State Farm Patent Puts Live Smart-Building Sensors Into Commercial Risk Selection

The U.S. Patent and Trademark Office published State Farm's application US 2026/0228840 A1 on August 6, 2026, describing AI models that combine live smart-building sensor readings from a portfolio of properties with claims history to generate location-specific risk recommendations (USPTO, August 2026), pushing commercial-property risk selection toward continuous scoring instead of a single point-in-time inspection.

A Portfolio of Buildings, Not a Single Sensor Placement

State Farm filed the underlying application, serial number 19/207,909, on May 14, 2025; the Patent Office published it fourteen months later under the title "Artificial Intelligence-Based Systems and Methods Utilizing Smart Building Data Analytics and Loss Reports." The abstract lays out a six-step process: the system receives smart building analytics data tied to a first group of buildings, receives claims data tied to a second group of buildings that overlaps the first, takes in a target location, runs both datasets through one or more trained AI models, and returns a recommendation for that location to a user's device (US 2026/0228840 A1, abstract). Independent claim 1 calls it a "building planning computer system" and repeats the same structure almost verbatim: input the sensor and claims data associated with buildings "each located at different locations," then generate recommendations "for the select location based upon the smart building analytics data and the claims data."

That plural framing, a first plurality of buildings and a second plurality of buildings, is what separates this filing from State Farm's earlier sensor patents. actuary.info's coverage of the carrier's water-sensor placement patent tracked a system scoped to optimizing device positions inside a single home, and its analysis of the self-updating pricing model patent covered a personal-lines rating engine that retrains itself on one policy at a time. The August filing instead trains its recommendations on aggregated analytics and loss experience pulled across many properties before applying the output to one. The building taxonomy in the specification confirms the portfolio is not residential by default: alongside "a house, an apartment, a townhome, a multi-family home, a condo/co-op, a manufactured home, a mobile home," the filing lists "a business," "a hospital," "fire stations, police stations," and even a "solar power plant" and "wind turbine plant" as buildings the system can analyze. Sensor inputs named in the claims and description include temperature, humidity, activity levels, electricity consumption, and maintenance frequency, the same categories of telemetry that a commercial building automation system already streams for HVAC and energy management, repurposed here as underwriting-adjacent signal.

From Point-in-Time Classification to a Continuous Feed

Commercial property underwriting has run for decades on a static classification framework: construction, occupancy, protection, and exposure, the COPE variables an underwriter or loss-control inspector captures at binding and typically refreshes only at renewal or during a periodic reinspection. Those variables are proxies. Construction class stands in for how a building burns; protection class stands in for how fast a fire gets suppressed; occupancy stands in for what kind of activity, and therefore what kind of liability and theft exposure, happens inside. The patent's continuous sensor feed reaches for the same underlying drivers through a different channel, replacing a periodic snapshot with a live stream.

Sensor metric named in the patent	Traditional COPE-style proxy	Frequency or severity driver it approximates
Temperature	Construction class, periodic inspection note	Pipe freeze, HVAC failure, fire ignition conditions
Humidity	Not typically captured outside claims history	Mold growth, latent roof or envelope leaks
Activity levels	Occupancy classification	Liability and theft exposure tied to actual foot traffic
Electricity consumption	Not typically captured pre-loss	Electrical fire risk, equipment strain, vacancy signal
Maintenance frequency	Loss-control survey findings	Deferred maintenance and neglect-driven loss potential

The gap that table exposes is not the sensor categories themselves; it is cadence. A COPE variable updates when someone reinspects the property, typically once a year at renewal. The patent's claim 1 describes a system that ingests smart-building analytics data continuously and can regenerate a recommendation, and by extension a risk score, at any point a policyholder's data updates. actuary.info's broader look at IoT sensor networks moving commercial property risk monitoring inward and its coverage of AI models reshaping commercial-lines risk selection at submission both describe the front end of this shift, sensors and AI touching the underwriting file. This filing is the mechanism that would let a carrier act on that data between renewals rather than waiting for the next reinspection cycle to catch a risk-state change.

The Loop From a Sensor Flag to Loss Control to a Premium

The specification does not stop at generating a recommendation. It describes the system tracking whether a policyholder acts on it, whether that means completing a maintenance task or making a structural change, and adjusting outcomes accordingly: "the insurance premium associated with the building may be adjusted" once a recommended change is verified, and the system can generate "one or more proposed insurance policies and associated premiums" for the location (US 2026/0228840 A1, description). The models are also designed to retrain on what happens next, updating based on changes made to a location following an insurance event or a completed recommendation. That closes a loop that starts with a sensor anomaly, routes through an AI-generated loss-control recommendation, depends on what the policyholder does or does not do about it, and feeds the outcome back into the training data that shapes the next building's score.

