Quanata Patents an ML Engine That Values What You Own

A trained machine-learning model reads a policyholder's personal data, compares it against historical policyholder records containing insurance claims, and returns a list of the items that person probably owns with a dollar value attached to each. Quanata, LLC won US Patent 12,700,041 B2 for that system on August 4, 2026 (USPTO Official Gazette). Claim 1 requires no photograph, no receipt, and no walkthrough of the house.

Key Takeaways

• The continuation drops the image requirement its predecessors carried. Claim 1 infers a household inventory from personal data run against historical policyholder records instead of analyzing images.
• Coverage C is usually set at 50% of the dwelling limit, with many insureds electing 70 to 75 percent. That is a form convention rather than an estimate of contents.
• 74% of Marshall Fire claimants were underinsured and 36% severely so, across roughly 5,000 policies written by 24 insurers.
• Policyholder corrections are behavioral responses, not ground truth, arriving from a self-selected subset of insureds with a directional interest in the answer.

Patent Details

Patent number	US 12,700,041 B2
Title	Systems and methods for generating and updating an inventory of personal possessions of a user for insurance purposes
Assignee	Quanata, LLC (San Francisco, CA)
Inventor	Kenneth Jason Sanchez (San Francisco, CA)
Application	18/397,932, filed December 27, 2023
Parent	Continuation of 16/793,810, filed February 18, 2020, granted as US 11,861,722
Pre-grant publication	US 2024/0127356 A1, April 18, 2024
Granted	August 4, 2026 (Official Gazette Vol. 1549 No. 1)
Classification	CPC G06Q 40/08 (insurance); G06Q 30/0278 (appraisal or valuation); G06F 16/9535
Claims	21

Predict, Price, Correct, Retrain

Claim 1 recites a short, closed sequence: receive personal data associated with a user; predict items possessed by that user using a model trained on historical policyholder records containing insurance claims; assign a predicted value to each item; display the items and values on a user device; receive adjusted values back from the user; and retrain the model on those adjustments.

The February 2020 priority date matters for reading the family. Quanata has been prosecuting this idea for more than six years, and the version that issued is the fourth pass, filed as a continuation five weeks before the parent issued. What changed is that the image requirement is gone.

The training corpus is the interesting specification. The model learns from records that contain insurance claims, which means the reference set for what a household like this owns is drawn from households that filed. Claims records are the richest itemized inventory a carrier holds, because a contents claim forces an insured to enumerate and value what burned or washed out.

That is also a selected sample. A carrier's itemized contents data concentrates in the perils that generate contents losses, and those are not evenly distributed. Wind and hail account for 40.7 percent of homeowners claims, water damage and freezing 27.6 percent, and fire and lightning 21.9 percent, while theft is 0.7 percent. A model trained on that mix learns the contents profile of flooded basements and burned attics more sharply than it learns the contents profile of a jewelry box.

The examiner's classification says something about what the office thought the claim was doing. CPC G06Q 30/0278 covers appraisal or valuation, attached alongside the insurance class: not administering a policy, but pricing goods. Claims reciting a specific trained model, a specific training corpus and a specific retraining trigger have been surviving where broad "determine a value using a computer" language has not, a divide the site tracked through the Section 101 reset.

Coverage C Is Derived, Not Measured

What the claim automates is not claims triage. It is Coverage C, the personal-property limit underneath every homeowners and renters policy, which carriers have long set by multiplying the dwelling limit by a rule-of-thumb percentage rather than by measuring anything.

That convention has a known failure rate. Among Marshall Fire claimants, 74% were underinsured and 36% severely so, across roughly 5,000 policies written by 24 insurers. A percentage-of-dwelling rule cannot be wrong in any individual case, because it is not an estimate of contents; it is a form default that happens to be applied to contents. The gap only becomes visible at total loss, when the insured is asked to enumerate what the limit was supposed to cover.

A model that estimates contents value directly severs the link between Coverage A and Coverage C, and the exposure base moves with it. That is the actuarial consequence worth sitting with. Personal-property premium today is, in effect, a fixed multiple of dwelling premium, so contents exposure is priced through the dwelling rating plan rather than on its own merits. Replace the multiple with a per-household estimate and the two coverages decouple: contents relativities become estimable in their own right, and the loss ratio on Coverage C becomes measurable against an exposure base that reflects what is actually in the house.

The direction of the correction is not neutral. If the Marshall Fire figures generalize even weakly, a model that estimates contents accurately raises limits on the underinsured majority, which raises premium and raises expected loss at the same time. Whether that improves the loss ratio depends entirely on whether the rate charged for the additional limit is adequate, and the rating plan that would price it does not exist yet, because nobody has needed contents relativities while contents rode on the dwelling limit.

Corrections Are Not Ground Truth

The retraining step carries the governance consequence, and it is the part of the claim that looks most innocuous. The system displays predicted items and values, receives adjusted values back from the user, and retrains on those adjustments.

Treated as a data pipeline, that is a feedback loop with a labeling problem. A policyholder correction is a behavioral response, not a measurement. It arrives from a self-selected subset of insureds, those motivated enough to review a generated inventory and change it, and every one of them has a directional interest in the answer. An insured who wants a higher limit corrects upward; one shopping on price corrects downward. Neither is telling the model what is in the house.

That biases the model in a way ordinary validation will not catch, because the corrections are simultaneously the training signal and the only available check on it. A model retrained on adjusted values converges toward what policyholders say they own when asked by their insurer, which is a different quantity than what they own, and the divergence compounds with each retraining cycle rather than washing out.

The exposure sits with the carrier rather than the vendor. Quanata is the renamed BlueOwl, part of the State Farm family, so the intellectual property sits with an affiliate rather than the writing company. State Farm wrote $31.46 billion of homeowners direct premium in 2024, which is the scale at which a systematically biased contents estimate stops being a modelling curiosity.

A limit-setting model whose training signal is the insured's own stated valuation is, structurally, a rating variable the policyholder helps set. The NAIC's AI model bulletin expects insurers to document how third-party AI outputs are validated, and a self-reinforcing valuation loop is a hard thing to validate against anything except itself.