Resources

Donegal Insurance Group has expanded its partnership with ZestyAI, adding Z-PROPERTY™ to deepen its understanding of the properties and exposures across its personal lines portfolio.

The expansion builds on a relationship that began in 2025, when Donegal selected ZestyAI's Roof Age solution to improve its own internal data and understanding of property risk across its homeowners business. Roof Age combines building permit records with more than 20 years of historical aerial imagery to pinpoint roof replacement events, delivering near-complete coverage of the contiguous United States with accuracy above 90%, turning a historically unreliable data point into one Donegal's underwriters could trust.

With Z-PROPERTY, Donegal is bringing that same accuracy and verification to the most critical property variables: roof condition, roof complexity, lot debris, and other features that impact risk and capture premium. That gives portfolio managers the evidence to make better risk decisions on policies, and gives Donegal a verified view of exposure across its personal lines portfolio.

"The quality of our decisions depends on the quality of the information behind them," said Jeffery Hay, Chief Underwriting Officer at Donegal Insurance Group.

"Our initial work with ZestyAI strengthened the roof age data available to our teams. Expanding the relationship gives us a more complete understanding of the properties we insure and greater confidence in how we evaluate, manage and reinsure risk."

"Donegal has experienced the benefits of our verified roof age data with regard to underwriting,” said Attila Toth, Founder and CEO of ZestyAI.

"Z-PROPERTY extends that standard across the entire portfolio. The carriers pulling ahead are the ones whose decisions rest on data they can defend."

ZestyAI has earned more than 500 regulatory approvals nationwide, giving insurers a trusted, market-tested foundation for putting property-level intelligence to work in the risk decisions that matter most.

When an insurer uses an AI model to price, underwrite, or settle a claim, who is accountable for the decision? At ITC Vegas 2026, leaders from across the insurance and technology ecosystem came back to a consistent answer: the carrier remains accountable, while the vendor shares responsibility for how the model is explained, governed, and used.

The discussion brought together Stephanie Kuczynski, Director of Risk Analytics at ZestyAI; Sandy Ulrich, SVP of Information Technology at Pharmacists Mutual; Rajat Sharma, Chief Revenue Officer at InsureMO; and Kobi Bendelak, CEO of Insurtech Israel, with Daniel Wolfe, Editor-in-Chief of Digital Insurance, moderating.

As AI becomes more deeply embedded in underwriting, pricing, and claims, that distinction matters more. Many of the models carriers rely on come from outside vendors, raising important questions around ownership, oversight, and accountability. Who is ultimately responsible for the decision? What does the vendor owe the carrier? And what does a regulator expect when a model influences an outcome?

Who is accountable when an insurer uses a vendor's AI model?

The carrier remains ultimately accountable for the decision, regardless of whether the model was developed internally or provided by a third party. But that does not absolve the vendor of responsibility. Carriers need enough transparency, traceability, and explainability to understand how the model works and defend its use.

ZestyAI's Stephanie Kuczynski framed that responsibility around three requirements: traceability, so a carrier can show how a score was produced; accountability, so ownership of the outcome is clear; and explainability, so the decision can be understood by a regulator or other external stakeholder. Those considerations have to be addressed before a model goes live, not after a decision is challenged.

Pharmacists Mutual's Sandy Ulrich drew an important distinction between running the technology and owning the business outcome. IT may support or operate the model, but the underwriting or claims leader acting on its output is still responsible for the decision that follows.

Rajat Sharma of InsureMO agreed that the carrier remains the final point of accountability, while noting that the maturity of governance still varies significantly across the market. Larger carriers are generally further along in formalizing these processes, while many smaller organizations are still building them.

Kobi Bendelak of Insurtech Israel added the vendor perspective. Early-stage technology companies may not have the same governance infrastructure as an insurer, which can shape what carriers reasonably require during vendor evaluation. It does not, however, change where ultimate accountability sits.

What does "human in the loop" actually mean for insurance AI?

Human oversight only works if the person reviewing an AI recommendation has the authority and expertise to challenge it. A human in the loop should own the outcome, not simply approve the output.

Pharmacists Mutual's Sandy Ulrich made that distinction explicit. The person reviewing an AI-driven recommendation needs enough context and judgment to question it when necessary. That requires training, not just a review step. Pharmacists Mutual has used intern programs and internal IT champions to help raise that baseline across the organization. Ulrich summed it up simply: train people to be thinkers.

