William May bought his Pacific Palisades home for $1.7 million. When the L.A. wildfires came, State Farm offered $1.35 million. Neighborhood home values had climbed roughly 50 percent since his purchase, according to Zillow's Home Value Index, putting the average Palisades property above $3 million by December 2025. May's insurer valued his destroyed home at less than he originally paid for it.
"They use this reductive method," May told Fast Company. "It's a phony way of calculating every screw, every bolt, and coming up with a profit for State Farm by undervaluing the house."
May blames Xactimate, a Verisk-owned software platform that most major property insurers use to generate repair and rebuild cost estimates. Verisk says AI handles only ancillary tasks like summarizing information and labeling photos, not cost generation, but the underlying database drives the estimates adjusters rely on, and homeowners across California, Alabama, and Illinois have filed lawsuits alleging the software systematically underpays claims. Los Angeles County launched a formal probe into State Farm's AI tools last November, demanding every internal document, memorandum, training manual, and policy directive related to how the company uses artificial intelligence in claims review.
May is not an outlier. A Harvard study covering 100 million mortgages says he is the median.
A Hundred Million Mortgages, One Ugly Number
Parinitha Sastry and co-authors at Harvard Business School, Columbia, and the Federal Reserve published the first large-scale study linking individual mortgage records to insurance policy data, a dataset covering approximately 100 million mortgage-backed homes, and their finding does not leave room for comfortable interpretation: the average American homeowner with a mortgage insures only 70 percent of what it would cost to rebuild their home. Between 2011 and 2020, that ratio dropped from 70 percent to roughly 50 percent, a decline that has almost certainly continued as construction costs kept climbing through the post-pandemic surge in labor and materials prices that shows no sign of reversing.
Construction costs kept climbing while coverage limits stayed flat. The Turner Building Cost Index rose 8 percent in 2022, 6 percent in 2023, 3.9 percent in 2024, and 4.1 percent in 2025, reaching a record high.
In Alabama, median coverage sits at 55 percent of replacement cost, and Mississippi falls below that. Louisiana homeowners pay $7,304 a year in premiums and cover only 58 percent of their rebuild cost, while Massachusetts homeowners pay $1,647 and cover 84 percent, a pattern so perverse it deserves its own name: the states where homeowners pay the most for insurance have the worst coverage ratios, because high premiums force people to reduce their coverage limits to keep total housing costs from consuming everything else in their budget. Every 1 percent increase in insurance prices cuts coverage by 0.3 to 0.8 percent, according to the Sastry research, creating a ratchet that tightens whenever insurers raise rates and never loosens when they do not.
Philadelphia Fed researchers tested these numbers against real destruction in their working paper on the 2021 Marshall Fire in Boulder County, Colorado, which examined roughly 5,000 claims from the blaze that destroyed over 1,000 homes. Seventy-four percent of affected homeowners were underinsured. In 36 percent of cases, coverage fell below 75 percent of replacement cost. Only 8 percent carried guaranteed replacement coverage, which is the only policy endorsement that would have fully covered a total loss.
When the Check Falls Short
Displacement, foreclosure, forced sale.
Philadelphia Fed researchers tracked Marshall Fire homeowners and found that every 10-percentage-point increase in underinsurance reduced rebuilding permits filed within a year by four percentage points. Underinsurance alone lowered total rebuilding permits by 25 percent over twelve months, and more than half of all destroyed homes changed hands within eighteen months of the fire, not because homeowners wanted to sell but because the insurance check did not cover the rebuild and they could not bridge the difference with savings, a construction loan, or federal disaster aid that arrives slowly if it arrives at all.
After Hurricanes Harvey and Irma in 2017, Sastry and co-authors tracked mortgage default rates among disaster-affected borrowers. Underinsured homeowners defaulted 38 basis points more often than fully insured ones, a number that sounds abstract until you remember that a basis point is one-hundredth of a percent and that 38 of them, concentrated in the months after a catastrophe, represent thousands of families losing homes they believed were protected by a policy they had been paying into for years.
Most homeowners do not know they are underinsured. Most will not find out until they file a claim, at which point the information arrives as a number on a letter from a company that has already decided what their home is worth.
What Insurers Built While Homeowners Weren't Looking
Insurance companies have not been idle. AI is projected to save the property insurance industry $35.77 billion annually by 2030, cutting processing costs by 50 to 65 percent and claims expenses by 20 to 30 percent, according to industry projections compiled by Storm Law Partners. Bain & Company's research for carriers projects that generative AI will produce a "30 to 50 percent decrease in total leakage," the industry's euphemism for the difference between what is paid and what is contractually owed per the policy terms, a number that shrinks when it catches genuine overpayment and does something altogether different when it trims legitimate claims that an overwhelmed adjuster would have approved on a second look.
