By Brandon Katz, Executive Vice President, Product & Strategy, KatRisk
There’s a lot of noise in our industry right now. AI is being positioned as the answer to some of catastrophe modeling’s hardest problems. My reaction is always the same: the physics still matter.
You cannot train a machine learning model on 40 years of loss history and expect it to predict a climate that no longer resembles those 40 years. And it isn’t only the hazard that has changed. The exposure has too. People and property have moved into new areas, development patterns have shifted, and the built environment an event strikes today is not the one it would have hit two decades ago. The same hailstorm over Dallas or Denver produces a very different loss now than it did then. Historical losses are doubly misleading: wrong climate, wrong exposure. Lean too hard on that history and you’re still looking in the rearview mirror, just with a more sophisticated dashboard.
So when people ask me what the next three to five years of catastrophe modeling look like, my answer is not that AI replaces physics-based science. It’s that AI makes that science faster, more scalable, and easier to use. Here’s where I see that happening.
Faster models, sharper calibration
The first shift is inside the models themselves, and it comes down to speed and calibration. Machine learning emulators trained on physics let us explore far more scenarios, at higher resolution, for a fraction of the compute, fast enough to sit inside live underwriting workflows. AI also accelerates calibration, tuning models against observations, claims, and event data at a scale and pace that used to take a research team months. The science doesn’t get simpler. It gets faster, and it gets tested against more information.
Vulnerability becomes an observation
The second shift is geospatial data, and I think it could prove even more consequential than the algorithms themselves. AI-driven feature extraction from aerial and satellite imagery means we can now observe what actually drives loss at the individual property: roof age and condition, defensible space, vegetation overhang, first-floor elevation. For most of this industry’s history, those characteristics had to be assumed or generalized. Increasingly, they can be observed directly. Combine that with physics-based hazard and you get true property-level pricing, and just as importantly, a quantified value for mitigation.
From reports to conversations
The third shift is the one I find most exciting, because it changes who gets to use a cat model. These models produce an enormous amount of information, and most of it goes unused. Extracting insight has always required a specialized analytics team writing queries. AI turns that into a conversation.
An underwriter or portfolio manager will simply ask: show me where I’m carrying too much correlation. Or: I have $100 million of premium to spend, where can I grow given my existing book while avoiding future accumulations, without any county’s 100-year loss ratio exceeding my threshold? That’s portfolio construction and optimization as a dialogue, with the AI doing the heavy lifting against the full stochastic output rather than a summary report.
But here’s the catch. Answering a question like that well isn’t one query. It’s hundreds of what-if scenarios run behind the scenes, each one a full portfolio analysis. That only works if the underlying model is fast enough to keep up with the conversation. KatRisk’s is the fastest on the market, processing 50,000-year event sets in minutes, which makes it likely the only model that can fully unlock queries like these. The models have always contained the answers. AI makes those answers accessible to the people making the decisions.
The real test is transparency
There is one constraint on all of this: transparency. Insurance is a regulated promise, and California’s PRID process is a preview of where the world is heading. If you want your model used in ratemaking, you have to be able to open it up and explain it. Black-box AI fails that test. Models need to be explainable, defensible, and open to examination.
That’s why I don’t think the winners of the next generation will be the ones with the most AI. The real opportunity is using AI where it adds value while preserving the scientific rigor and transparency that catastrophe modeling demands. It’s a race we at KatRisk are very comfortable running, because we’ve been the transparent, science-first alternative from day one. The quality of the answer will always depend on what’s underneath it. The physics still matter.