Stephanie Kuczynski argued that vendors share responsibility for making that oversight possible. A model provider cannot simply deliver a score and step away. The carrier needs to understand how to interpret the result, trace how it was produced, and explain it when a regulator or other stakeholder asks why a decision was made. If the carrier cannot explain the model, the vendor has not finished the job.

Rajat Sharma of InsureMO added that the level of human oversight should depend on the maturity of the use case. In more established applications of AI, a human reviewer may add less value. In areas such as underwriting, where both the technology and the surrounding processes are still evolving, human judgment remains more important.

Kobi Bendelak offered a different view, arguing that the pace of AI development may eventually make traditional human review impractical and that AI agents could take on more of that oversight role.

How do regulators approach AI in insurance, and how does it differ by state?

For insurers, AI oversight is still highly state-specific. The questions regulators ask can vary by jurisdiction, but the focus is increasingly consistent: how does the technology affect the policyholder, and can the carrier explain the decision?

Stephanie Kuczynski spoke from ZestyAI's experience working with state regulators on behalf of carrier customers. There is no single U.S. rulebook for AI in property insurance, and each department of insurance brings its own expectations. In some states, regulators may focus on the age or source of the data behind a property score. In others, scrutiny may center on how the model affects rates, eligibility, or underwriting decisions.

Across those differences, Kuczynski said the conversation increasingly comes back to the consumer. Regulators want to understand how a model influenced an individual outcome, which makes traceability and explainability critical. A carrier should be able to trace a decision back to the relevant inputs and explain the result in plain language.

That becomes especially important when a model contributes to an adverse action. If a policyholder is non-renewed, surcharged, or declined, the carrier needs an appeal process that gives a qualified person the ability to review the decision and change the outcome when appropriate.

Regulatory readiness also has to be considered before deployment. Insurers need to know whether a model can be used in the states where they write business and whether that status remains current as requirements evolve.

This is where regulatory experience becomes a meaningful part of vendor evaluation. ZestyAI's models hold more than 500 regulatory approvals across state departments of insurance, supported by direct engagement with regulators in the markets where carrier customers operate. That infrastructure helps carriers respond more quickly and confidently when questions arise about how a property-level score was produced or used.

Kobi Bendelak added that the challenge is not limited to the U.S. In Europe, AI regulation can be even more complex, and startups may not fully confront those requirements until late in the sales or implementation process.

How should insurers manage data before deploying AI?

AI accountability starts with the data behind the model. If the inputs are inconsistent, incomplete, or difficult to trace, every downstream governance question becomes harder to answer.

Pharmacists Mutual's Sandy Ulrich emphasized the need for a single source of truth. Before deploying AI, insurers need confidence in the accuracy of their data and a clear understanding of which decisions depend on it. That becomes especially important as carriers experiment with large language models and other tools that may draw from multiple internal systems.

Rajat Sharma of InsureMO extended that point to the broader data environment. It is not enough for data to be clean for transactions; it also has to support regulatory scrutiny. That means building a data strategy that preserves the information needed to reconstruct how a decision was made, rather than relying on institutional memory after the fact.

Stephanie Kuczynski connected that directly to explainability. Traceability starts with the inputs: the carrier should be able to identify which imagery date, parcel boundary, model version, or other underlying data produced a particular score. Without that record, explaining the outcome to a regulator becomes much more difficult.

Ulrich also drew an important distinction between internal and customer-facing uses of AI. Tools used for internal enablement may warrant one level of governance, while applications that directly affect customers require a higher standard. The same technology can carry very different risks depending on where and how it is used.

What changes in AI accountability over the next 6 to 12 months?

The near-term shift is likely to be less about whether insurers use AI and more about how rigorously they evaluate, govern, and operationalize it.

Stephanie Kuczynski expects vendor vetting to become more demanding across the market. Large carriers already tend to require model validation, documentation, and regulatory review as part of procurement. Over the next year, more mid-size and smaller carriers are likely to adopt similar expectations as regulators ask tougher questions about how AI is selected and used.

For vendors, that raises the bar. The ability to explain how a model works, document its inputs and outputs, support regulatory review, and demonstrate that it can be governed in practice will increasingly become part of the buying decision.

Sandy Ulrich expects the most immediate gains from AI to come through efficiency, with teams first using the technology to streamline workflows before moving toward more complex decision support as confidence grows.

Rajat Sharma pointed to a similar evolution in process design. AI is already taking on a larger share of tasks such as software development, but he expects the bigger change to come when organizations begin redesigning workflows around AI rather than simply inserting it into existing processes.

