Beyond chatbots: How AI agents are transforming insurance

SALMA GALVÁN: Hi, Rick. Thank you so much. I'm glad to participate.

RICK HAYES: And welcome, Amanda.

AMANDA HUG: Thanks, Rick. Happy to be here and excited about the discussion.

RICK HAYES: And welcome, Eddy.

EDDY TRIVEDI: Thanks, Rick. Looking forward to this discussion.

RICK HAYES: So before we start, I'd like each speaker to say a little bit about what they do at WTW and their experience. So, Salma, why don't you kick us off with a brief introduction.

SALMA GALVÁN: Sure. I'm part of the Mexico City's office, so I work in ICT. Essentially, I do outsourcing projects, so we support the North American offices for the life team.

RICK HAYES: All right. Thank you for that. Amanda.

AMANDA HUG: Hi, everyone. I'm Amanda Hug. I am a director out of our New York City office, where I lead our group of life actuaries there. I am a recent past president of the Society of Actuaries and excited about this conversation because a lot of the work that I do is to help our clients modernize and automate and upskill their actuarial functions, and AI is definitely a piece of that. So excited about what we're chatting about today.

RICK HAYES: Wonderful. Thank you for that. And Eddy, we'll finish up with you.

EDDY TRIVEDI: Hi. It's Eddy. I'm a senior director with WTW. I'm based in Atlanta office. I've been with the firm for more than 20 years. And over the course of 20 years, I have really focused on financial modeling-related projects and most recently, the impact of AI on financial modeling-related activities.

RICK HAYES: All right. Well, thank you all for that. Now, let's get started with the AI discussion. So I've got a few questions, and I will rattle them off. And you guys can feel free to jump in. First off, can you explain what AI agents are and how they differ from traditional AI systems or chatbots?

AMANDA HUG: Rick, this is Amanda. I can take that. So I think this is a great primer for folks who may not have exposure to AI agents. Hopefully at this point, every single person listening has exposure to AI chatbots like ChatGPT, or Claude, or whichever, Microsoft Copilot. And so we're very familiar with interacting with those.

And when you interact with a chatbot, you're typically, it's a single prompt and a single query with a single answer. And that answer might be quite long, but you really have to be with it every step of the way and ask those directive questions. So the difference with an AI agent is it has the ability to plan and perform multi-step, complex tasks.

AI agents can take a goal and really break it down and then execute a series of actions to accomplish that goal. And so very different than the call and response of a ChatGPT. You can think about it more like an employee, where you might assign them a big project with a large goal, and then they're going to need to figure out how do I accomplish this through smaller steps.

Another big distinction with chatbots is that AI agents can use tools, so they can interact with the outside world. So again, much like an employee might do, they can query databases, send emails, or perform other automated tasks. In addition, AI agents have memory, so they can learn from past interactions and improve over time. So while chatbots are powerful, AI agents really just take it to that next level.

RICK HAYES: Wonderful. Thank you for that. So what would an AI agent paradigm look like in an insurance company then?

AMANDA HUG: Yeah, so in an insurance company imagine your org chart today. You have many boxes and layers and all those boxes are filled with people.

And so the visual that I think of is you still have your org chart, you still have your boxes, and you still have a whole lot of people, but you might have a new world where some of those boxes are filled with AI agents, or maybe some new boxes are added to create a team that didn't exist before that's made up of AI agents.

And one of the key benefits of this AI agent paradigm is you can create agents that have specific roles within an organization. So just like in an insurance company, you have roles for underwriting, claims processing, customer service, fraud analysis. Whatever it might be, you have some specialization. And so we'll eventually have AI agents that are tailored to each of those functions.

And the cool thing is that those agents won't just handle their specialized task, but they can interact with each other, they can delegate tasks, and they can collaborate again, much like employees do within an organization.

And so one example might be, maybe there's an AI agent on the fraud analysis team or the claims team, and they're trying to process the claim and make sure that it's legitimate. They might reach out to the underwriting team of agents and say, hey, can you help me? I need some information about how this was underwritten. What were the answers on their application? And so they'll collaborate in that way.

And so the advantage of building these AI agents to be very specialized, instead of say, having one large monolithic agent that can do everything, is it's easier to build, easier to maintain, and you can really optimize it for these specific tasks.

RICK HAYES: Thanks, Amanda. That was very helpful. Now, maybe a little bit more focused in terms of accessing and incorporating data. So how does Retrieval Augmented Generation enhance the capabilities of the AI agents?

SALMA GALVÁN: Thank you, Rick. This is Salma. I know this may sound like a difficult concept like Retrieval Augmented Generation, but essentially, I want to break it down to the three words that it means. So retrieval comes from retrieving information from external data sources.

Augmented, it comes from the fact that all this information that is retrieved, you can improve your prompt. So the AI agent will take all this information as context and as more information. And then generation, obviously, because we are using, and we are generating language based on that data.

So instead of relying only on the static knowledge the model learned during training, RAG enables an AI agent to access real time or domain-specific information, such as underwriting guidelines, policy details, or regulatory updates. And then it incorporates it into its response.

So this is critical in insurance because we know that regulations, coverage terms, and pricing models, they change frequently. For example, imagine an agent helping a broker quote a policy.

With Retrieval Augmented Generation, it can pull the latest underwriting criteria from the company's internal knowledge base, and combine that with market data to provide accurate recommendations instantly.

Similarly, a claims assistant could retrieve recent fraud detection patterns or local compliance rules before advising on a claim. So obviously, the result is more accurate. It's context-aware answers. We have fewer hallucinations. And we have a system that adapts without retraining.

