Why Are AI Property Models Becoming Core to P&C Underwriting?
AI-driven property models in U.S. home insurance are shifting from experimental tools to a baseline capability for underwriting — and the economics of consecutive record catastrophe years are accelerating the move. P&C Specialist reporter Vrushank Nayak detailed the shift in a May 2026 analysis, Ignore AI Property Models at Your Own Risk?, featuring perspectives from ZestyAI co-founder and chief product officer Kumar Dhuvur and senior director of regulatory & government affairs Bryan Rehor.
Read the full article in P&C Specialist →
What's actually driving the shift?
Per Dhuvur, the pressure is economic, not technological. The U.S. property insurance market absorbed roughly $89B in insured natural catastrophe losses in 2025 and $108B in 2024 per Swiss Re Institute — what Dhuvur described as "the most expensive operating environment in the history of U.S. property insurance," with no visible path back to historic loss levels. In that environment, the territory-level averages and patchy property-level inputs underwriters have historically relied on aren't fine enough to differentiate risk anymore.
What changed on the supply side is just as important. Aerial and satellite imagery, combined with computer vision, now generate consistent structure-level signals at portfolio scale. Roof condition, structural characteristics, visible damage, and prior exposure can be evaluated across an entire book in the same way they'd be assessed on a single inspected property. That capability didn't exist seven or eight years ago.
How widespread is AI property model adoption today?
Dhuvur estimates roughly half of U.S. carriers now use AI property models in some form, with underwriting the most common use case. Pricing adoption is slower, in part because of regulatory complexity. The P&C Specialist analysis of state rate filings identifies major national carriers — Nationwide, Liberty Mutual, Allstate, and State Farm — among those citing third-party AI property models in homeowners filings, alongside specialty and regional carriers using a wider mix of providers.
Why don't rate filings always reflect actual AI model usage?
This is where Bryan Rehor's regulatory perspective in the article matters most. Filing citation counts can understate real adoption because filing rules differ state to state. Some states require carriers to resubmit their full rating manuals with each filing; others require only the portions that have changed. In change-only states, carriers may continue using an AI model year over year without re-citing it in every filing. Colorado, for example, doesn't require the AI model to be mentioned in every rate filing, but does require carriers to document and govern any model in use, because regulators can audit at any time.
The implication: market-wide AI property model adoption is harder to read from filings alone than the raw citation counts suggest.
What's the cost of moving slowly?
Dhuvur framed the strategic risk as a compounding adverse-selection problem. When some carriers price properties at their true individual risk and others continue pricing on territory averages, the precise pricers systematically capture the better risks and the territory pricers absorb the worse ones. Over time, that gap compounds into loss ratios, retention, and growth — and the carriers most exposed are the ones whose competitors have already moved.
As AI property models shift from differentiator to default, the question isn't whether to adopt. It's how to govern adoption credibly enough to satisfy regulators while still capturing the underwriting gains.
Read the full article
Ignore AI Property Models at Your Own Risk? →
P&C Specialist, May 4, 2026, by Vrushank Nayak. Includes additional commentary from Patrick Schmid at the Insurance Information Institute and other experts, plus detailed analysis of which carriers and states are citing AI property models in current rate filings.