Kobi Bendelak took a longer-term view, predicting that increasingly autonomous AI systems will become a much bigger part of the industry conversation over the next year.

One takeaway from each panelist

The panel closed with one final takeaway from each speaker, and together they captured the main themes of the discussion:

• Stephanie Kuczynski, ZestyAI: Accountability is shared, but the carrier remains the final point of responsibility. Vendors need to be transparent about how their models work and support carriers in explaining and governing their use.
• Sandy Ulrich, Pharmacists Mutual: Human oversight only works when people are equipped to challenge the technology. As she put it: train people to be thinkers.
• Rajat Sharma, InsureMO: Strong AI governance starts with strong data. Every other control depends on having accurate, traceable information underneath the model.
• Kobi Bendelak, Insurtech Israel: As AI systems become more autonomous, the industry will need to rethink what oversight and accountability look like.

For carriers, the practical implication is clear: a vendor's accountability should be demonstrated, not assumed. ZestyAI's models hold more than 500 regulatory approvals, and more than half of the top 100 carriers use them. Both depend on the ability to show how a property-level score was produced, explain how it should be used, and support carriers when regulators ask questions.

See how carriers deploy ZestyAI models.

FAQ: AI accountability in insurance

Who is accountable when an insurer uses a third-party AI model?
The carrier remains ultimately accountable for the underwriting, pricing, claims, or other business decision made using the model. The vendor also has important responsibilities, including providing the transparency, traceability, documentation, and support the carrier needs to understand and govern how the model is used.

What does AI explainability mean in insurance?
Explainability is the ability to communicate, in understandable terms, how a model contributed to a decision. That includes understanding the relevant inputs, how the model uses them, and why a particular outcome followed. Traceability complements explainability by preserving a record of the data, model version, and other information behind an individual result.

What should insurers require from an AI vendor?
Insurers should look for more than technical performance. A strong AI vendor should be able to document how its models work, provide traceability into individual results, support regulatory review, establish clear data governance and continuity practices, and train carrier teams to interpret and explain the technology appropriately.

What does “human in the loop” mean in insurance AI?
Human oversight is meaningful only when the person reviewing an AI-driven recommendation has the knowledge and authority to challenge it. The appropriate level of oversight will depend on the use case, but simply adding a human approval step does not by itself create effective governance.

How does regulation of AI in insurance differ by state?
Insurance regulation remains highly state-specific, and departments of insurance may focus on different aspects of an AI model or its use. Insurers need to understand whether a model can be used in the states where they operate and be prepared to explain how it affects individual policyholders. Working with a vendor that has established regulatory experience can make that process significantly easier to manage.

What is the foundation of AI accountability in insurance?
Reliable, traceable data. Insurers need confidence in the accuracy of the information feeding a model and the ability to reconstruct which inputs, data sources, and model versions contributed to a decision. Without that foundation, explainability and governance become much harder.

See how insurers are putting ZestyAI to work.

What do we build, what do we buy, and what do we partner on? That question was at the center of a conversation at ITC Vegas 2026 with Sai Vishnubhatla (USAA), Richard Pluschau (Falcon Risk Services), Itai Ben-Zaken (Honeycomb Insurance), and ZestyAI's own Founder and Chief Product Officer Kumar Dhuvur, moderated by Eugenio Gonzalez of Plug and Play.

The discussion made one thing clear: AI does not fit neatly into the traditional build-vs.-buy playbook. Models and capabilities are evolving quickly, specialized talent is difficult to scale, and insurers face growing expectations around transparency and governance. Just as importantly,

How do insurers decide what AI to build in house?

The panelists approached the build decision differently, but each came back to the same underlying question: is this capability important enough to justify owning it over the long term?

At USAA, that starts with a three-part litmus test. Sai Vishnubhatla said the company considers whether a capability is already commoditized, whether it has the talent to build it internally, and whether that talent will still be relevant as the technology evolves. That last point is especially important in AI, where models and capabilities are changing quickly. Building something in house is not just a question of whether you can develop it today, but whether you can continue to support, govern, and improve it over time.

That is also why USAA does not think about the decision simply as build versus buy. Vishnubhatla described the preferred relationship as an ally: a partner that brings expertise while helping the internal team become more capable in the process.

For Falcon Risk Services, the decision is shaped heavily by the operational burden that comes with ownership. Richard Pluschau noted that companies of different sizes will naturally draw the line in different places. Falcon, for example, has no interest in owning and operating a data center or taking on the staffing required for 24/7 uptime and support. What matters more is retaining IP and control over the capabilities that are strategically important, without taking on infrastructure the company does not need to own.

