An Algorithm Graded 110 Million Roofs from Satellite Photos. Your Insurer Used the Score. You Never Saw It.
Linda Bennett got a letter from State Farm. Not a bill, not a premium adjustment, not a courtesy notice about coverage changes ahead of renewal season. An ultimatum: replace your roof within the specified timeline or lose coverage entirely, at an estimated cost of $28,290.
Bennett had lived in her Santa Ana, California home since 1993, three decades and change without a roofing crisis, without a leak, without a claim filed against the very policy now threatening to vanish. No inspector ever climbed her roof. No adjuster knocked on her door, peered through binoculars, or parked a truck in her driveway. State Farm evaluated her roof remotely, likely through aerial or satellite imagery captured from altitude, and rendered its verdict without setting foot on the property.
Her situation is not unusual. It is rapidly becoming the default experience for homeowners across the country who discover their insurer has been watching their roof from space and making coverage decisions on what it saw.
How It Works
A company called Cape Analytics, now owned by Moody's, has built an AI system that has scored the roof condition of more than 110 million residential buildings across the United States, a figure representing 98 percent of the country's housing stock. Aerial and satellite imagery flows into convolutional neural networks trained on labeled roof defects: missing shingles, tarps draped over damage, discoloration patterns that signal aging, material degradation visible only from above, ponding water pooling on flat sections, amateur patching jobs that suggest deferred maintenance.
More than 100 insurance carriers have adopted Cape's Roof Condition Rating system, and over 300 regulatory filings in more than 40 states authorize its use in underwriting decisions, a scale of deployment that makes it effectively invisible precisely because it is so universal. Cape pioneered this capability in 2016 and 2017, covered the continental United States by 2018, and demonstrated by 2019 that its AI-derived roof scores could predict insurance losses more accurately than homeowner-reported roof age, which had been the industry's primary metric for decades.
Cape is the largest player in a crowded field, but it is far from alone. Nearmap, Verisk, and Betterview all offer competing geospatial property intelligence products that promise insurers faster, cheaper, more objective evaluations than a human inspector could deliver in a week of windshield surveys and ladder climbs. A consortium called the Geospatial Insurance Consortium provides aerial imagery infrastructure to carriers directly, partnering with a firm called Vexcel for AI-based analysis of the photographs it captures. What emerges is a technology supply chain that is deep, well-funded, operationally mature, and almost entirely invisible to the people whose homes it evaluates every time a satellite passes overhead or a fixed-wing aircraft completes its grid pattern.
When the Algorithm Gets It Wrong
In Texas, Alaina Callahan received a demand from her insurer to replace her roof or face non-renewal, a demand that might have been routine if her roof weren't five years old and in demonstrably good condition. She asked to see the aerial images the company had used to reach its conclusion, the specific photographs and algorithmic scores that had triggered a decision with thousands of dollars in consequences for her household.
She never got to see them. "I had no recourse as a homeowner," Callahan told NPR. "None whatsoever."
In California, Cindy Picos was dropped by CSAA Insurance after an aerial photograph allegedly showed her roof was failing, a determination made without anyone visiting the property, climbing the structure, or asking the homeowner whether recent repairs had been completed. She commissioned an independent, in-person inspection. Her inspector concluded the roof had at least ten more years of useful life, a finding that contradicted the algorithmic assessment by a full decade of projected service. CSAA did not reverse its decision.
Since 2023, at least a dozen homeowners have filed formal complaints with the Texas Department of Insurance over the use of aerial photos in underwriting decisions, and in filings with the state, five insurers operating in Texas confirmed they use third-party aerial photographs processed by AI to evaluate properties. In one documented case, an insurer based its non-renewal decision on imagery of the wrong property entirely.
These are the cases that generated complaints, the ones with homeowners motivated enough to navigate a state regulatory process. Countless others simply accepted the non-renewal letter and quietly moved to a more expensive policy, or went uninsured, or let the coverage lapse because the replacement they couldn't afford was for damage that may not have existed in the first place.
A $1.4 Billion Parallel
Beyond aerial imagery, a lawsuit in Oklahoma is revealing a separate but philosophically aligned initiative that raises similar questions about what happens when insurers deploy sophisticated analytical tools against policyholders without disclosure. A Tulsa police officer named Hursh filed suit after State Farm denied his hail damage roof claim, and the case, Hursh v. State Farm, alleges that State Farm hired Accenture to construct something called the "Fire Model Enhancement" program, which plaintiffs describe as a system of artificial industry standards engineered specifically to justify denying full roof replacement claims, even when policyholders had purchased replacement cost coverage that entitled them to exactly that.
According to the plaintiffs' analysis, State Farm reduced its wind and hail claim payouts by approximately $1.4 billion in 2021, the first year it implemented the program. State Farm has not publicly confirmed or denied the figure. In late June 2026, the Oklahoma Supreme Court denied State Farm's attempt to block discovery, and Judge Amy Palumbo ordered the company to produce its documents related to the Accenture program, a ruling that may eventually reveal the internal mechanics of how an insurer builds analytical infrastructure to systematically reduce what it pays.
FME and AI aerial scoring are different technologies with different mechanisms: one allegedly creates internal standards to reduce payouts after a claim is filed, while the other automates property evaluation to make underwriting decisions before a claim exists. But they share a structural feature that matters more than any technical distinction. In both cases, the homeowner is the last person to learn what criteria are being applied to their property, and they have no meaningful opportunity to contest the outcome before it reshapes their financial life.
