US Patent 12,694,576, granted to State Farm on July 28, 2026, does not cover a water sensor. It covers the machine learning model that decides where to install one, trained on historical water-damage claims data and building codes, for a peril that drove 27.6% of all homeowners insurance claims in 2022, up from 19.6% two years earlier (I.I.I., citing ISO/Verisk data). Any carrier can bundle a sensor kit. Few can show a patented method for where the sensors go.

That distinction is the whole story. Water-sensor giveaways are already a commodity: State Farm's own ADT bundle ships up to seven sensors plus a smoke detector and door and window sensors for a discount of up to 6% off the homeowners premium (Coverager), and rivals from Ting to Notion run comparable programs. What none of them has previously had is a patented, claims-trained algorithm that tells a technician or a homeowner exactly where a sensor needs to sit to catch the specific failure modes a carrier's own loss history says are most likely at that structure. Patent 12,694,576, titled "Determining Optimal Water Sensor Placement Using Machine Learning," was filed September 25, 2023 and published in 2024 as application US 2024/0319683 A1 before issuing this week with 20 claims spanning method, system, and non-transitory-medium coverage (USPTO grant, US 12,694,576 B2).

What the claims actually cover

The independent method claim describes obtaining data about a structure, including its geographic location and applicable building codes, feeding that data along with historical water-damage claims data into a trained machine learning model, and generating an output that identifies one or more optimal locations for water sensor placement relative to the structure (USPTO grant, US 12,694,576 B2). The model is trained specifically on claims records that include the geographic location of the structure tied to each historical claim, so the placement recommendation for a ranch house in a freeze-prone Minnesota zip code can differ from the recommendation for a slab-foundation home in humid coastal Texas, even if both structures are nominally the same square footage. Twenty claims break down into ten method claims, nine system claims covering the processor and memory architecture needed to run the model, and one claim for the non-transitory computer-readable medium storing the executable instructions, a claim structure that mirrors how carriers now routinely wrap a single inventive method in method, system, and medium coverage to close off workaround paths.

State Farm did not stop at the placement algorithm. A companion patent, US 12,567,182, filed the identical day, September 25, 2023, and granted five months earlier on March 3, 2026 to inventors Kyle Malan and Sean Kingsbury, covers a separate invention: using augmented reality to visualize the optimal sensor locations the placement model identifies, overlaying instruction markers on a live camera view so an installer or homeowner can see exactly where to mount each device (Google Patents, US 12567182B2). Filing both applications on the same date is a deliberate two-layer IP strategy: one patent locks up the decision (where should a sensor go), the other locks up the delivery mechanism (how does a person find that spot in the physical world). A competitor that reverse-engineers an AR-guided installation app still runs into the placement-logic patent underneath it, and a competitor that licenses or builds around the placement algorithm still needs its own way to communicate the output to an installer.

Loss prevention as underwriting input, not a marketing add-on

Most insurer sensor programs today, State Farm's ADT bundle included, place devices using generic manufacturer guidance: under sinks, near water heaters, behind the washing machine. That is reasonable but untargeted. Patent 12,694,576 replaces the generic rule with a model trained directly on the carrier's own claims experience, which means the training signal is not "where do leaks generally start" but "where did claims actually originate at structures with comparable characteristics in comparable locations." That closes a loop that has historically run one direction: claims data informed reserving and pricing, but rarely fed back into where loss-control equipment gets physically installed. actuary.info's coverage of the shift from AI underwriting moving from selection to prevention flagged this same pattern emerging across commercial lines: tools that intervene before a loss occurs, not merely before a policy binds, change which loss-cost assumptions still hold for the book they touch. A claims-trained placement model is that shift applied to a specific, high-frequency personal-lines peril.

The scale of what is being targeted is not small. Non-weather water damage accounts for roughly 20% of all property insurance losses, costing insurers an estimated $8.24 billion a year in homeowners claims and a further $3.57 billion in commercial property claims (Location, Inc., analysis of carrier loss data). Water damage and freezing claims average $13,954 per claim across 2018 through 2022, with a claim frequency of 1.61 per 100 insured homes each year, meaning roughly one in 60 insured homes files a water or freezing claim annually (I.I.I., citing ISO/Verisk data via a December 2023 NAIC report). A placement method that measurably improves detection rates against a loss category this large is not a minor product feature. It is a frequency lever applied to one of the two costliest, most volatile causes of loss on the homeowners book, after wind and hail.

