The Automated Life Underwriting Patent Swiss Re Built and Sold

The machine learning chain at the center of US Patent 12,705,675 does not read an applicant. It reads a contract. Granted on August 11, 2026, the claims describe a series of models that parse an existing life policy into a parse tree, pull out its premium and exclusion parameter values, and hand the result to a hidden pattern detection unit "configured to trigger automated under-writing processes" (US 12,705,675 B2, claim 1). The applicant of record is a Schaumburg term life brokerage whose entire portfolio is this one file.

Key Takeaways

• 20 claims issued, two of them independent, and the granted independent claim never names life insurance. That narrowing waits for claim 10, which limits the invention to term life and health risk-transfer structures.
• Rejected under Section 101 on October 28, 2025, the claims were allowed on April 7, 2026 after the applicant deleted the policy-generation step from claim 1 and added one architectural limitation in its place.
• $156 billion of life benefits written across roughly 350,000 families is the book AccuQuote brings to an engine whose dependent claims identify "offers providable to a policyholder" inside an existing contract.
• $580.8 million deployed buying 1,310 life policies in 2025 is the business of Abacus Global Management, which took this patent as loan collateral in April 2025 and bought AccuQuote outright that October.

What the Granted Claims Cover

The file is a bypass continuation of PCT/EP2023/069634, filed July 14, 2023, which claims priority to Swiss application 000839/2022 (USPTO Patent Center, application 18/544,018). Claim 1 covers the system, claim 12 the corresponding method.

Patent number	US 12,705,675 B2
Title	Automated, parameter-pattern-driven, data mining system based on customizable chain of machine-learning-structures providing an automated data-processing pipeline, and method thereof
Applicant of record	AccuQuote, Inc. (Schaumburg, Illinois)
Inventors	Thomas Young, Rory Creedon (both Zurich)
Swiss priority	000839/2022, filed July 14, 2022
US filing	December 18, 2023 (application 18/544,018)
Granted	August 11, 2026
Claims	20 total, 2 independent
Classification	G06Q 40/08, G06N 20/00, G06Q 50/18, G06Q 10/0635; USPC 705/4

Granted claim 1 recites a knowledge extraction engine having "a customizable chained series of machine learning modeling structures, wherein each machine learning modeling structure of the chained series ... passes an output to a next machine learning modeling structure." What the independent claim does not do is name life insurance. The insurance subject matter sits in the dependents, where claim 10 limits the policies to "life risk-transfer structures at least including term life risk-transfer structures and/or health risk-transfer structures." Claim 1 itself is architecture.

That shape is the residue of a fight the file wrapper documents. The examiner issued a final rejection on October 28, 2025 holding all 20 claims ineligible under 35 U.S.C. 101, on the ground that they "cover concepts of fundamental economic principles or practices ... but for the recitation of generic computer component."

The applicant then struck the insurance-specific output from claim 1, which had required "generating appropriate new digital risk-transfer policies," and replaced it with the topology, each model passing its output to the next. The examiner allowed the case on April 7, 2026, quoting that limitation back.

The applicant surrendered a claim to generating insurance policies and received a claim to a pipeline shape, which cut the opposite way from how it reads: granted claim 1 now covers a chained document-parsing model that trips an underwriting trigger in any line of business. It matches the site's review of the Section 101 reset, where eligibility survived on recited computational structure rather than insurance utility.

A Parser Pointed at an In-Force Book

The dependent claims describe a document-processing pipeline rather than a mortality model. Parser structures split a policy into "characters, words, and string of words" and store a parse tree. Identifier structures locate the language "defining condition parameters indicating offers providable to a policyholder," and linker structures map what they find "to one or more parameterized queries which are executed against a standardized database of customer data." Claim 17 enumerates the extracted fields: deductible values, risk-transfer type definitions, policy limits, exclusions, and riders.