Honeycomb Insurance takes a similarly strategic view, but from the perspective of a disruptor. Itai Ben-Zaken said the key question is whether the company truly needs to own a capability. Honeycomb would rather buy something that already exists than recreate it internally, but that comes with an important counterweight: if the company buys too much of what makes the product distinctive, does it risk giving up its own differentiation?

Kumar Dhuvur framed the decision in terms of prioritization. A carrier may have 20 things it could build, he said, but choosing to build one means deciding that it belongs among the organization’s top one or two priorities. If it does not rank that highly, building it internally is difficult to justify.

And the commitment does not end when development is complete. Maintenance means continuing to add capabilities as the technology evolves, not simply keeping the system running. For models that touch rating, that commitment can also include navigating regulatory approval across multiple states.

What costs does the build business case leave out?

When the panel turned to total cost of ownership, time to value, and ROI, the discussion surfaced several costs that are easy to underestimate in a build business case.

Operationalization is one of them. Falcon's Richard Pluschau pointed to master data management as an example. The upfront cost of building the capability is visible, but the longer-term costs of governance, data quality, support, and service-level expectations can be much harder to account for. As Pluschau put it, if you build it, you "better damn well have the mechanisms in place to support it."

Regulatory approval can be another major factor. ZestyAI's Kumar Dhuvur noted that if a model affects rating or underwriting, the business case has to account for the regulatory process as well. That can add a year or more before the model begins producing value. A vendor that already has direct standing with regulators through a rating and advisory organization may have already absorbed much of that work across states.

Then there is time itself. USAA's Sai Vishnubhatla called it the most valuable resource in the room, while Pluschau noted that Falcon is willing to give up some flexibility if it means getting to value faster. The point was simple: the business outcome matters more than owning every part of the solution.

Vishnubhatla added an important caveat. Even when an insurer chooses to buy or partner, the investment should still leave the organization stronger. The internal team should learn from the work and build new capabilities along the way. A partner that does all the thinking may deliver a solution, but it does not necessarily make the carrier more capable. That is another reason USAA prefers the word ally.

Dhuvur brought the discussion back to the core question: is this capability central enough to the product to justify building it yourself? If not, the fastest path to value may be to partner rather than build.

What should insurers demand from an AI vendor?

Honeycomb's Itai Ben-Zaken framed the decision around a useful question: what can we afford not to do on our own? The answer helps insurers preserve internal resources for the capabilities that truly differentiate them, while looking to partners for areas where ownership creates less strategic value.

But choosing to partner is not the same as simply buying a license. Much of the discussion focused on what a durable AI partnership actually requires, particularly when the technology becomes embedded in critical business processes.

Start simple and build trust over time. ZestyAI's Kumar Dhuvur recommended beginning with straightforward, API-based integrations, allowing both the value of the partnership and the trust behind it to grow over time.

Keep the roadmaps aligned. Falcon's Richard Pluschau highlighted the risk of a carrier and vendor gradually moving in different directions. A vendor may expand broadly while the carrier needs greater depth in a particular area. Avoiding that divergence requires transparency on both sides and regular conversations about where each roadmap is headed.

Test for reliability and plan for continuity. Ben-Zaken pointed to reliability, hallucination rates, business continuity planning, and rigorous testing before AI reaches production. Because replacing an embedded partner can be costly and disruptive, insurers need to evaluate not only how a solution performs today, but how resilient that relationship will be over time.

Protect ownership and access to data. Pluschau emphasized the contractual protections that become critical in a long-term partnership, including concentration risk, code escrow, and clear data transfer agreements. As he put it, "That's our inventory." If a carrier ever needs to move to another provider, its ability to retain and transfer that data matters enormously.

Finally, a true partner should be willing to share accountability for the deployment. USAA's Sai Vishnubhatla argued that vendors should co-own the risk rather than simply deliver the technology and step away. That extends to governance as well: it should be built into the relationship and the product from the beginning, not treated as a compliance exercise added later.

That distinction — between providing technology and sharing responsibility for its success — is ultimately what separates an AI vendor from an AI partner.

Where does AI differentiation actually come from?

A recurring theme across the discussion was that differentiation does not come from the model alone. It comes from how effectively the organization is set up to use it.

USAA's Sai Vishnubhatla made that point through the lens of the operating model. USAA can build or buy almost anything, he noted, but the more important question is whether the organization is ready to put that technology to work effectively. Without the right processes around it, even a strong model can end up layered onto an inefficient workflow rather than improving it.