Eight Years Without Explanation
In February 2025, Cape Analytics launched version 5 of its Roof Condition Rating system, and the marquee improvements were "reason codes" and "confidence scores," features that tell the underwriter which specific defect triggered a particular rating and how certain the model is about its assessment. These are genuine improvements that give underwriters and, potentially, homeowners a clearer picture of why a roof received the score it did.
They are also an implicit acknowledgment that for the previous eight years, from 2016 through 2024, the system that scored 110 million roofs and informed hundreds of thousands of underwriting decisions did not routinely provide either one. An insurer could receive a score, act on it by canceling a policy or demanding repairs, and neither the insurer's own underwriter nor the homeowner could easily determine which specific feature of the roof image drove the result or how confident the algorithm was that the defect it identified was real rather than an artifact of shadow, angle, or resolution.
Cape's v5 documentation also notes improvements in handling tree cover and shadow, two known sources of false positives that create apparent defects where none exist. Trees obscure roofs, and shadows from those trees, from chimneys, from adjacent structures, and from the angle of the sun at the moment the satellite passed overhead can all produce visual patterns that a neural network trained on labeled defects may interpret as material degradation, ponding water, or structural damage. These are not edge cases in residential America; they describe the typical condition of most suburban rooftops on a summer afternoon. That the fifth major version of the system is still refining its handling of trees and shadows suggests earlier versions were making consequential errors on properties with mature landscaping, errors that may have triggered non-renewals for homeowners who never knew a shadow was the problem.
No One Wrote the Rules
No federal regulation specifically governs the use of AI-derived property assessments in homeowners insurance underwriting, a gap that matters because the regulatory framework for insurance was built for a world where inspection meant a human being visiting a property, knocking on a door, and forming a judgment that could be questioned, appealed, and contested in the same language both parties spoke. AI aerial scoring breaks that assumption entirely: the evaluation happens at scale, the methodology is proprietary, and the homeowner may never learn it occurred until the non-renewal letter arrives.
"Just because a technological opportunity exists doesn't mean it can or should be used without guardrails and consumer protections," says Doug Heller, Director of Insurance at the Consumer Federation of America. "Using images that were gathered without consumer awareness, or let alone consent, is really problematic."
State Farm told reporters that aerial imagery is "one of several tools" it may use when reviewing a property, including images from manned fixed-wing aircraft, satellites, and, in some cases, an on-site inspection. A statement that is technically accurate and practically meaningless to a homeowner who has just received a non-renewal notice with no attached evidence, no explanation of methodology, and no appeals process that doesn't require them to hire their own inspector and pay out of pocket to prove their roof is fine.
What You Should Do
If you receive a non-renewal notice citing roof condition, do not accept it passively. Request the specific images and data your insurer used to evaluate your property, because while not every state requires the insurer to provide them, making the request creates a paper trail and sometimes produces information that reveals the assessment was based on outdated imagery, the wrong address, or conditions that had already been repaired before the algorithm ever looked.
Commission an independent roof inspection from a licensed contractor or certified home inspector, and insist on a written report with dated photographs that document the roof's actual condition at the time of the dispute. If the independent assessment contradicts the insurer's AI-derived evaluation, submit it to the insurer in writing and simultaneously file a formal complaint with your state's department of insurance, because regulatory agencies track complaint volume by company and by issue type, and sustained complaint patterns are one of the few mechanisms that trigger regulatory scrutiny of underwriting practices that would otherwise go unexamined.
If you are buying a home, ask whether the property has been flagged by any geospatial underwriting platform before you close. Your insurance agent can often run a preliminary check that reveals what the algorithms think of the roof, and a poor score may not mean the roof is actually failing; it may mean a mature oak was casting a shadow in the image the neural network analyzed last August. But whatever the cause, a low score will affect your ability to insure the home and at what price, and discovering that after closing is significantly worse than discovering it during due diligence.
If you are building a new home, document the roof installation with dated photographs from ground level and, if possible, from a consumer drone flying directly above the completed surface. Upload them to a cloud service that timestamps the files with metadata you control. When the AI eventually scores your brand-new roof, and it will, you will have contemporaneous evidence of its actual condition that predates whatever satellite or fixed-wing imagery the algorithm ingests. Think of it as a property appraisal for the algorithm age: proof that what the machine sees from 15,000 feet may not be the whole truth about the structure below.
Limitations
The $1.4 billion figure from the Hursh v. State Farm case is a plaintiff estimate. State Farm has not disclosed its own accounting of FME's financial impact, and the case is in discovery, not at trial. Cape Analytics' error rates for roof condition scoring have not been independently audited by any consumer protection agency or academic institution. The "12+ complaints" figure from the Texas Department of Insurance represents formal filings only and almost certainly understates the prevalence of disputed aerial assessments. We could not determine how many homeowners were non-renewed based on AI aerial scoring nationally in any given year because no regulator collects or publishes that data. State Farm's statements about its use of aerial imagery are drawn from press reports, not from direct responses to questions we submitted. Five Texas insurers confirmed using AI-analyzed aerial photos in state filings, but the extent and methodology of their use varies and has not been independently reviewed.