State Farm Patents the Placement Logic Behind Water Sensors, Not the Sensors

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.

Key Takeaways

• The patent covers placement, not hardware. The model takes a structure's geographic location and applicable building codes plus historical water-damage claims data and outputs optimal sensor locations.
• 27.6% of all homeowners claims in 2022 were water damage and freezing, up from 19.6% two years earlier, at an average $13,954 per claim.
• 1.61 claims per 100 insured homes annually, so roughly one in 60 insured homes files a water or freezing claim each year.
• Twenty claims across method, system and medium, with a companion patent filed the same day covering augmented-reality visualization of the model's output.
• The effect is a truncated right tail, not a uniform frequency drop, which breaks the stationarity assumption a blended water-peril development triangle rests on.

Patent Details

• Patent number: US 12,694,576 B2, "Determining Optimal Water Sensor Placement Using Machine Learning"
• Filed: September 25, 2023
• Published: 2024, as application US 2024/0319683 A1
• Granted: July 28, 2026
• Claims: 20, split as ten method, nine system and one non-transitory computer-readable medium
• Companion: US 12,567,182, filed the same day and granted March 3, 2026 to inventors Kyle Malan and Sean Kingsbury, covering AR visualization of the placement output

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 along with historical water-damage claims data into a trained machine learning model, and generating an output identifying one or more optimal locations for water sensor placement (USPTO grant). The model is trained specifically on claims records carrying the geographic location of the structure tied to each historical claim, so a ranch house in a freeze-prone Minnesota zip code can draw a different recommendation from a slab-foundation home in humid coastal Texas at the same square footage.

The companion filing is the more revealing part. US 12,567,182, filed the identical day and granted five months earlier, covers using augmented reality to visualize the locations the placement model identifies, overlaying instruction markers on a live camera view. One patent locks up the decision, the other the delivery. A competitor reverse-engineering an AR-guided installation app still runs into the placement logic underneath it; one building around the placement algorithm still needs a way to get the output to an installer.

From a Generic Discount to a Development-Triangle Problem

Most insurer sensor programs place devices using generic manufacturer guidance: under sinks, near water heaters, behind the washing machine. State Farm's own ADT bundle ships up to seven sensors for a discount of up to 6% off the homeowners premium, and rivals run comparable programs. The training signal here is different: not where leaks generally start, but where claims actually originated at comparable structures in comparable locations.

That distinction is what turns a marketing bundle into a filed rating argument. Regulators generally require a premium credit to correspond to a demonstrable reduction in expected loss, and a generic smart-home discount rests on vendor marketing or third-party efficacy studies the filing carrier does not control.

A carrier can instead point to its own claims history as the training input, document the optimization, and produce before-and-after experience on policies where sensors sat at model-recommended locations. The exposure justifies the effort: non-weather water damage runs about 20% of all property insurance losses, an estimated $8.24 billion a year in homeowners claims and $3.57 billion in commercial property. Commercial lines has the same gap in reverse, with ISO rate plans carrying no explicit variable for sensor deployment, forcing IoT credits through schedule modifications.

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 mechanism worth isolating is not that sensors reduce claims but how. A slow leak behind a wall or under a slab runs undetected for days or weeks before surfacing, by which point remediation, mold mitigation and structural repair push a claim well past the $13,954 average. A sensor sited where the training data says a leak is statistically most likely to originate catches the same leak within hours, before it propagates through drywall, flooring or framing.

The result 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 remediation events resolve instead as small early-caught ones: a shut-off valve trips, a plumber replaces a fitting, and the claim closes for a few hundred dollars. Frequency may barely move, because a caught leak still generates a claim record, just for a fraction of the amount.

That is a reserving problem before it is a pricing opportunity. Chain-ladder and other development-triangle methods assume the claims-generating process is stationary, so that the relationship between an early-period claim count and its ultimate value in prior accident years holds for the current one. A book where an increasing share of policies carry optimally placed sensors violates that directly: loss development factors built on a claims mix dominated by delayed-discovery, high-severity losses will overstate ultimates on the sensor-covered subset, which closes early and cheap.

Nor does it wash out with gradual adoption, because the shift is nonlinear. A book 5% sensor-covered behaves close to baseline. A book crossing 30% or 40% penetration, plausible within a few renewal cycles if a defensible discount drives faster uptake than a generic bundle, needs either a triangle segmented by sensor status or an explicit adjustment to LDFs derived from legacy unsegmented data. A single blended selection carries redundant reserves on the covered segment while potentially under-reserving the unprotected one.

The Telemetry Asset and the Governance It Attracts

Deploying sensors at model-optimized locations generates a class of data State Farm did not previously hold at scale: geotagged, structure-specific telemetry showing when and where water events were detected relative to the model's predicted highest-risk points. That feeds the training process directly, retraining placement against confirmed detections rather than only historical claims, and it is separately an exposure-level asset, since knowing where in a structure a triggering event happened and how fast it was caught is granularity generic claims data does not carry.

It is also proprietary in a way vendor-sourced property data is not. The aerial-imagery firms whose data moats depend on licensing attributes to multiple carriers can be matched by any rival signing the same contract. A competitor here would need the model, the fleet and years of confirmed-detection telemetry.

The retraining loop is exactly what creates the governance exposure. 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, and that framework does not distinguish a licensed vendor model from one a carrier builds and patents internally. Both face documentation expectations on training data, retraining cadence and disparate-impact testing.

So the asset and the liability are the same mechanism. Regulators are already pushing insurers to document not just a model's output but the version that produced it at the time of filing. A discount justified by a 2026-vintage placement algorithm is not automatically still justified once the model has been substantially retrained on two or three years of new detection telemetry, and the better the feedback loop works, the faster the filed justification goes stale.