Falcon's Richard Pluschau made a similar point with a familiar hierarchy: people, process, and technology — in that order. Technology comes last because the value of the tool depends on the organization around it.

That aligns closely with what we see at ZestyAI. The carriers that get the most value from property risk models are not simply the ones with strong data science teams. They are the ones that have changed how underwriters, agents, and pricing actuaries use risk insights in day-to-day decisions. The model matters, but the operating model determines how much value it creates.

What are insurers missing in AI vendor evaluation?

One of the final points raised in the discussion was that many insurers still evaluate AI too broadly. ZestAI's Kumar Dhuvur noted that carriers often use a single AI questionnaire for technologies as different as computer vision, machine learning, and generative AI, even though each comes with different considerations.

That means the questions should change depending on the technology: where and how it is being used, which external stakeholders may care about that use, and how the vendor will adapt as regulatory and governance expectations evolve.

It is a useful way to frame the broader build, buy, or partner decision. The question is not simply who writes the code. It is which capabilities and risks an insurer wants to own, which it is better positioned to share, and whether the organization is ready to put the technology to work.

FAQ: Build, buy, or partner for AI in insurance

Should insurers build, buy, or partner for AI?
Insurers should generally reserve in-house development for capabilities that are strategically important enough to justify the talent, maintenance, governance, and regulatory burden that comes with ownership. For other capabilities, buying or partnering can provide faster access to specialized expertise, proven technology, and ongoing support without requiring the insurer to build and maintain everything internally.

What is the difference between buying AI and partnering with an AI vendor?
Buying typically means acquiring access to a technology or capability. A strong partnership goes further, with shared roadmaps, ongoing collaboration, knowledge transfer, governance support, and greater accountability for how the technology performs in practice. For insurers, that can be especially important when AI becomes embedded in underwriting, pricing, or other critical decisions.

What costs are often missed in an AI build-vs.-buy analysis?
Development cost is only part of the equation. Insurers also need to account for ongoing model maintenance, data governance, support and service levels, specialized talent, regulatory filings, and the opportunity cost of internal teams. A partner with established infrastructure, domain expertise, and regulatory experience may already have absorbed many of those costs.

How can regulatory approval affect an in-house AI model?
For models used in rating or underwriting, regulatory review can materially extend the path to value. Insurers should factor in not only development time, but also the work required to support filings, explain model methodology, respond to regulators, and maintain approvals across states. Working with a partner that already has regulatory experience can reduce that burden and accelerate deployment.

What should insurers look for in an AI partner?
Insurers should look beyond technical performance alone. A strong partner should offer reliability, clear data ownership, business continuity planning, roadmap transparency, governance support, and the ability to evolve as regulatory and business requirements change. The strongest relationships also help the insurer build internal capability rather than simply outsourcing the work.

Where does AI differentiation come from in insurance?
The model itself is only part of the advantage. Differentiation comes from how effectively an insurer integrates AI into its people, processes, governance, and decision-making. The right technology partner can help accelerate that integration while allowing the insurer to focus internal resources on the capabilities that are most strategic to the business.

The build, buy, or partner decision ultimately comes down to where an insurer creates the most value by owning a capability itself — and where the right partner can accelerate progress.  

ZestyAI's Risk Decision Platform brings together property-level data, AI models, and agentic AI capabilities to help insurers make faster, more informed decisions across underwriting, pricing, risk management, and regulatory strategy.

See how insurers are putting ZestyAI to work.

ZestyAI today announced that its risk models have surpassed 500 regulatory approvals across 317 carrier rate and rule filings as its regulatory footprint has expanded to 44 states. The approvals span perils responsible for the vast majority of U.S. property losses, including wildfire, severe storm, hail, non-weather fire, and non-weather water.

The milestone comes as insurers and regulators work to maintain coverage availability in markets increasingly affected by weather volatility. By assessing individual properties rather than broader geographic averages, ZestyAI’s models can recognize mitigation, align price more closely with risk, and give carriers and regulators greater visibility into how that risk is assessed.

Each approval represents a state Department of Insurance reviewing a ZestyAI model within a carrier filing and accepting its use in rating or underwriting. The approvals span personal and commercial property lines and reflect both new adoption and carriers expanding ZestyAI models into additional states, perils, and programs.

They include the first AI model approved as part of a carrier rate filing for wildfire in California, where property-specific intelligence has since supported new capacity in a market where homeowners have faced constrained options. More than half of the top 100 U.S. P&C insurers now rely on ZestyAI across underwriting, rating, reinsurance, and regulatory strategy.

