How AI Wildfire Models Price California Home Insurance

There's a flat, overhead photo of your roof. Taken on some Tuesday you don't remember. And in a data centre somewhere, a model is studying it... the trees around the house, the slope of the street... and deciding what you pay to insure the place. Or whether anyone will insure it at all.

And that's a thing that exists? Not a pitch deck?

California's insurer of last resort, the FAIR Plan, has been running an AI wildfire model since twenty twenty-one. It's called Z-FIRE, from a company called ZestyAI. Last June they signed a four-year extension.

So what's it actually looking at? Be specific.

More than two thousand historical wildfires. Satellite and aerial imagery. Topography. And then your specific property... not your zip code, your parcel. When the FAIR Plan announced it, they said, quote, we can no longer rely on thirty-year-old technology to understand modern climate phenomena.

Which is a polite way of saying the old maps were wrong.

Coarse, more than wrong. Hurricanes and earthquakes, insurers have modelled for decades. Fire is messier. Vegetation, wind, human intervention... and the flight path of one burning ember.

An ember. The model has to guess where a single spark lands.

Essentially. And hold onto that word, precision, because by the end of this it stops being good news. First, the number that explains why anyone bothered. The FAIR Plan's exposure... the total value it's on the hook for... was one hundred fifty-three billion dollars in twenty twenty. By September twenty twenty-four, four hundred fifty-eight billion.

Wait. Triple? In four years?

Policies in force more than doubled. Two hundred three thousand, up to four hundred fifty-two thousand. That's the safety net catching everyone the private market dropped.

So the AI didn't push those people out. It showed up after.

Both things are true at once, and that's the honest tension here. The case for the models is that precision is what lets insurers come back. If all you can see is a mountainside, you write off the mountainside. If you can see one house with a cleared yard and a fire-resistant roof... you can price that house.

Okay. That's a better argument than I expected.

And Z-FIRE does credit mitigation. Fuel management, fire-resistant materials. California's insurance department had been pushing insurers to assess risk property by property, and this was the answer. There's even an MGA called Delos writing policies in fire-prone California, using a model built with the Spatial Informatics Group, a think tank that's worked with Cal Fire for decades.

Their whole business is going where nobody else will go.

Their chief data officer, Shanna McIntyre, describes it as exclusive access to that research for insurance purposes. The bet being: good risks exist inside bad neighbourhoods, and the machine can find them.

Then the same machine finds the bad risks inside the good neighbourhoods.

Nobody markets that half. A model precise enough to rescue your neighbour is precise enough to single you out. And you may never see the image, or the reasoning.

Does any of this help people after a fire? Or is it purely a pricing tool?

No, this part is genuinely good. After the Los Angeles County wildfires in twenty twenty-five, Moody's ran AI image analysis on before-and-after satellite and aerial pictures, sorting structures into destroyed, partially damaged, untouched.

From orbit. Before an adjuster can get down the road.

And they trained it to tell a house from a shed. Sounds trivial. It isn't. If a parcel has a home, a garage and a workshop, you can double-count losses badly. Getting it right means faster, cleaner claims.

Huh. Same technology doing the cold thing and the kind thing.

And it's spreading past fire. Insured losses from severe convective storms... hail, thunderstorms... hit sixty-one billion dollars in twenty twenty-four, second-highest on record, according to Aon. Texas has more hail claims than any state. Texas approved an AI storm model from ZestyAI.

Let me push back. Are these models actually better? Or just more confident?

Fair, and the evidence is real but narrow. The FAIR Plan hired the actuarial firm Milliman to test whether Z-FIRE could genuinely separate the best risks from the worst. Milliman said it outperformed the alternatives by a wide margin. But flood models built to measure the same risk still disagree with each other. Nobody has solved uncertainty. They've measured it more finely.

Here's what I can't shake. Insurance was supposed to be pooling. Everyone pitches in, the unlucky one gets covered.

And that's where precision stops being good news. Perfect prediction is the end of pooling. If a model knows exactly which house burns, there's nothing left to share. The question isn't whether the AI works. It's what happens to the deal when it does.

And none of us get a vote on the picture of our roof.

Not yet. So when your renewal lands and the number has jumped... it may be that no person decided that. A model looked down at your house, counted your trees, and moved you out of the pool.