CAPE Analytics sells AI-powered aerial and satellite imagery analysis to insurers, deploying computer vision that detects roof condition, tree proximity, swimming pools, trampolines, yard debris, and what it describes as "hundreds of other property-specific risks," all from overhead imagery that homeowners never consented to and cannot inspect for errors. Its own white paper frames the value proposition: helping insurers avoid "excessive exposure when quotes are too low." EagleView launched Horizon in May 2026, an agentic GeoAI engine combining aerial imagery with property intelligence and over 20 AI-powered analysis tools that allow carriers to score, filter, and flag properties at a scale that would require tens of thousands of human adjusters to replicate.
Homeowners encounter the output of these systems as a number on a claims letter, nothing more. A carrier survey by the National Association of Insurance Commissioners found that 88 percent of auto insurers currently use or plan to use AI to evaluate claims, and home insurance adoption is likely comparable, though no equivalent survey has been published. Policyholders never see the model, the training data, the confidence intervals, or the assumptions that produced the estimate they are expected to accept.
Counter-Tools Are Finally Arriving
Two independent adjusters in the Southeast, Ben Mandell and Mark Vinson, launched InsuranceClaim123 in early 2026, offering what they describe as a "CarFax for insurance claims." For $295, homeowners upload their insurer's build-back estimate and receive a report within three to five business days that identifies missed damage, questionable depreciation calculations, inconsistent repair scopes, and cases where the carrier proposed repairs instead of full replacement, which is one of the most common ways insurers reduce payouts without technically denying a claim. Mandell, a former homebuilder turned adjuster, built the tool with Vinson, who studied computer science, after watching homeowners accept lowball estimates simply because they had no independent way to evaluate whether the number was fair.
InsuranceClaim123 is not AI in the way insurers use the term. It blends limited automation with building permit data, material pricing databases, photographs, and human adjuster judgment, and it costs less than 2 percent of what a public adjuster typically charges to challenge a claim. It cannot force an insurer to pay more, but it can tell you whether the number on that letter is reasonable before you spend real money fighting it or, worse, accept it and discover the shortfall mid-rebuild when the framing contractor's invoice exceeds your remaining coverage by $40,000.
Public adjusters remain the most powerful individual counter-tool available. Working for the policyholder rather than the carrier, they organize documentation, translate policy language into actionable claims strategy, and present complete scopes designed to prevent the insurer from omitting line items that an untrained homeowner would not recognize as missing. They are also expensive, typically taking 10 to 15 percent of the final payout.
Parametric insurance offers a structural alternative that bypasses the estimation problem entirely. Instead of dispatching an adjuster to argue about damage after the fact, parametric policies use satellite imagery, IoT sensor data, and weather feeds to trigger automatic payments when predetermined conditions are met for a specific property. No estimate dispute, no AI valuation gap, no adjuster roulette. Monica Palmeira of the Greenling Institute calls it "a really useful tool to make sure folks have some kind of baseline coverage in a way that can be deployed very efficiently," though she notes the obvious catch: parametric policies pay fixed amounts regardless of actual damage, so a catastrophic loss still requires a traditional policy underneath.
The Regulatory Response: 12 States, One Pilot
In March 2026, the National Association of Insurance Commissioners launched the first formal government examination of insurer AI in claims processing. Twelve states participate: California, Colorado, Connecticut, Florida, Iowa, Louisiana, Maryland, Pennsylvania, Rhode Island, Vermont, Virginia, and Wisconsin. The pilot runs through September 2026, with a nationwide rollout vote scheduled for the NAIC's November fall meeting.
The NAIC's AI Systems Evaluation Tool covers four areas: the extent of AI use within the carrier, internal governance structures, high-risk system details, and the data those systems rely on. One provision matters more than the rest: insurers must take full responsibility for AI purchased from third-party vendors. You cannot blame Verisk when your algorithm underpays a wildfire claim. Existing insurance law applies to AI-driven decisions the same way it applies to human ones.
NAIC President Scott White framed it carefully at the 2026 Spring Meeting in San Diego: regulators "don't want to stand in the way of innovation that generally serves consumers" but want AI used "transparently, fairly and in ways that hold up to scrutiny."
The industry's response was less diplomatic. A joint letter from trade groups filed in December 2025 argued the program is "voluntary for regulators while compulsory for companies" and that carriers could face penalties from a tool that had not been finalized. The pilot launched anyway.
Separately, Florida State Representative Hillary Cassel has sponsored legislation requiring human review of AI-generated insurance decisions. In Illinois, a class-action suit alleges that State Farm's claims algorithms disproportionately impact Black policyholders, causing delays and underpayments. The case is pending.