Read end to end, the system takes a book of existing policies and asks, contract by contract, what better terms the person holding each one could be offered today. The actuarial consequence sits in who is able to say yes. AccuQuote brings $156 billion of life insurance benefits written across roughly 350,000 families and almost $340 million of lifetime premium (Abacus Global Management, October 6, 2025). Point claim 4's identifier structures at a book that size and the output is a ranked list of policyholders whose current contract is worse than what the market would write them now.

Only part of that list can act on it. A policyholder whose health has deteriorated since issue cannot replace coverage on better terms and keeps the old contract. Every pass of the engine therefore withdraws healthier-than-average lives from a block and leaves the impaired in place. That is anti-selective lapse, and it moves the surviving block's mortality against the pricing basis without anything changing about the underlying lives.

Term pricing is where this bites. Level term premiums are supported by a lapse assumption normally calibrated on surrender behavior driven by affordability, life stage and agent activity, none of which correlate strongly with insurability. An engine that reads contracts and surfaces replacement candidates makes lapse a direct function of who can still pass underwriting, so a mortality basis trended off historical persistency understates the residual block. The site traced the same validation gap from the opposite direction in mortality slippage in AI-driven life underwriting, where a model chose whom to accept rather than whom to approach.

Placement in the grant cohort sharpens the point. Of 304 US patents carrying the insurance classification granted through August 11, 2026, 51 also carry G06N 20/00 for machine learning (USPTO Open Data Portal), and that cohort is dominated by personal-lines property and casualty. MassMutual's own mortality-scoring file is not even in it, because the office classified it under health informatics for turning diagnosis codes into an underwriting classification. One patents what to compute about a person; this one patents which files ever reach an underwriter, the less-examined half of an accelerated program and the half NAIC market conduct guidance now reaches.

The Chain of Title

The recorded assignments read as a sequence with one very busy day in the middle. Young and Creedon assigned to Swiss Reinsurance Company Ltd. in November 2023. Sixteen months later, on April 4, 2025, three instruments were executed together: Swiss Re assigned the application to AccuQuote, AccuQuote granted a license back to Swiss Re, and AccuQuote pledged the same application to Abacus Global Management "as lead lender," all visible through USPTO assignment search.

Abacus released the security interest that August. Seven weeks later it did not need the collateral, because it bought the borrower, announcing the acquisition of AccuQuote on October 6, 2025. The patent granted ten months after that, still naming AccuQuote as applicant.

What Abacus does with capital is the complication. The company deployed $580.8 million buying 1,310 life policies in 2025, up 53% and 26% respectively, against total revenue of $235.2 million (Form 8-K, March 12, 2026). A life settlement is the purchase of an in-force policy from an insured whose life expectancy has shortened since issue, and the return on it improves when the insured dies earlier than the premium schedule assumed.

So the two halves of one company now select on opposite tails of the same population. The brokerage side holds an engine built to find policyholders healthy enough to be re-underwritten into a better contract. The settlements side buys the contracts of people who are not. Both depend on the same intermediate product, a structured reading of what an in-force policy actually says.

Swiss Re did not leave when it sold. The license was executed the same day as the conveyance, so a reinsurer that prices mortality for ceding carriers, a distributor positioned to generate replacement offers against their in-force blocks, and a buyer of the policies that cannot be replaced all hold an interest in one parsing engine at once.

Further Reading

• MassMutual's Newly Granted Patent Turns Medical Claims Into a Mortality Score - The other life-side patent of this grant cycle, and the contrast case: a risk model rather than a document pipeline.
• Mortality Slippage Tests AI Life Underwriting at Scale - What happens to priced mortality when a model, not an underwriter, decides which applications accelerate.
• NAIC Embeds Life AI Exam Guidance Into Market Conduct Reviews - The examination framework that now reaches predictive models sitting ahead of the underwriting decision.
• USPTO Section 101 Reset: What Changed for Insurance AI Patents - The eligibility framework that shaped the January 2026 amendment in this file.
• The AI Patent Race in Insurance - The hub tracking carrier and vendor AI patent portfolios across the sector.