For decades the insurer priced your ZIP code. This year it started pricing your building.

Walk any commercial block and pick two shops that look the same to an insurer: same street, same ZIP code, same line of business. To the typical insurance pricing engine they are one and the same risk, charged the same blended rate to transfer that same-looking risk, even though one has a hardened roof and a clear yard and the other has neither. For more than a hundred years the insurance carrier had no cost effective way to tell these risks apart, so it did not try. This week, in three separate corners of the insurance market, the tools that finally tell one building from the next moved out of the pilot deck and into the machinery that sets prices. That is the whole thing we are building, arriving from the other direction.

The advanced analytic tools and aerial imagery data sets that tell one building from the next (from the sky) just moved from pilot decks into insurance rate filings. The careful operator is the one these tools are built to find, and the careless neighbor is the one who has been getting a discount on the careful operator's dime.

The unit of pricing is shrinking

In August, FM bought FortressFire, a firm that combines physics-based fire modeling with machine learning to judge whether one specific structure will ignite, not whether its neighborhood might. FM's own read on the deal is blunt: traditional carriers use coarse geographic segmentation, but what you want is structure-level modeling that prices the individual property risk. The gap between the old way and the new is widening into an adverse-selection problem for anyone still pricing the map, not the roof. This is not a science project. More than 20 states have already accepted AI-driven wildfire models in rate filings, per the Insurance Institute for Business and Home Safety and the American Property Casualty Insurance Association. The money is there because the losses are: wildfires and severe storms were 83% of 2025's $107 billion in insured catastrophe losses, by Swiss Re's count.

Granularity cuts both ways

A model that can reward the hardened building that has proactively prepared for, and mitigated, wildfire loss risk can surcharge the exposed one, and it will. The insurance carrier that can see and compare the individual structures takes the good risks at a fair price and leaves the blended-rate carrier holding the rest, which is what adverse selection means in practice. The old bucket was comfortably profitable (until it wasn't) precisely because it hid something Friend sees again and again: the careful operator, with a well managed property, subsidized the careless one, and neither could tell. Structure-level pricing ends the subsidy. It is good news for the operator who has done the work and a bill for the one who has not, and for the first time the carrier can act on the difference.

The arithmetic, on one building

Take a bakery in a wildfire-exposed county with an $8,000 premium property policy. Priced by ZIP code, it pays the territory average, the same as the shop two doors down with a wood-shake roof and brush to the wall. Priced by structure, with a Class-A roof and a cleared five-foot zone, the same bakery looks more like a $5,600 risk, while the careless neighbor looks more like $10,400. The $2,400 the careful owner was overpaying did not go anywhere useful. It quietly covered the neighbor. The data and the model that separate them are not expensive to run. It has simply, until now, not been run.

The industry is finally paying to tell one building from the next, peril by peril, and every place it does, the careful operator stops subsidizing the careless one. The read was always the missing piece. It is starting to get built.