Eleven point eight million California homes, one AI model, and a finding that should unsettle everyone who has ever received a nonrenewal letter from their insurer: roughly $1 trillion worth of properties that FEMA classifies as "low risk" actually face elevated wildfire danger when you bother to look at the individual lot instead of the surrounding zip code.
ZestyAI ran those numbers using their Z-FIRE model, which ingests satellite imagery, measures vegetation density within specific distance bands around each structure, catalogs roof materials, grades topographic slope and aspect, and cross-references suppression response times against a training set of more than 2,000 historical wildfires. Over one-third of California's insurance market now uses Z-FIRE for underwriting decisions, which means the algorithm that determines whether you can insure your home is looking at your dead juniper hedge while FEMA is still looking at a polygon drawn around your county.
Who Insures What Nobody Else Will
Delos Insurance announced in April 2026 that it expanded wildfire coverage eligibility to 12 million California properties, adding roughly 1 million homes in areas traditional carriers had abandoned. Delos attributes this expansion to refined wind data modeling and suppression efficacy analysis powered by Google Cloud. Most striking: the company reports zero fire losses to date across its entire California book, a result that would be statistically implausible under zone-based underwriting but becomes explicable when the model distinguishes between a stucco home with cleared brush and a wood-sided home with junipers touching the eaves on the same street.
Compare that track record to the state of the traditional market. California homeowner insurance costs have climbed 41 percent since 2020. Enrollment in the FAIR Plan, the insurer of last resort that exists because nobody else will write the policy, surged 43 percent as carriers like State Farm and Allstate pulled back from wildfire-prone regions. The FAIR Plan itself needed an additional $1 billion just to cover claims from the January 2025 Los Angeles wildfires, which caused total losses estimated between $150 billion and $275 billion according to Delos and ROAR Partners' joint analysis.
A 43 percent spike in last-resort enrollment is not a market correction but a market failure, and AI underwriters are filling the vacuum by doing something traditional actuarial tables never attempted: scoring the property instead of the neighborhood.
What the Algorithm Actually Sees
If you are building a home in fire-prone territory, or buying one, the specific choices that now carry direct insurance consequences through AI scoring are more granular than most buyers realize. Six variables dominate property-level wildfire models, and every one of them is something a builder or homeowner can control.
Roof material. Class A fire-rated surfaces like concrete tile, standing-seam metal, or asphalt composite score lowest risk, while wood shake roofs score highest. A reroofing project that costs $15,000 to $25,000 can shift your AI risk category enough to move you from FAIR Plan territory into the admitted market, where premiums run 40 to 60 percent lower.
Defensible space. California already mandates 100 feet of defensible space clearance under Public Resources Code 4291, but AI models measure compliance from satellite imagery rather than relying on self-reported inspections. Zone 1 (0 to 30 feet from the structure) must be essentially combustible-free. Zone 2 (30 to 100 feet) requires reduced fuel loads with adequate spacing between tree canopies. Models like Z-FIRE capture whether you actually did the work or just filed the paperwork.
Building materials. Fiber cement siding, stucco, and brick outperform wood lap siding in every model. Vent screening matters enormously: 1/8-inch mesh on attic and soffit vents prevents ember intrusion, and the presence or absence of that $200 upgrade is visible in high-resolution satellite imagery when analysts know where to look.
Vegetation proximity, deck materials, and topographic position round out the scoring. Composite decking versus wood. Tree canopy distance from the roofline. Whether your lot sits on a slope that channels wind uphill toward your structure or on a ridgeline with exposure on three sides. None of these factors appear in a FEMA flood zone map, and none of them appear in the crude wildfire hazard severity zones that California's Office of the State Fire Marshal last updated comprehensively in 2007.
Novel Calculation: The Mislabeling Rate
ZestyAI's finding that $1 trillion in "low risk" California properties face elevated wildfire danger deserves context. California's total residential real estate value sits around $10 trillion. If the mislabeled properties represent roughly 8 to 10 percent of the state's housing stock by value, and if those properties carry insurance priced for the FEMA-assigned "low risk" classification rather than their actual exposure, then every major wildfire event in those areas generates claims that the premium pool was never designed to absorb. Insurance is a bet on accurate risk classification. Mispricing 8 percent of a $10 trillion portfolio is how you get a FAIR Plan that needs emergency billion-dollar infusions after a single fire season.
Why This Should Worry You Anyway
AI underwriting models are opaque. Homeowners cannot see their property-level risk score, cannot understand which specific variable pushed them into a higher tier, and have no appeals process when an algorithm decides their vegetation clearance falls short of whatever threshold the model learned from its training data. Older homes with deferred maintenance and properties in lower-income areas where landscaping budgets are tighter may be systematically scored as higher risk regardless of the owner's actual fire preparedness practices, because the model sees the current state of the roof and the yard, not the owner's intentions or financial constraints.
There is a deeper problem. Every one of these models was trained on historical wildfire behavior. The Lahaina fire in August 2023 killed 101 people in a town that had no meaningful historical wildfire precedent. Urban conflagrations driven by extreme wind events and densely packed structures behave differently from the brush fires and forest fires that dominate the training data, and a model that scores a property as "low risk" because no fire has burned there before may be encoding the same blind spot that FEMA's zone maps already suffer from, just with more impressive satellite imagery and a higher confidence score.
If You Are Building in Fire Country
Specify Class A roofing, fiber cement or stucco siding, 1/8-inch ember-resistant vent mesh, and composite decking from the start. These choices add $8,000 to $15,000 to a typical 2,000-square-foot home and will likely pay for themselves within three years through lower premiums in AI-scored underwriting markets. Design defensible space into the landscape plan before the first grading permit, not as an afterthought when the insurance quote arrives. And download WyldSafe, a free AI-powered app that uses computer vision to assess your property's wildfire vulnerabilities and identify specific improvements that could shift your insurability.
Kingstone Insurance entered California in Q2 2026 using ZestyAI's Z-FIRE model for underwriting. Farmers Insurance introduced a 22 percent discount for wildfire mitigation improvements effective September 2026. Delos continues expanding through 14,000 broker partners statewide. New insurers are entering this market specifically because AI lets them see risk that traditional models could not price. If you build to the standard that these algorithms reward, you will have more carriers competing for your policy, not fewer.
Your zip code is a legacy artifact. Your roof is the new underwriting document.
Limitations
Delos Insurance's "zero fire losses" claim cannot be independently verified from public filings. ZestyAI's $1 trillion mislabeling figure relies on proprietary methodology that has not been externally audited. Specific model weights for roof materials versus vegetation versus topography are not publicly disclosed by any AI underwriting vendor. FAIR Plan enrollment increases likely reflect multiple factors beyond carrier retreat, including population growth in wildland-urban interface areas and increased awareness of wildfire risk after the 2025 LA fires. No longitudinal data exists to confirm whether AI-scored underwriting outperforms traditional actuarial methods over a complete multi-decade fire cycle. Premium savings estimates for fire-resistant construction are based on admitted-market versus FAIR Plan rate comparisons and will vary by carrier, location, and policy terms.