Takeaways:
- Predictions generated with AI are hard to explain
- Agentic AI requires constant human oversight
- Usefulness of AI will be function-dependent
Insurers are looking for insights
This article is excerpted from a longer interview and edited for clarity. This is the first of two parts.
Is AI effective at predicting risk outcomes for insurers?
AI is able to generate predictions, but it's not the best way to generate predictions. We need predictions that are stable. We need predictions that are explainable. Machine learning could be a black box, and you can white-box it with AI, but I'm not saying machine learning is fully explainable.
There's a number of reasons why you don't want generative AI to predict risk. The only way to actually do this for an industry like ours is to provide deterministic and non-deterministic artifacts in a way that makes sense, so you will have machine learning, generalized linear models (GLMs) and traditional statistical models that are doing connections, and enhance that with the ability for these models to say if there's something funny going on, the data is not working the way it should and you should be looking at this.
Just telling AI to do a prediction is not a good use, because you will not be able to explain it. If you asked it 50 times, you would get 50 different answers. It's mind blowing how some of those answers are very close to what you would do with more traditional predictive knowledge or machine learning.
How should insurers then look at agentic AI systems?
The only agentic systems that actually deliver results are the ones where you're constantly shaping them, because you're working with the system to identify edge cases, where your golden data sets need to evolve, and the expertise of your people. The heuristic needs to flow into that system seamlessly. You need to monitor human signals and AI signals and degradation. These systems can only become good enough for an industry as specialized as ours if humans are part of the system.
What does this mean for the potential of AI for insurance companies?
We're very far away from systems that take over and make the decisions, with humans having no role. That's why when CEOs say AI is the reason for letting people go, I say
Are there functions that AI just can't do by itself, that require human participation?
Reserving is an example of a specialized actuarial function that AI cannot do on its own. It is a very complex process involving multiple human-based processes, multiple experts, several data sources and formats. It is a place where AI can generate productivity gains in some parts of the process, but not across all tasks.
For a process or function like that to leverage AI in a way that translates into productivity gains, it is key to embed human expertise and translate human decisionmaking into something a machine can read and learn from. This process is not something you only do once, but constantly have to iterate. Human experts cannot be replaced, and are needed as part of a continuously learning system that has people and AI working symbiotically.
What will the impact of AI on the insurance industry look like a year from now?
The impact will depend a lot on the function, the type of insurer and the market. Different parts of the industry will transform at different paces and adopt AI at different rates. It is function-dependent. Actuarial functions will change very differently from consumer-facing ones. The complexity, regulatory scrutiny and need for explainability in actuarial work means the transformation there will be more structured and incremental, while customer service and distribution functions may see faster and more visible AI adoption.
How much AI-driven transformation a market experiences is closely linked to how much sophistication is needed to compete in it. A highly dynamic, competitive market like direct personal lines motor insurance in the U.K. is already pushing hard towards digitization and AI-enabled pricing, underwriting and customer experience. The pace of change there is intense.
The honest answer is: it depends on where you sit. Some parts of the industry will look very different. Others will look much the same. The push towards horizontal AI adoption will be homogeneous across the industry, soon realizing that true transformation needs a lot of work in vertical, insurance-specific AI systems that require not only investment, but cultural change, the ability to constantly monitor, shape and interact with AI systems to scale, leverage and protect human expertise effectively.
Within the next year, we will see the most change in leveraging Gen AI and agentic AI systems to accelerate, but not replace, automation, digitization and data liberation.
