Agentic AI in insurance: hype or real value?

Agentic AI is one of the loudest buzzwords in enterprise tech and insurance is no exception. But behind the hype, the reality is more complex. Giles Mazars, Group Chief AI Officer at vector8, breaks down what agentic AI really means today, why it’s not yet ready for production and where it might eventually add value.

Let's start with the basics. What is "agentic AI" in its current understanding? The term seems to be generating a lot of noise.

The term "agentic" itself is quite controversial within the research community, as we've had discussions around coordinating agents in AI for a while now. However, as of today, the common understanding of agentic AI is giving a machine learning system a problem to solve, and letting it solve the whole problem, not just a piece of it. This means the system goes through a reasoning process and codes the necessary tools to achieve the objective with the least possible human intervention. In the new AI world, especially with LLMs, people are often mixing up definitions, and agentic is often used to describe asking a large language model to handle reasoning and execute tasks using provided tools.

Why is the term "agentic" controversial in the research community?

There is extensive literature on coordinating agents, defining what an agent should be and how they are controlled. People often use the term "agentic" without fully understanding its current meaning. Additionally, the definition of an agentic system is very broad, essentially meaning any software that tries to accomplish an objective with tools and autonomy could be described as "agentic".

How are large language models (LLMs) currently being viewed through this "agentic" lens, and what are their limitations?

Today, agentic AI is largely looked at through LLMs because they can identify and use available tools to solve a problem. They can execute tasks with these tools, get results and move to the next step. However, there are significant problems. The reasoning capabilities of LLMs are still extremely limited. It's only been in the last few months that LLMs have been able to perform simple "chain-of-thought" reasoning by themselves. Thinking a bank or insurance company will fix real production issues based on technology that has only been working in research for the past three months is probably not realistic.

Given these limitations, is the hype around agentic AI distracting businesses from more immediate, solvable challenges?

I don't think so, not in the reality of discussions with customers. While there's certainly a lot of talk in the media and on LinkedIn, our customers are very conscious of what they can and want to do. They largely have a plan; it's not a crazy panic. My feeling is that people are educated, trying to understand what they don't know, and using what they do know to make good decisions, focusing on building an AI strategy. They understand there are steps to get there, and agentic is seen as a future, less mature technology.

How do agents, skills and accelerators fit into vector8's approach to AI, and how do they relate to the "agentic" idea?

At vector8, we've developed a blueprint for implementing and scaling AI, moving clients towards an agentic future. We call this our AI Implementation System, or AIS. The ultimate goal for us is to build AI with longevity for our clients, avoiding isolated point solutions. We do this by creating modular, composable building blocks that build AI services that in time are orchestrated by agents to automate part or full workflows.

What are the current limitations and challenges of building for agentic AI in an insurance context?

Building for agentic AI is currently out of reach because it's not a mature technology. The pieces needed for it are still being developed, and it's not yet well-tested or robust for production environments. Challenges include controlling and monitoring agents, synchronising tasks and managing the costs associated with tasks launched by agents. While interesting for research, in business-to-business discussions, it's more practical to focus on orchestrating tasks and improving current workflows.

What are the key areas where the current understanding and implementation of agentic AI still need significant development?

To summarize, agentic AI with LLMs is an incredibly interesting path, and there's a lot we could do. However, we still need to work on controlling it, finding ways to explain its actions, improving reasoning capabilities and shaping it to be used safely in business. We are not there yet. There's too much time spent talking about it in articles and posts, while in B2B discussions, there are many other crucial things to focus on, such as orchestrating tasks, handling human validation, and determining where to put humans in the loop or AI in the loop of humans.

Ready to build AI that works in the real world?

We help insurers design AI systems that are practical, reusable and ready to grow. Whether you're early in your journey or looking to scale with confidence, we focus on what works today and what sets you up for tomorrow.

Talk to our team to see what’s possible.

‍