That pattern, a claims or sensor signal feeding a model that then feeds back into pricing or reserving, is not isolated to this filing. actuary.info's coverage of AI-driven workers' compensation claims triage and loss development and of a separate carrier's patent reaching from claims classification into total-loss reserve estimation both describe closed loops on the claims side of the ledger. This filing extends the same architecture upstream, into loss control and risk selection, before a claim exists at all. The mechanism is more defensible on paper than it is simple to operate: a recommendation engine that adjusts premium based on verified remediation needs a reliable way to confirm the remediation actually happened, not merely that a sensor stopped flagging it, a verification step the patent claims but does not detail.

Data Lineage and Rate-Filing Questions

Once an AI-generated recommendation starts shaping "proposed insurance policies and associated premiums," it stops being a voluntary loss-control nudge and starts looking like a rating input, and rating inputs draw regulatory scrutiny that maintenance tips do not. Twenty-four states had formally adopted the NAIC Model Bulletin on the Use of Artificial Intelligence Systems by Insurers as of March 2025, requiring insurers to maintain "a written program for responsible use of AI systems that make or support decisions related to regulated insurance practices" (Quarles Law, March 2025). The bulletin's governance expectations reach directly into what this patent describes: insurers must document controls addressing "data currency, lineage, quality, integrity, bias, minimization, and suitability" across an AI system's lifecycle, and they retain responsibility for diligence on third-party data feeding those systems. actuary.info's broader survey of state-by-state AI regulation adoption and the federal preemption fight covers how unevenly that requirement is currently enforced.

A smart-building recommendation engine trained on portfolio-wide sensor and claims data is a harder lineage problem than most AI underwriting tools already in production, because the sensor feed does not originate inside the carrier's own systems. It comes from building automation platforms, third-party IoT vendors, and whatever subset of commercial policyholders chose to install connected equipment in the first place. That last point cuts against the model's own training data before any bias-testing program gets involved: buildings wired for continuous monitoring skew toward newer construction, larger accounts, or already risk-conscious ownership, an adoption pattern that is itself correlated with lower loss potential. A recommendation engine trained disproportionately on that self-selected population risks calibrating its baseline "composite risk score" (patent description) against a sample that looks nothing like the broader commercial book a rate filing has to cover, unless the carrier corrects explicitly for who opted into the sensor program and who did not.

Where This Sits in a Widening AI Patent Portfolio

State Farm has filed 326 AI-related patents since 2014, and together with USAA and Allstate accounts for 77% of all AI patents filed by U.S. property-casualty insurers over that period (Evident Insights, December 2025), a concentration actuary.info examined in detail in its analysis of what that 77% share means for the rest of the market. Building a proprietary portfolio-level scoring system, rather than licensing one, puts State Farm on the carrier side of a build-versus-buy contest playing out across commercial property risk technology, where vendors like ZestyAI and Verisk sell comparable property risk models to dozens of carriers at once, a dynamic actuary.info covered in its reporting on the competition between property-risk AI vendors and the carriers evaluating them. A patented, in-house system is harder for a competitor to replicate and cannot be pulled out from under State Farm by a vendor contract renegotiation, at the cost of having to build and validate the data pipeline, including the third-party sensor lineage problem above, entirely in-house.

The commercial incentive behind that choice is visible in State Farm's own numbers. Its homeowners and commercial multiple peril line ran a combined ratio around 108 in 2025, only 2.5 points better than 2024 despite earned premium jumping more than 13% for the second consecutive year, largely on the back of the January 2025 Los Angeles wildfires (Carrier Management, February 2026). That result sits well behind the broader market: Fitch Ratings pegged the U.S. commercial lines sector's aggregate 2025 combined ratio at 94%, the best result in more than fifteen years, as commercial property rates softened for the first time in nine years (Fitch Ratings, 2026 outlook). A carrier running double-digit points worse than the sector average on the line this patent targets has a direct underwriting-profit reason to sharpen risk selection with continuous data rather than wait for a softening market to fix the line on its own. The addressable base for that data is also growing on its own trajectory: commercial buildings account for roughly 11.8% of the global installed base of IoT devices, and the installed base of smart commercial-building IoT devices was projected to grow at a 13.7% compound annual rate between 2023 and 2028, reaching an estimated 3.25 billion devices (ResearchAndMarkets, May 2023), meaning the sensor population a system like this one draws on keeps expanding independent of anything State Farm does.

A Narrower Lag Between Risk State and Rate

What this filing changes, if it moves from application to product, is the interval between a risk changing and a carrier knowing it. A static COPE variable holds until the next renewal or reinspection; a continuous sensor feed can flag deferred maintenance, an electrical anomaly, or an occupancy shift the same week it happens. That shrinks the gap actuaries have historically priced around, the assumption that a risk assessed at binding stays reasonably representative until the next annual look. Reserving and pricing actuaries reviewing a model built on this architecture will want two things documented alongside the familiar construction and occupancy variables: how frequently the underlying score actually updates in production versus what the patent claims, and what share of the rated book the training population's sensor-adoption skew leaves uncorrected. Neither question is answered by the patent language itself, and neither will be answered by anyone outside the company until the underlying product, and the rate filing built on it, becomes public.