“Five hundred approvals represent the industry deciding, one rate filing at a time, that AI belongs in how insurance risk is underwritten, priced, and reinsured,” said Kumar Dhuvur, Founder and Chief Product Officer at ZestyAI.

“We built these models to be examined. Every score can be explained, every input can be traced, and a homeowner’s mitigation shows up in the result. That is what regulators expect from AI in insurance, and it is what we deliver.”

ZestyAI works alongside regulators as the standards for AI in insurance take shape. The company participates in National Association of Insurance Commissioners (NAIC) proceedings and has presented to regulators on how AI models are built, validated, and explained. It works directly with state departments of insurance as their expectations for fairness, transparency, and consumer protection evolve, and it files its models proactively through its in-house rating and advisory organizations so regulators review them before any carrier puts them to work.

Every ZestyAI risk score can be traced to the property characteristics behind it, including mitigation actions taken by the policyholder. That gives the insurance value chain a consistent view of risk: the same intelligence a carrier uses to price a policy can be understood by the regulator reviewing the filing and the reinsurer evaluating the portfolio.

Risk Theory and ZestyAI today announced a partnership supporting Jupiter Platinum Home, Risk Theory’s new primary excess & surplus (E&S) homeowners program for high-net-worth properties in California. Built for difficult-to-place homes, including those with significant wildfire exposure, the program uses ZestyAI’s Z-FIRE™ for individual property-level wildfire risk assessment, bringing greater precision to underwriting and risk selection.

The program addresses a persistent coverage gap for California homeowners whose properties fall outside the appetite of the admitted market or other E&S products, including those facing wildfire-related non-renewals. Coverage is available for homes with dwelling limits starting at $750,000 and up to $25 million in total insured value.

Traditional wildfire scoring approaches often rely on broad geographic or territory-level factors, making it difficult to distinguish how risk varies from one property to the next. Z-FIRE delivers a more granular view by applying machine learning to evaluate defensible space, vegetation proximity, topography, building materials, and surrounding fire behavior patterns at the individual structure level – combining hazard and vulnerability into a single, unified risk assessment.

Jessica Furrow-Young, Executive Vice President of Underwriting at Risk Theory, said:

“Jupiter Platinum Home was built for risks that are increasingly difficult to place in California, including homes with elevated wildfire exposure. For us, that means understanding the differences between individual properties rather than treating wildfire exposure as a simple yes-or-no decision. Z-FIRE gives our underwriters insight into both the hazard surrounding a home and the characteristics that influence how that individual structure may perform, helping us deploy capacity thoughtfully while delivering a strong solution for our agency partners and their clients.”

Attila Toth, Founder and CEO of ZestyAI, said: 

“Expanding capacity in California requires more than simply taking on risks the traditional market has pulled back from. It requires the ability to differentiate those risks with precision. Risk Theory built property-level risk selection into its underwriting approach from the outset. That level of insight can help insurers approach even challenging wildfire exposures with greater confidence and discipline.”

Z-FIRE is approved across all Western wildfire markets and was the first AI-based wildfire model approved as part of a carrier rate filing in California. ZestyAI’s broader portfolio of risk models has secured more than 200 regulatory approvals nationwide.

California's homeowners insurance market is showing signs of recovery after three years of retreat. The California FAIR Plan added roughly 12,000 residential policies in Q2 2026, its fifth consecutive quarter of declining additions, down from more than 54,000 new policies in a single quarter at the peak in early 2025. Nine insurers, including six of California's 10 largest home insurance groups, have now committed to expand in the state.

Capacity is returning. The more important question is whether insurers now have the risk intelligence to make that capacity durable.

For carriers, that means being able to distinguish wildfire risk property by property, account for both wildland fire and urban conflagration, incorporate mitigation into risk decisions, and support those decisions with models that can withstand regulatory scrutiny.

ZestyAI Founder and CEO Attila Toth made the full argument for this recovery, and what it will take to sustain it, in a bylined viewpoint for Insurance Journal. Read Attila's perspective in Insurance Journal.

Here, we look more closely at the risk-modeling shift behind that recovery and what it could mean for catastrophe-exposed insurance markets beyond California.

