Artificial intelligence offers significant potential for the insurance industry, but getting reliable results requires more than simply asking the right question. Understanding what AI does well, where it can fall short, and how human expertise fits into the process is critical to putting the technology to practical use.

Here, Chris Abbott (Gen Re’s Head of Analytics and Innovation) talks with Tim Fletcher (Senior Emerging Issues Specialist) about the practical business uses and limitations of AI in insurance.

Tim: You work extensively with AI, designing solutions to business tasks. How do you go about it?

Chris: In many ways, the process isn’t that complicated. I collaborate with our experts to design AI workflows to ask the right questions and deliver outputs aligned with how a knowledgeable professional would think through a task. Then I take that large-scale task and put it into a general AI chat, just as you would with any other question that you wanted to ask AI – and I’ll see where it fails. This helps me identify what the AI does well, and what it does poorly in that generic chat. From there I’ll break the question into smaller problems.

So, I’ll take a big item and split it into 10 pieces, and then run each of those individually through that same chat and see what happens. From there I have a better picture of what AI is best at answering. I can then go into a programming editor and build a tool that incorporates rules and guidelines around fixing the failures.

Tim: It sounds like a lot of refinement and modification is needed.

Chris: Yes, but it’s necessary. It doesn’t work if you put the whole problem into an AI chat and expect it to solve every step of the problem at once. Going down that path typically doesn’t yield reliable results. What we do is provide the model with a string of iterative smaller problems that can be rigorously checked.

Tim: Give us an example of a solution that you worked on at Gen Re, where you applied those principles?

Chris: Sure. One of the challenges we have at Gen Re is that we handle many different contract documents, in many different languages, which experts in the legal field must read and explain. Using AI, we can quickly translate those complex documents into English and present them back in a way that the information within those contracts is easier to find and understand.

I want to stress here that the process we have is secure and client confidentiality is maintained throughout.

Tim: As you develop a solution, how do you work with the business side to help you validate and/or modify it?

Chris: Absolutely. We collaborate very closely with appropriate team experts throughout the process to make sure that the best questions are asked of the AI tool, and that we understand what output a user needs from it.

Put simply, we build a programmatic workflow that mimics the thought process of a person who is familiar with the business topic and how they might answer someone’s questions.

Tim: And this is a process that can be applied to different business functions?

Chris: Yes, the workflow programming process can be applied to different areas of the insurance business, including claims management and underwriting. But remember, this is an application and not an expert. It is an untrained tool that is simply very good at solving the small tasks that we give it, in a very specific order.

Importantly, when applying AI to a multi-dimensional issue like claims management, we wouldn’t ask one complex question and hope for one perfect answer.

Tim: Give us an example of how that works.

Chris: We go through a lot of testing to understand the average user’s expectations and experience with the AI application. Take the question “I’m looking at a new claim based on the attached description. Help me evaluate this new claim.”

That prompt will give a very different answer from “I’m a claims executive and I’m worried about nuclear verdicts. I’m looking at a new claim based on the attached description to determine if I should escalate it to pursue settlement. Help me evaluate this new claim from Pennsylvania.”

Neither prompt is wrong. In the shorter prompt, you leave the door open for AI to add interpretation, some models are fantastic at this. In the longer prompt, you’ll get a targeted response about specific laws in Pennsylvania.

This is where user testing and understanding adds the most benefit. If a user can go into a generic chat agent and work through these nuances and preferences on their own, they’ve saved me a tremendous amount of time as a programmer.

Tim: We hear a lot about AI “hallucinations” and the need to keep a human being in the loop. Have you encountered this phenomenon?

Chris: AI tools have been known to generate incorrect, misleading, or completely fabricated information and present it confidently as a fact. The problem is that these tools have proved to be very good at guessing what you would expect to hear back. So, a hallucination could be a wrong, made‑up answer – or it could be a subtle tweak of some kind of intention that you unwittingly passed to the model. Again, this risk can be minimised by recognizing and understanding that these pitfalls exist, planning for them, and ensuring that people are reviewing the results of an AI tool for accuracy.

Tim: How do you reduce the risk of AI hallucinating?

Chris: Hallucinations are becoming less common as the models evolve. Still, it is part of our responsibility, as programmers and as humans, to stay in the loop and make it possible for users to pause, validate, and refine output before it is sent along to another path.

Also, AI models are increasingly involved in checking for hallucinations. An autonomous AI checking agent can be part of the verification process, evaluating output against certain criteria until it deems that the criteria has been met and the output is ready for human review and assessment.

Tim: What’s the biggest surprise you have encountered when working with AI?

Chris: The biggest surprise for me is that these tools are not as intelligent as they appear to be. They’re designed to make their output appear very convincing, so a lot of rigorous checking of what’s coming out of the model is needed. And it’s important to thoughtfully identify use cases where it’s appropriate to involve AI and those that are not. I don’t think tasks we do very well ourselves make good use cases for AI.

Tim: What is an example of something that AI can do really, really well?

Chris: That contract example I mentioned earlier is a good illustration of where AI can be genuinely helpful. It can scan lengthy, complex documents very quickly, identify key provisions, and help legal experts get to the relevant information faster. That can make the review process much more efficient. But I wouldn’t want AI making the judgment call. It should be used to support the lawyers by identifying and organizing information, not to replace their expertise or legal judgment.

Tim: How will the deployment of AI tools in insurance evolve over the next five years?

Chris: I’m watching, and somewhat concerned about, how the cost associated with using AI applications is increasing. I don’t think the pricing model that exists today for AI will survive into the future. It’s important because the decision criteria for when to use AI is heavily influenced by how cheap the models are. I think that the number and type of use cases we consider to be appropriate for AI will change as the cost starts to increase.

From a broader perspective, I expect that AI usage will expand across the insurance industry; however, that could create its own challenges. For example, perhaps in the future AI‑empowered clients will expect Gen Re to know and understand a lot more than we do today when it comes to their average reinsurance transaction!

Original printing in IN magazine, Fall 2026.