Claim 1 of US Patent 12,700,043 B2 opens with a weather service reporting a hazard inside a geographic radius, then sends the system to pull geotagged photographs and video from public feeds for one specific address inside that radius. A model trained in two stages on both feeds returns the status of that property. State Farm won the grant on August 4, 2026 across 21 claims.
The word running through the 188-paragraph specification is parametric. Claim 21 has the engine retrieve trigger parameters for an insurance contract and display a claim when the model finds them met; claim 9 initiates that claim without an in-person inspection. The per-site loss estimate lands before an adjuster sees the roof, and on one claimed path before the policyholder has reported anything.
Key Takeaways
- 21 claims issued on application 18/597,587, filed March 6, 2024 against a provisional from a year earlier. Examiner Edward Chang mailed a final rejection in November 2025; State Farm cleared it with a continuation request in January 2026.
- Five to ten miles is the proximity band at which a neighbor's claim becomes a trigger to assess a policyholder who has filed nothing, letting the engine originate a loss estimate ahead of first notice.
- 40% to 60% are the confidence thresholds the specification names as escalation points, with cost thresholds of $50,000, $100,000 and $200,000 doing the same job on the severity axis.
- $5.6 billion of hail claims paid in 2025, up roughly 12% on 2024, is the book this engine sits in front of, with Texas alone at $1.4 billion (Claims Journal, April 2026).
- $3.1 billion underwriting loss on $39.2 billion of homeowners and commercial earned premium in 2025 (State Farm, February 2026) is the pressure behind automating per-site damage assessment.
Patent Details
| Patent number | US 12,700,043 B2 |
|---|---|
| Title | Determining an Event and Resulting Damage Using Crowdsourced Data |
| Assignee | State Farm Mutual Automobile Insurance Company, Bloomington, Illinois |
| Application | 18/597,587, filed March 6, 2024 |
| Priority | Provisional 63/488,685, filed March 6, 2023 |
| Pre-grant publication | US 2024/0303748 A1, September 12, 2024 |
| Granted | August 4, 2026 (Official Gazette week 31) |
| Claims | 21, including the parametric-trigger claim 21 |
| Classification | CPC G06Q 40/08; USPC 705/4; art unit 3696 |
| Inventors | Rick Lovings, Jody A. Thoele, Erik Skyten and five co-inventors |
| Continuation | Application 19/732,131, filed July 6, 2026, pending |
What Claim 1 Requires
Two data legs, named separately. The first hazard event source has to be a weather service; the supplemental sources have to include a public data source, which claim 2 opens to crowdsourced, donated or public data and claim 3 ties to the timing of the event and the specific location.
The specification is direct about the problem. Insurers rely on National Oceanic and Atmospheric Administration measurements "taken for the general vicinity of an insured home, business, or vehicle at the city or county level," and "there is often a difference between these measurements and the actual weather event for a specific location" (US 12,700,043 B2, August 2026). The crowdsourced leg closes a city-to-address resolution gap in the authoritative feed.
The training limitation is where the grant was won. Claim 1 requires a model trained in a first stage on historical weather data from the weather service, then in a second stage on picture or video data from the supplemental sources, and selected from a pool trained for different hazard types.
That specificity arrived after a fight. Art unit 3696 rejected the application in May 2025, rejected it again in November, and allowed it only after the January 2026 continuation request, which is the drafting pattern the post-Recentive eligibility reset now rewards.
The worked example is a parametric homeowners policy paying $10,000 when hail at the residence reaches two inches in diameter. The engine answers whether that threshold was met at that address, and claim 21 wires the answer straight to a displayed claim.
Where the Loss Estimate Lands in the Triangle
Claim 10 lists three triggers: a policyholder submitting a claim, the weather data satisfying a criterion, or a different policyholder submitting a claim on a second property within a predetermined distance of the first, which the specification puts at five or ten miles. The third trigger points the engine at policyholders who have reported nothing.
The conventional order runs event, reported claims over a tail, inspection, estimate, case reserve, with the actuary carrying the unreported remainder as pure IBNR. Here the estimate can precede first notice, and the specification puts claim processing at under three days, within 24 hours, or during the event itself.
Two effects fall out and they run in opposite directions. Faster reporting compresses the development pattern, so age-to-age factors calibrated on historical catastrophe reporting lags overstate ultimate when applied to an accelerated triangle, the same distortion showing up in AI-accelerated workers compensation triangles. The neighbor trigger does something else: it converts losses that would never have been reported into reported claims, raising ultimate counts rather than moving them forward.
Severity governance sits in the dependent language. The specification escalates to a physical inspection when modeled cost clears $50,000, $100,000 or $200,000, and routes to a human when confidence falls below 60%, 50% or 40%. Those dials decide how much of a severity pick a model sets rather than an adjuster.
Against a book that paid $5.6 billion of hail claims in 2025, with the top ten states carrying $4.2 billion of that, the setting is not a rounding decision. Severe convective storm accounted for roughly $26 billion of the $46 billion in global insured catastrophe losses in the first half of 2026 (Gallagher Re, July 2026), so this patent is drafted around the peril carrying more than half the industry's catastrophe bill. Allstate's 2026 grant on catastrophe claim volume and adjuster allocation optimises the deployment of people once claims arrive; State Farm's claim 10 runs the other way and manufactures the claim.
The Model's Own Output Re-enters the Training Set
Paragraph 160 of the specification is the part worth reading twice. Once the engine determines the actual status of a property, it compares that actual cost to the model-determined cost, and if the comparison satisfies a criterion it generates a new historic event record containing the model-determined cost. Claims 6 and 17 make retraining on those updated records an express limitation.
The corpus the engine learns from can absorb the engine's own estimates as though they were settled amounts. A systematic model-to-actual bias never surfaces as a residual, because the residual is what decides whether the record gets written. Severity picks calibrated on that corpus inherit the tilt, and so does any catastrophe IBNR built on them.
The tilt would not be random across events. Claim 1 selects one model per hazard type, so a hail model carrying a downward bias carries it into every hail event it touches, in the same direction, across every state in the book. That is a correlated reserving error on a homeowners line that lost $3.1 billion in 2025 even as State Farm's overall property-casualty book turned a $1.5 billion underwriting gain on $111.6 billion of earned premium.
The authoritative leg is thinning at the same time. Claim 1 hard-requires a weather service as the first source and the specification names NOAA, which retired its Billion-Dollar Weather and Climate Disasters product on May 8, 2025 after 403 events and more than $2.9 trillion in damage since 1980. The leg the patent treats as ground truth has stopped growing while the crowdsourced leg expands, and claim 2 admits crowdsourced, donated or public data with no veracity limitation attached.
At the lowest score criterion named, a configuration can act on an estimate the model itself rates at just above two-in-five confidence, with claim 9 removing the inspection that would have caught the miss. Every one of those estimates, once compared and recorded, is eligible to become training data.
Further Reading
- Allstate Patents a Catastrophe Claim Volume and Adjuster Allocation Engine - The rival approach, optimising response after claims arrive.
- The AI Patent Race in Insurance: Complete Guide - Hub page on carrier and vendor patent strategy.
- State Farm Patents a Pricing Model That Rewrites Itself - The same retraining loop, on the rating engine.
- A Near Miss on Mexico’s Earthquake Cat Bond - Basis risk when a parametric trigger and the actual loss disagree.
- AI Claims Speed Is Biasing Workers Compensation IBNR - What faster reporting does to a development pattern.