So in short, Retrieval Augmented Generation turns AI agents into dynamic tools that stay relevant and trustworthy in a fast-changing insurance environment.

RICK HAYES: Excellent. All right. So next, can someone discuss the role of prompt engineering and designing in effective AI agent interaction?

SALMA GALVÁN: Sure. So well, if you have used AI, everyone has used prompts. That's essentially the way you communicate with the model. So prompt engineering is the art of crafting instructions that guide how an AI model responds. So it is essential because the way you frame a prompt determines the quality and relevance of the output.

A well-designed prompt provides context. It defines the role. It specifies the format of what you are expecting to get from the agent.

So, for example, instead of asking, explain this policy, you might say, act as an insurance advisor and summarize this life policy in three bullet points for a customer with no technical background, which is more clear. So that clarity ensures that the agent delivers useful, compliant, and customer-friendly responses.

And the good thing also about this is that prompt engineering is also iterative. You can refine prompts based on feedback and performance. So ultimately, it's what makes the interaction feel natural and accurate, turning a powerful model into a reliable assistant.

RICK HAYES: Very interesting. So maybe more broadly now. So what areas within the actuarial modeling or actuarial space do you think AI can make the most significant impact and why?

AMANDA HUG: Yeah so at WTW last year, we did an analytics survey with our clients to find out what areas they were already using AI for or planning to use AI in the future. And the most common, by far 60% of clients said that they were using it for code writing and model generation. And so that could include writing code, debugging, testing, translating, reviewing. So certainly, a big use case in the actuarial space right now for code.

They also expressed an interest in using AI to generate documentation. And that's going to be an area where these models excel because it's a text-based task. And then there was some interest in using AI for testing and to a lesser extent, data cleansing. So really, coding is kind of where actuaries are using AI right now. Hopefully, that will expand in the future, but that's where we are.

RICK HAYES: Thanks, Amanda. So what are some of the AI initiatives that WTW has been involved in this far.

EDDY TRIVEDI: Thanks, Rick. So at WTW, recently AI as a great tool to add to the arsenal of financial models. We recently added an AI assistant to our financial modeling solution, RiskAgility FM. The AI assistant works with Microsoft's Azure OpenAI model's data client licenses. This ensures that any data or queries passed to the AI assistant are secure and stay within the client's environment.

AI assistant allows a modeler to ask basic questions related to financial modeling. It goes beyond that. It allows a modeler to leverage AI to explain the existing model code. It can also assist the modeler in fixing modeling issues, and more importantly, it allows the modeler to use the AI assistant to write new code for any model updates enhancements.

In addition to this AI assistant, we have also created separate tools that allows a user to write model documentation and migrate models from Excel to RiskAgility FM. Actuaries are great, but I have to admit that when it comes to documenting the work, they can do more. They can do better.

So with our tool, you can just point it to the model and it'll tell you what the model does, provides a great first draft of the model documentation. A lot of actual work when it comes to financial modeling is still done in the prototyping phase is done in Excel.

So once a client is ready to move from prototyping phase in Excel to more robust modeling, our tool can help the clients migrate the existing Excel models to RiskAgility FM and continue their work in a more robust solution. There is more to come in 2026, where we integrate agentic AI more deeply with the solutions, so stay tuned.

RICK HAYES: Thanks, Eddy. All right, so now, considering the background you've provided me and your general experience, I'm curious to know whether you believe AI is going to meaningfully enhance the life and annuity industry or if it's just too early to tell. So maybe we can go around the room and get your honest feedback on that.

EDDY TRIVEDI: So I think AI is a useful tool. Again, there's some hype around AI. AI is not magic. It cannot do everything for you, but it can make you far more productive and more efficient than what it has done, what we can do with AI. We've already seen evidence of that when it comes to model migration, model documentation, and even writing new code. So it is real. It is helpful, but it is not magic.

AMANDA HUG: Yeah, I can build on that. I mean, I think the examples that Eddy shared and that we've done here at WTW and frankly, that we've seen some of our clients doing in real life are extremely powerful.

Think how much time actuarial students have spent on taking an Excel-based model and then putting it into an actuarial software. That takes a lot of time. So to have AI be able to do the first kind of pass at that is a huge time savings.

And similarly, think about documentation. That can take days. And if you do it well, you have to keep it maintained over time. Again having AI do that first pass at it is really going to free us up to work on what we actually care about and more interesting analytical and value-added work.

So I don't think any of the three of us would be on this podcast if we didn't think AI will meaningfully enhance the life insurance industry. So we're still in the early phases, particularly on the actuarial space. But I would say in the next two to three years, I'm really excited to hopefully see exponentially more use cases to improve our industry.

SALMA GALVÁN: Yeah, I agree with you both. I also think that AI is growing pretty fast, and I think we've seen a lot of improvements based only on the last two or three years, so I'm really excited to see what comes next.

I'm really excited to test all that because I mean, it's great that we are able to be the first ones to use all that technology and see all the things that we can do. So I'm really excited about the future on AI.

RICK HAYES: Glad to hear you all so bullish. Well, thank you all for contributing to the discussion about the implications of AI agents in the actuarial space. Thank you, Eddy.

EDDY TRIVEDI: My pleasure.

RICK HAYES: And thank you, Salma.

SALMA GALVÁN: Thank you, Rick. It was a fun discussion.

RICK HAYES: And thank you, Amanda.

AMANDA HUG: Thanks for inviting me.

RICK HAYES: And thanks to our audience for listening to this episode of (Re)thinking Insurance. We'd love to hear from you and understand which areas you think AI has the most potential to change. Thanks, everyone. Have a great day.

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, 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.