Key takeaways

• California FAIR Plan residential additions fell from more than 54,000 per quarter at the early-2025 peak to roughly 12,000 in Q2 2026: five consecutive quarters of decline.
• Nine insurers, including six of California's 10 largest home insurance groups, have committed to expand in the state.
• More than 91% of the area burned in the January 2025 Eaton and Palisades fires had been rated High or Very High risk by ZestyAI's Z-FIRE™ before the fires occurred.
• Durable capacity depends on property-level, regulator-approved risk intelligence, not a return to the risk assumptions that failed before.

Is homeowners insurance capacity returning to California?

Yes. Several indicators suggest that California's homeowners insurance market is beginning to stabilize.

The clearest signal comes from the California FAIR Plan. Residential policy additions have now declined for five consecutive quarters, while commercial policies in force have fallen for two consecutive quarters. At the same time, major carriers are making commitments to expand their California books.

Travelers announced its expansion in April 2026. Farmers removed its policy cap in late 2025. Mercury committed to writing 38,000 new policies, including in distressed areas. The recovery is early and uneven, but the leading indicators, FAIR Plan additions, carrier commitments, actual policy counts, now point toward expansion rather than retreat.

Why did insurers pull back from California in the first place?

The pullback had been building for years. Major carriers had paused or restricted new homeowners business in California well before January 2025, after a decade of wildfire losses kept exceeding what their risk views had priced for.

The January 2025 Eaton and Palisades fires made the gap especially visible. The fires destroyed nearly 16,000 structures, took 31 lives, and produced roughly $40 billion in insured losses.

They also exposed a fundamental modeling challenge: wildfire risk is no longer only a wildland problem.

Eaton and Palisades became urban conflagration events, with fire spreading from structure to structure through dense residential areas. Risk models built primarily around vegetation, terrain, and historical wildfire boundaries can miss important elements of how fire propagates through the built environment.

For insurers, the consequence is straightforward. If the risk view does not adequately differentiate the properties most likely to experience loss, carriers have fewer tools to determine where they can responsibly write, how they should price risk, and where mitigation can meaningfully change an outcome.

Better insurance availability therefore depends in part on better risk selection.

What has changed in California wildfire risk modeling?

One of the most important changes is regulatory.

Under California Commissioner Ricardo Lara's Sustainable Insurance Strategy, insurers can now use forward-looking catastrophe models to inform catastrophe loads in rate filings, coupled with commitments to increase coverage in distressed areas.

That creates a path for carriers to base decisions on a more current and granular view of catastrophe risk rather than relying exclusively on historical loss experience.

California had already demonstrated that modern AI models could be incorporated into the regulatory process. ZestyAI's Z-FIRE™ became the first AI-powered wildfire model approved as part of a California carrier rate filing and has since been approved in additional carrier filings.

That regulatory history matters because sophisticated models are only useful to insurers if they can also be explained and defended.

A modern risk model must do more than produce a score. Carriers need to understand the factors influencing risk, reproduce results, document their methodology, and establish an audit trail that regulators and other stakeholders can evaluate.

Why does property-level wildfire risk matter?

Wildfire exposure can change dramatically from one property to the next, even within the same neighborhood.

Property characteristics such as roof condition, surrounding vegetation, defensible space, nearby structures, building density, topography, and other characteristics of the built and natural environment can influence how a property responds when a wildfire reaches a community.

Property-level models are designed to identify those differences rather than assigning the same broad risk view across an entire ZIP code, territory, or hazard zone.

That distinction becomes particularly important when insurers are trying to expand selectively into areas they previously considered too difficult to write.

With greater granularity, a carrier can identify properties where the underlying risk is more favorable, distinguish higher-risk structures from their neighbors, recognize mitigation, and make underwriting and pricing decisions at the level at which the risk actually exists.

What did Z-FIRE identify before the Eaton and Palisades fires?

ZestyAI's analysis of the Eaton and Palisades fire footprints illustrates what a property-level approach can reveal.

More than 91% of the area that ultimately burned had been classified as High or Very High risk by Z-FIRE before the fires occurred.

Within the burned area, properties in Z-FIRE's highest risk tier were roughly 50% more likely to be destroyed than properties in its lowest risk tier.

Those results matter for more than retrospective validation. They illustrate the kind of segmentation insurers need before an event occurs: a view of risk granular enough to differentiate individual properties and identify meaningful variation within the same catastrophe-exposed market.

How are insurers using property-level wildfire models in California?

Property-level wildfire intelligence is already being used across a significant share of the California market.

More than 40% of California homeowners premium is written by carriers using ZestyAI's view of wildfire risk. The California FAIR Plan uses Z-FIRE in its rating, and carriers are applying the model to support underwriting, rating, portfolio management, mitigation, and regulatory strategies.

