ZestyAI announced on April 15, 2026 that Scott Stephenson has joined its board. Stephenson ran Verisk Analytics as Chairman, President and CEO from 2013 to 2022, taking its market capitalization from roughly $10 billion to more than $40 billion. He did not go to another incumbent, and he did not stay to advise Verisk's own AI work. The interesting question is what the platform he joined is asking regulators and actuaries to accept.

Key Takeaways

  • From roughly $10 billion to more than $40 billion in market capitalization is what Stephenson oversaw at Verisk, the company whose cooperative data pool and ISO loss costs underpin most US property rate filings.
  • 99.7% of US properties covered at a stated 95% or better verified accuracy, with carriers using ZestyAI models insuring $3 trillion of insured value.
  • More than 200 regulatory approvals nationwide, which is the actual barrier to entry: a property-level score is worth nothing in pricing unless a regulator accepts it in a filing.
  • Carriers insuring roughly 40% of the California homeowners market use Z-FIRE, the first AI-based wildfire model approved as part of a California rate filing.
  • Nearly one million previously uninsurable properties were covered in 2025 using these models, double the 511,000 of 2024.

What Stephenson Left, and What It Was Built On

Stephenson joined Verisk pre-IPO and led it from 2013 to 2022, during which revenues more than doubled, the geographic footprint tripled, and market capitalization moved from roughly $10 billion to more than $40 billion. Before that he was a senior partner at The Boston Consulting Group.

What he built is worth stating precisely, because it defines the thing being challenged. Verisk at its core is a data aggregation and statistical rating business. Carriers submit loss experience, Verisk aggregates and analyzes it, and the resulting ISO loss costs become the starting point for individual carrier rate filings. That produced reliable territory-level risk classification for decades, and it is backward-looking by construction and geographic in its granularity.

Verisk is not standing still against the challenge. Synergy Studio consolidates more than 110 catastrophe models into one cloud environment, Model Context Protocol connectors route ISO Indications and XactRestore into conversational access, and seven new AI modules shipped in Q1 2026 alongside 30% aerial imagery revenue growth over two years. But the strategy is layering AI onto the existing pool rather than replacing the unit of analysis.

The Unit of Analysis Changes, and So Does What Needs Validating

The substantive difference is not model architecture. It is that the score attaches to an address instead of a territory.

ZestyAI states coverage of 99.7% of US properties at 95% or better verified accuracy, with carriers using its models insuring $3 trillion of insured value and a claimed 62x segmentation lift over traditional rating approaches. Z-SPARK, launched March 2026 against the $25 billion annual non-weather fire loss problem, claims 30x greater risk differentiation than territory-based approaches by reading building materials, maintenance condition, neighboring structures and local fire response capacity.

The number that actually gates any of this is smaller and less exciting. More than 200 regulatory approvals nationwide is what converts a property-level score into a rating variable, and it is the barrier no amount of model quality substitutes for. Z-FIRE was the first AI-based wildfire model approved as part of a carrier rate filing in California, and carriers insuring roughly 40% of the California homeowners market now use it, with approvals across New Mexico, Oregon and Utah and filings pending in Montana and Nevada.

For the actuary signing the filing, the segmentation claims are the vendor's and the responsibility is not. A 62x lift and a 30x differentiation are statements about a model's discriminatory power on the developer's validation data, not about how the score performs against a specific carrier's own loss experience in a specific book. Those are different exhibits, and only the second supports a rate indication.

That distinction was manageable when the model refined pricing on a book the carrier already had experience for. It is the whole problem when the model is used to write business the carrier has never written.

The Book Grows Fastest Where the Experience Is Thinnest

The commercial milestones disclosed alongside the appointment describe a company working: cash flow positive on roughly $62 million of total funding, 26 new carrier clients including Applied Underwriters, California Casualty, Lemonade and Marsh, and 12 expanded relationships including Berkshire Hathaway, the California FAIR Plan and CSAA.

One of those milestones is different in kind. Insurers using ZestyAI enabled coverage for nearly one million previously uninsurable properties in 2025, double the 511,000 of 2024, concentrated in catastrophe-exposed regions.

Read that as an actuarial statement rather than a market one. Those properties are, by definition, risks the carrier previously declined and therefore has no loss experience for. They are being written on the strength of a vendor score whose differentiation is validated against the vendor's data, in the perils and geographies where loss volatility is highest and a single event correlates across the whole cohort.

The credibility problem is structural and it compounds. A carrier expanding into previously uninsurable wildfire-exposed property has thin experience at the outset, and the experience it accumulates arrives slowly in frequency and violently in severity, which is the worst combination for validating a segmentation claim. Standard practice would weight the model's indication against the carrier's own emerging experience. Here there is no own experience to weight against, and the fastest-growing part of the book is the part where that is most true.

None of which argues the models are wrong. Territory-level rating manifestly failed these properties, since the alternative on offer was no coverage at all. But the growth figure and the accuracy figure are describing the same expansion from opposite ends, and only one of them has been tested by a loss year.

Further Reading

Sources