an agentic system. think, you know, in the future people will not use, you know, one single provider for everything. They will have an agentic system which is using lots of tools along the way. So our tools will very much play nicely in agentic systems. So we're making sure they can be controlled by and can control other agents in an appropriate way.

But I think the other thing that we're thinking about is coming back to what said earlier about human accountability and sort of the interface layer. So we think as the insurer target operating models change, there will be more of a need for a sort of different type of interaction. So maybe less emphasis on the manual hand crafting of models and the turning the handle and more emphasis on the surfacing of insight and recommendations for review.

So there'll be more of an interrogative sort of accountability layer, if you like, I think. And I think that will change the nature a little bit of what we have to produce. So think we're pretty busy. We've to work it out on a few of those things. So think the things that we can now dream up that we want to build have increased. And we weren't short of ideas anyway. Thankfully, the one thing I sort of touched on slightly earlier as well, which is there is some amazing productivity gains we're seeing in software engineering itself.

Duncan Anderson: Using some of the AI coding tools across the software development lifecycle, not just coding, know, QA, product design, UX and everything. There are material uplifts we're finding in the use of AI tools. It's not a golden panacea that some might have you believe, nevertheless, there's very material increases in efficiency we're seeing from that as well, which sort of mitigates all these, the increased workload from all these new ideas we've got.

Charlie Samolczyk: But maybe I'd go so far as to say there will be winners and losers potentially in this space, depending on how people adopt or how agile they can be. Who do you think, if you think about the next phase, who might win in this phase? what's going to separate insurers from those who truly transform from those who don't?

Duncan Anderson: Yeah,

Duncan Anderson: Yeah, it's funny, actually, I'm reminding of a lunch I with a chap who know who he is. I think he was quoting someone else as well. But one of my favorite quotes with AI is that with AI, you're either at the table or you're on the menu. And I think that, you know, that resonated a lot over the last year or so. Yeah, but what does that really mean? I think the key thing now is doing stuff. It's execution.

So I come back to the thing, it's a human change management issue. think the technology is there. Technology is going to get better and better and better, but that technology is no longer the constraint. With the technology and the AI tools that are out right now, there is so much that can be done and the barrier is getting organizations to adapt and change. There's heaps of proof of concepts going on all over the place, loads of companies experimenting, but the ones that will win are the ones that execute and actually do stuff.

So think that's the, you know, properly executed change management programs, I think is the key thing. you know, how could you tell if someone's getting there? Well, I think, you know, we have this vision that, you know, if there's an analytical team with 50 people today, maybe that's actually done by five people and 50 agents in the future. So, you know, what will the winners look like? Maybe it's the ones that are actually managing to get material work done by the agents in practice rather than in a proof of concept thing.

Charlie Samolczyk: As we've talked through this, I think one thing comes through very clearly. Agentic AI isn't just another technology layer bolted onto what insurers already do. It's a shift in how decisions are made, how control is exercised, and where expertise is applied. We started talking about faster models and better insight, but what we've really been describing is a move away from slower, potentially committee-driven control cycles towards something much more responsive potentially a world where agents handle, I think you said the drudgery, but the routine activity, they surface the exceptions and then that we can potentially connect the technical functions in ways that simply weren't possible before. And you touched on this isn't about replacing insurance expertise, if anything it elevates or it augments it and the value shifts away from just manually producing outputs and towards having skills where you design systems and you set the guardrails and you potentially challenge the outcomes that you're getting from the agents, but you still own the judgment and the ethics and the accountability and the transparency. And that technology though moving fast, I think it still comes down to the people and the governance and the operating models. And as you pointed out, I think in my intro, I think what it's not a technology problem to solve anymore. It's more of a leadership and a change problem and that ensures that when, they won't be the ones with the most pilots necessarily or the flashiest demos, they'll be the ones who are willing to make a start and to redesign how their technical functions actually run and think about the role that humans and agents play in that. So first off, thank you very much for joining me today in the insight. think maybe as one, I'll let you kind of have the closing question. mean, obviously you've lived through lots of shifts. You talked about, you know, PCs on desks all the way through to aggregators and machine learning and automation. Thinking about a agentic AI in that context, if you had to give insurance leaders kind of one piece of advice as they look ahead and, you know, what should they start doing now, emphasis on now to kind of make sure that they're on the right side of it.

Duncan Anderson: I think map out what needs to be done and start executing bit by bit. Think about where your expertise lies. think significant expertise in the future becomes more valuable than it is today. So I think, know, agent existence will sort of multiply up that expertise. A lot of experts today maybe spend a lot of their day not being expert, but actually doing, you know, drudgery things, more handle turning things. So I think those that really understand insurance and can sort of provide expert interpretation, judgment and strategic decisioning will become more valuable. think there's an issue to consider around training in the long run as to how we continue to have experts in the future if there's less learning done by doing the drudgery. But I think it's about, you know, planning out a sensibly ambitious program and then bit by bit executing on it and trying to get value incrementally rather than waiting for ages for one massive system. But maybe I'll repeat what I said at the very beginning. I've seen a lot of changes in my career, but I do think this is the biggest, widest reaching and most fundamental change to ensure technical functions that I've seen. I think some very interesting but challenging years ahead.

Charlie Samolczyk: Thank very much Duncan. Appreciate you joining.

Duncan Anderson: Thank you.