New market entrants are also using property-level intelligence as part of their California strategy. Windward Risk Managers, DUAL North America, and Kingstone Companies are all underwriting California wildfire risk with Z-FIRE.

Windward Risk Managers, for example, entered the California homeowners market using Z-FIRE to evaluate wildfire risk at the individual-property level. Its expansion illustrates one way additional capacity can return: not by assuming that an entire distressed market has become less risky, but by identifying where risk can be understood and selected more precisely.

Only 11% of California homes fall into Z-FIRE's High Risk wildfire tier, and they are scattered through the same territories as everyone else. Territory-level rating and underwriting averages them together with their lower-risk neighbors. Property-level scoring separates them.

How can better wildfire models support insurance availability?

Better risk models cannot create capacity on their own. Rates, reinsurance, regulation, capital, claims experience, and market strategy all influence an insurer's willingness to write.

But risk intelligence affects nearly every one of those decisions.

A more precise view of risk can help insurers identify insurable properties within high-hazard markets, differentiate risk within a portfolio, recognize mitigation, communicate portfolio quality to reinsurers, and support rating and underwriting approaches with documented evidence.

This creates an alternative to the bluntest form of catastrophe management: withdrawing from entire territories because a carrier cannot confidently distinguish the risks within them.

For homeowners, that distinction can ultimately affect whether a property is considered writable at all.

What can other catastrophe-exposed states learn from California?

California may be the most visible example of the insurance industry adapting to changing catastrophe risk, but it is not the only market facing the problem.

Wildfire, severe convective storms, wind, hail, and other perils are creating similar pressure across the United States. As loss patterns change, insurers increasingly need risk views capable of differentiating exposure at the individual-property level rather than relying primarily on historical averages and broad geographic proxies.

The broader lesson from California is therefore not simply that capacity can come back after a period of retreat.

It is that sustainable capacity requires insurers to understand risk differently.

The models used to decide which properties to write, how to price them, how mitigation changes risk, and how portfolios will perform under future catastrophes have to keep pace with the risk itself.

California is beginning to test what that looks like in practice.

Frequently asked questions

What is Z-FIRE?

Z-FIRE™ is ZestyAI's AI-powered wildfire risk model. It scores wildfire risk at the individual property level for every address in the U.S., covering both wildland fire and urban conflagration. The model produces two scores: a neighborhood score that measures the likelihood a property will fall inside a future wildfire perimeter, and a property score that measures the likelihood the structure ignites, based on roof characteristics, defensible space, vegetation, topography, and surrounding building density. Z-FIRE is trained on the industry's largest wildfire loss database, spanning more than 2,000 fire events over 20 years, and incorporates the latest structure-ignition and materials research from the Insurance Institute for Business & Home Safety (IBHS). It outperforms traditional regional models by 44X in predicting property-level risk and is approved by regulators for rating and underwriting in every wildfire-prone state in the western U.S.

Is Z-FIRE approved for use in California?‍

Z-FIRE was the first AI-powered wildfire model approved as part of a California carrier rate filing and has subsequently been approved in additional carrier filings. ZestyAI's models have earned hundreds of regulatory approvals nationwide.

Does the California FAIR Plan use Z-FIRE?‍

Yes. The California FAIR Plan uses Z-FIRE as part of its rating approach to evaluate wildfire risk at the property level.

How is property-level wildfire modeling different from traditional wildfire hazard maps?‍

Traditional hazard views often classify risk across larger geographic areas based heavily on characteristics such as vegetation, terrain, and historical fire activity. Property-level models evaluate risk at the individual address and can incorporate characteristics of the structure, surrounding environment, nearby buildings, and mitigation to differentiate risk between neighboring properties.

Why are catastrophe models important to California's insurance recovery?‍

California insurers need to understand where they can responsibly expand while continuing to manage severe wildfire exposure. More granular, forward-looking catastrophe models can help carriers differentiate risk, recognize mitigation, manage portfolios, support regulatory filings, and identify properties that may be writable even within high-hazard areas.

A more precise path back into catastrophe-exposed markets

California's insurance recovery is still in its early stages. Rate filings still take time. Reinsurance capacity still matters. Another severe wildfire season could test the progress that has been made.

But the market is moving in a different direction than it was three years ago.

The next phase will depend not only on how much capacity returns, but on whether insurers have better tools for deciding where that capacity goes.

For California — and for every catastrophe-exposed insurance market watching it — more precise, property-level risk intelligence is increasingly becoming part of that answer.