Takeaways:
- Creating AI models doesn't come naturally to insurers
- Explainability, dependability and concentration pose risks
- Third-party data raises risk of operational cost increases
Using AI has already proved tricky for
This article is excerpted from a longer interview and edited for clarity.
What do you see as some of the big risks in using AI?
A lot of insurance companies don't have an operating model to develop AI and maintain it. They have never been in the business of generating internal tools. It's a new paradigm, so you need a function that actually manages the lifecycle of these AI tools. Even if they're internal, they need to comply with regulation. They need to consider information security and cyber security. If you're doing things right, adoption is increasing and you will have a lot of overlap in different agents and agents that evolve more than others, that are not consistent among themselves, even though they should be. It's an additional thing that you need to monitor.
One of the biggest risks today is that, because of the lack of understanding of the technology, we're building things that do not scale. We're building things that have errors in them that are difficult to spot or expensive to spot, and people are not combining Gen AI and agentic systems with other things that are more cost-effective and adequate for our industry.
A number of risks are about how you navigate fears and anxiety about this technology, how to generate a cultural change to understand what you consume, the technology you have to interact on a daily basis. That generates a gap between the most efficient and highest talent that we have. It's very difficult to differentiate, and so you can generate disengagement for some people, and their productivity will drop.
Then you have all of the governance layers around ethical uses of AI. How much black boxing can we have, if at all? How much of that white-boxing exercise with AI is effective?
What risks are caused by the black box nature of certain AI systems?
There are several risks caused by the black box nature of certain AI systems. The first and most fundamental one for an industry like insurance is explainability. Insurers are legally required to be able to explain their pricing, reserving and underwriting decisions to regulators. A black box model fundamentally cannot support that requirement — which is precisely why the architecture matters.
The second risk is ecosystem dependency. There is a dependence on an ecosystem you don't control — you're building all of this, you're calibrating your models, and the brains are in the hands of five companies. You cannot control what is happening to the updates. They change, then things break, they don't work, they work differently. You're depending on models you don't own, and they're becoming very costly.
The third risk is concentration. When the entire industry is building on the same small number of foundation model providers, you have a systemic fragility that goes beyond any single company's exposure. That is why companies are using Anthropic, Microsoft and others, trying to figure out how to do this agnostically. The level and the pace at which technology is evolving is making that choice extremely difficult. It's a high complexity environment with a very high pace.
Are there risks to relying on AI providers who can sharply increase costs once an insurer is on board?
We can expect that to happen, and some of these companies have already announced that the way they charge the licenses and tokens is changing. That is a risk. It's not a new risk in the sense of the last 15 years we have been actually innovating on data that we don't own.
If your life insurance products are dependent on data from an Apple Watch, if one day Apple decides that the data from your Apple Watch is becoming more expensive, then perhaps your product ceases to work in the way you want it. It's not a new problem.
There is a lot of the data that we consume and that we use to enrich that we don't own. Now we're using methodologies that have the same risks, but as long as you architect things to minimize that risk and minimize the cost of those models, leveraging them where they make sense and not across the board, that's the right way to navigate this.