From a generic discount to a defensible rating factor

Insurance regulators generally require that a premium credit correspond to a demonstrable reduction in expected loss, not simply reward equipment ownership. A generic "smart home discount" filed today typically rests on vendor marketing claims or third-party efficacy studies the filing carrier does not control, which is a thin actuarial foundation and one every competitor selling the same commodity sensor can match. A discount built on a patented, claims-trained placement methodology is different in kind. State Farm can point to its own historical claims data as the training input, document the specific optimization the model performs, and, over time, produce its own before-and-after loss experience on policies where sensors were installed at model-recommended locations versus generic locations. That is closer to the actuarial basis behind a documented loss-control credit than to a marketing bundle, and it is not something a competitor without access to State Farm's claims history or an equivalent claims-trained model can replicate simply by buying the same sensor hardware.

This mirrors a gap actuary.info identified in commercial lines: current ISO rate plans carry no explicit rating variable for monitoring history or sensor deployment, forcing carriers to route IoT credits through schedule modifications rather than a filed base-rate variable, largely because insurers have lacked a defensible, quantifiable basis for how much a given sensor deployment reduces expected loss (actuary.info coverage of IoT sensors in commercial property underwriting). A patented placement algorithm trained on claims data is precisely the kind of documented methodology that could eventually support a genuine rating variable rather than a discretionary schedule credit, since the carrier can show the factor is tied to a specific, reproducible process rather than an underwriter's judgment call.

Metric Figure Source
Share of all homeowners claims, 2022 27.6% (up from 19.6% in 2020) I.I.I. / ISO-Verisk
Average water/freezing claim severity, 2018-2022 $13,954 I.I.I. / ISO-Verisk
Claim frequency, 2018-2022 1.61 per 100 insured homes annually I.I.I. / ISO-Verisk
Annual homeowners losses, non-weather water $8.24 billion Location, Inc.
Annual commercial losses, non-weather water $3.57 billion Location, Inc.
Existing generic IoT discount, State Farm/ADT bundle Up to 6% off premium Coverager

The distributional shift: compressing the severity tail

The actuarial mechanism worth isolating is not that sensors reduce claims, that much is intuitive, but how they reduce claims, and that shape matters more than the headline frequency reduction. A slow leak behind a wall or under a slab foundation typically runs undetected for days or weeks before it surfaces as visible damage, by which point remediation, mold mitigation, and structural repair can push a claim well past the $13,954 average into the tens of thousands of dollars. A correctly placed sensor, sited at the specific point in the structure where the model's training data says a leak is statistically most likely to originate rather than at a generic manufacturer-recommended spot, catches that same leak within minutes or hours of onset, before it has time to propagate through drywall, flooring, or framing.

The effect on the loss distribution is not a uniform downward shift in claim size. It is a truncation of the right tail. Claims that would have developed into large, delayed, multi-month remediation events under a generic or absent sensor program instead resolve as small, early-caught events: a shut-off valve trips, a plumber replaces a fitting, and the claim closes for a few hundred dollars rather than tens of thousands. Frequency may not fall as much as severity distribution shifts, since a sensor that catches a leak still generates a claim record, just for a fraction of the loss amount. A carrier deploying a claims-trained placement model at scale should expect the mix of its water-peril claims to shift materially toward the low-severity end of the distribution, with the large, structurally destructive claims becoming disproportionately rarer relative to the small, caught-early claims that dominate frequency counts.

Why that shift breaks a stationarity assumption

This is where the patent becomes a reserving problem, not just a pricing opportunity. Chain-ladder and other development-triangle methods implicitly assume the process generating historical claims, the mix of small versus large losses, the pace at which claims develop and close, is stationary: that the relationship between an early-period claim count and its ultimate value observed in prior accident years will hold for the current accident year. A book where an increasing share of policies carry claims-trained, optimally placed sensors violates that assumption directly. Loss development factors calculated on a pre-sensor claims mix, dominated by delayed-discovery, high-severity losses, will systematically overstate the ultimate value of claims arising on a sensor-covered subset of the book, because those claims are increasingly caught and closed early, cheap, and fast, a fundamentally different development pattern than the triangle was built to describe.