What This Means If You Are Building a Home
If you just built a $500,000 home, you know exactly what it cost. You have the invoices, the permits, the lien releases, the final draw. Your insurer's algorithm does not use your actual construction cost as the rebuild estimate. It uses its own database, its own cost assumptions, its own depreciation schedule. On the day you move in, your Coverage A limit may already be below what you spent to build the house you are standing in.
If you used non-traditional construction methods, the gap gets worse. Our previous reporting on AI valuation models found that AI valuation models have statistically meaningless training data for 3D-printed, modular, and mass timber homes. The same problem hits insurance estimation: ICON has built approximately 200 3D-printed structures worldwide, and there are not enough comparable sales or reconstruction events to train a model. FHA requires lenders to use conventional site-built comps when fewer than two comparable non-traditional sales exist. Insurance estimating databases face the same void.
Run the coinsurance math on a concrete example. You build a home for $500,000 and carry $350,000 in dwelling coverage because your insurer's algorithm estimated replacement cost at $375,000 and your agent rounded down, and then you suffer a $200,000 kitchen fire. Your coverage ratio is 70 percent (350/500), which falls below the standard 80 percent coinsurance threshold. The insurer applies the coinsurance formula: (actual coverage / required coverage) × loss = payout. That is (350,000 / 400,000) × 200,000 = $175,000. You eat $25,000 on a partial loss because the algorithm underestimated your home by $125,000 on the front end.
Total loss is worse. You get $350,000. Your home cost $500,000. The insurance check covers 70 percent. You carry the remaining $150,000 through savings, a construction loan at current rates, or you sell the lot and leave.
Strongest Counterargument
Insurers argue that AI makes pricing more accurate, not systematically lower. Verisk says Xactimate's cost database is "market-based, transparent, and rooted in human-validated data" and that AI handles only photo labeling and information summarization, never cost generation. CAPE Analytics frames its aerial intelligence as helping carriers avoid both overpricing and underpricing, creating fairer assessments for everyone. If AI genuinely improves risk accuracy, some homes should get cheaper coverage while others get more expensive coverage, but all should be priced closer to their actual risk. The National Association of Insurance Commissioners explicitly states it does not want to block innovation that serves consumers.
That argument holds in theory. In practice, the Sastry dataset shows coverage ratios declining for a decade straight, across every state, and the only direction the insurance industry's AI investment moves the leakage number is down. Down means paying less, and the question regulators in twelve states are now trying to answer is whether "less" means "less than owed" or "less than previously overpaid." The data available to homeowners is not sufficient to tell the difference, and the data available to insurers is not shared.
What to Do About It
Pull your declarations page. Find Coverage A. Multiply your home's square footage by local per-square-foot construction costs from the CoreLogic Marshall & Swift database, which your agent can access, or use recent local contractor bids as a proxy. Divide Coverage A by the rebuild estimate. If the ratio falls below 0.8, you will trigger the coinsurance penalty on a partial loss. If it falls below 1.0, a total loss leaves a gap you carry yourself.
Ask your agent about extended replacement cost coverage, which adds 25 to 50 percent above the stated dwelling limit, and guaranteed replacement cost coverage, which eliminates the cap entirely. Guaranteed replacement cost is disappearing from many markets, particularly in wildfire zones and Gulf Coast states. If it is available in your area, take it. If it is not, the extended endorsement is the next best protection.
If you receive a claims estimate that feels low, spend the $295 on InsuranceClaim123 before spending $15,000 on a public adjuster. If the audit confirms the shortfall is real and substantial, then escalate. The order matters. Most homeowners either accept the first number or litigate. Very few verify it independently, and the insurance industry has relied on that gap for decades. AI has widened it. A $295 report will not close it. But it will tell you whether you are fighting a $5,000 discrepancy or a $150,000 one, and that determines everything that follows.
Limitations
The Sastry/Harvard study uses data through 2020. Coverage ratios may have stabilized or declined further since then; no updated dataset has been published. Xactimate's algorithmic methods are proprietary, and we cannot independently verify claims about systematic bias in its cost estimates. The NAIC's 12-state pilot launched with auto insurance as its primary focus; a dedicated home insurance examination has not been formally scheduled, though the evaluation framework applies to all lines. InsuranceClaim123 is too new for any independent evaluation of its accuracy or effectiveness. The state-level premium and coverage data from MoneyGeek uses a standardized 2,500-square-foot home at $250,000 dwelling coverage; your numbers will vary with actual home size, location, and construction type. The coinsurance calculation assumes standard policy language; your contract's specific provisions may differ.