A reserving actuary working a homeowners water-peril triangle cannot simply assume this washes out in the aggregate as sensor penetration grows gradually. The shift is nonlinear with adoption: a book that is 5% sensor-covered behaves close to the historical baseline, while a book that crosses 30% or 40% sensor penetration, plausible within a few renewal cycles if a patented, defensible discount drives faster adoption than a generic marketing bundle would, needs either a separate triangle segmented by sensor status or an explicit trend adjustment applied to the LDFs derived from the legacy, unsegmented data. Carriers that continue to reserve water peril off a single blended triangle risk carrying redundant reserves on the sensor-covered segment of the book while potentially under-reserving the still-unprotected segment if the blended selection masks a widening bifurcation between the two.

A new telemetry asset, and who controls it

Beyond the placement algorithm itself, deploying sensors at model-optimized locations generates a new class of data State Farm did not previously own at scale: geotagged, structure-specific telemetry showing exactly when and where water events were detected relative to the model's predicted highest-risk points. That telemetry is a feedback loop the patent's own training process can consume, retraining the placement model against confirmed detection events rather than only against historical claims data, which should make each successive model version more precise than the last. It is also, separately, a pricing and reserving asset in its own right: a carrier that knows not just that a claim occurred but where in the structure the triggering event happened, and how quickly it was caught relative to a model-predicted risk point, has exposure-level granularity that generic claims data alone does not provide.

That puts State Farm in a different competitive position than the third-party property-data vendors actuary.info has covered extensively, firms like the aerial-imagery and computer-vision providers whose data moats depend on licensing structure-level attributes to multiple carriers. State Farm's sensor telemetry is proprietary and internally generated rather than vendor-licensed, which means no competitor gains visibility into it through the normal channel of buying the same third-party data feed. A rival carrier without an equivalent claims-trained placement patent and an equivalent installed sensor base cannot replicate this data asset by signing a vendor contract; it would need to build the underlying model, the sensor fleet, and years of confirmed-detection telemetry from scratch, or negotiate a license State Farm has no obligation to grant.

Governance and documentation exposure

A proprietary, claims-trained placement model used to justify a rating credit will draw the same model-governance scrutiny that has increasingly attached to third-party vendor models. As of mid-2026, 25 states and the District of Columbia had adopted the NAIC's AI Model Bulletin in full or substantially similar form (NAIC AI Bulletin adoption tracker, Q2 2026), and that framework does not distinguish between a model a carrier licenses from a vendor and a model it builds and patents internally; both face documentation expectations around training data, retraining cadence, and disparate-impact testing. actuary.info's coverage of the broader model-drift and rate-filing governance gap found regulators pushing insurers to document not just a model's output but the version of the model that produced it at the time a filing was made (actuary.info coverage of AI pricing model drift and rate-filing governance). If State Farm periodically retrains the sensor-placement model on newly confirmed detection telemetry, a filed IoT discount tied to that model's output needs the same version-tracking discipline: a discount justified by a 2026-vintage placement algorithm is not automatically still justified once the underlying model has been substantially retrained on new data two or three years later.

What competitors are left to do

Every carrier running a sensor discount program now has to answer a version of the same question a patent grant always poses: build around it, license it, or compete on a dimension the patent does not cover. Building around 12,694,576 likely means developing an independently derived placement methodology, one not trained the same way on claims data correlated to building codes and geographic location, which is a meaningful research investment for carriers that currently just follow manufacturer install guides. Licensing is plausible if State Farm chooses to monetize the patent rather than hold it purely defensively, though nothing in the grant signals that intent yet. Competing on a different dimension, faster installation, cheaper hardware, broader smart-home ecosystem integration, sidesteps the patent but leaves the underlying actuarial question unanswered: whose discount can actually demonstrate, with a documented and reproducible methodology, that it reduces expected loss. For now, State Farm is the only carrier that can point to a granted patent as that documentation.

Further Reading

Sources