Sixfold’s Patent Shows How Underwriting Manuals Are Becoming Code

When Sixfold AI filed U.S. Patent 12,561,746 in March 2024, three of the seven named inventors were the company’s own co-founders: CEO Alexander Schmelkin, along with Brian Moseley and Jane Tran. The patent was granted on February 24, 2026, the same month the company closed a $30 million Series B led by Brewer Lane, with strategic investment from Guidewire and continued backing from Bessemer Venture Partners and Salesforce Ventures. The co-founders’ names on the claims and the investors’ names on the cap table represent the same bet: that the technology for converting carrier underwriting judgment into machine-executable rules is both novel enough to patent and commercially significant enough to fund at scale.

The patent’s formal title is “Extracting rules and determining risk parameters from an underwriting manual.” What the claims actually protect is a specific pipeline: a transformer-based neural network, initially trained on multi-carrier underwriting data and subsequently fine-tuned for a specific carrier, that ingests an unstructured underwriting manual, classifies its contents into three categories of rules, and outputs those rules in a machine-readable format that can score and triage incoming submissions automatically. Underwriters interact with the extracted rules through a review interface, can modify them, and those modifications feed back into the model’s training cycle. The manual no longer lives in a shared drive. It lives in the model, updated continuously by every decision the underwriting team makes.

Reviewing AI patent filings across the underwriting vendor space over the past two years, the signal that keeps surfacing is that manuals, appetite memos, and referral notes are being treated not as process documents but as proprietary training assets. The Sixfold patent makes that treatment explicit in the claims language.

Patent Details

Patent Number	U.S. 12,561,746 (B2)
Title	Extracting rules and determining risk parameters from an underwriting manual
Filed	March 29, 2024
Granted	February 24, 2026
Assignee	Sixfold AI Inc.
Inventors	Alexander M. Schmelkin, Brian Moseley, Jane Tran, Ian Namit Hirschfeld, Gregory Hess Tourville, Laurence Brouillette, Samuel Levine
Primary Claim	Method of receiving an underwriting manual, processing with carrier-specific neural networks trained on multi-carrier data, determining insurance rules and risk parameters, generating machine-readable rules (code signals, risk signals, Q&A), displaying a UI for modification, and performing additional training based on user modifications

The Three Layers of Rule Extraction

The patent’s architecture for rule extraction defines three discrete categories, and the selection of those three categories reveals what Sixfold’s engineering team understood about how underwriting manuals actually work in production. Manual text does not organize itself conveniently into tables. It arrives as dense, cross-referenced prose that combines business classification criteria, risk factor thresholds, application requirements, and referral triggers in the same paragraph.

The first category is “code signals”: standardized business classification identifiers, specifically SIC and NAICS codes and their associated risk characteristics. An underwriting manual for a small commercial general liability account does not say “we decline NAICS code 236115.” It says “we do not write residential new construction.” The system’s job is to translate that natural-language appetite restriction into the corresponding NAICS classification, and then flag any incoming submission whose standardized code falls into that bucket. Claims 7 and 8 in the patent explicitly cover this: “identifying, in the extracted insurance rules, one or more standardized code signals, wherein the standardized code signals comprise SIC and NAICS code signals.”

The second category is “risk signals”: underwriting-specific factors that travel through the manual as conditional statements. Claims history for a general liability account. Health history for a group life submission. Occupancy type and construction class for property. Loss control survey results for workers compensation. These are the substantive terms on which the carrier prices and selects risk, and they appear in manuals in forms that require domain-specific interpretation. A transformer model fine-tuned on insurance training data handles the vocabulary of insurance risk factors well precisely because general-purpose language models cannot reliably distinguish a “claims-made” trigger from an “occurrence” trigger, or parse the significance of “habitational” as a class modifier for commercial property.

The third category is “questions and answers”: structured interrogatories that apply to specific application types. For a commercial auto submission, the manual may require answers about fleet composition, driver demographics, and loss history format. For a professional liability submission, the required answers concern coverage structure, deductible selection, and retroactive date. The system extracts these as explicit question-answer templates, which then generate the AI-assisted application review screens that underwriters use to evaluate incoming submissions.

The three categories together reproduce the logical structure of an underwriting decision: identify the business (code signals), assess the risk (risk signals), and gather required information (Q&A). What the patent protects is the automated translation of those three steps from manual prose to machine-executable rule sets.

SIC and NAICS Extraction: The Small Commercial Classification Gap

The emphasis on SIC and NAICS code extraction in the patent’s claims is not incidental. It targets the specific failure mode that makes small commercial underwriting expensive at scale: appetite leakage through misclassification.

Small commercial submissions, generally defined as accounts under $25,000 in annual premium, arrive with inconsistent or missing classification data. A plumbing contractor files as “building maintenance.” A food manufacturer describes its operations as “food distribution.” A staffing agency selects “office” as its primary occupancy. In a manual underwriting workflow, an experienced underwriter catches these mismatches through context clues in the application, supplemental questionnaire responses, and third-party data lookups. In a high-volume digital submission flow processing thousands of accounts per month, misclassifications pass through appetite filters unchecked and generate quotes for risks the carrier never intended to write.

Sixfold’s system addresses this by running SIC and NAICS identification as part of the submission analysis layer, not as a pre-filter applied before the application text is read. The model sees the full application, including the business description, prior loss narrative, and any supplemental responses, and assigns standardized codes based on semantic content rather than accepting the applicant’s self-reported classification. When the assigned code conflicts with an appetite restriction extracted from the manual, the system flags the conflict before the underwriter review stage rather than after a quote has been generated and expectations set.

By early 2026, the company had processed more than one million submissions across 40-plus lines of business for insurers representing $265 billion in gross written premium, including Zurich North America, Guardian Individual Markets, Generali GC&C, AXIS, and Skyward Specialty. Skyward Specialty reported a 35% reduction in quote response time across eleven underwriting teams. Zurich North America reported up to two hours saved per submission across more than 200 underwriters. Matthew Richardson, Global Head of Operations and IT at Generali GC&C, described the operational progression concisely: “We’ve gone from testing it to requiring its input for every quote.”

That shift from optional to required is a workflow change, not a guideline change. The guidelines still sit in the manual. The system’s contribution is making those guidelines executable at the point of each individual submission rather than at the point of underwriter review. The distinction matters more than the productivity metrics suggest, because the execution layer is now vendor-owned infrastructure rather than internal process.

The Feedback Loop: When Underwriter Decisions Become Training Data

Claim 9 in the patent is the one actuaries and audit committees should read most carefully. It covers “processing prior applications and user feedback with enhanced neural networks.” The user interface allows underwriters to modify the rules extracted from the manual. The system does not record these modifications as simple overrides; it uses them as training examples for the model’s next refinement cycle.

This creates a learning loop that has no natural stopping point. The model starts with rules derived from the manual text. Underwriters modify rules based on judgment, market conditions, and specific accounts they have seen. The model incorporates those modifications as labeled training examples. Subsequent rule extractions reflect the accumulated judgment of everyone who has interacted with the system, not just the original manual author.

In a deployment that has been running for two or three years, the model’s understanding of carrier appetite is no longer primarily derived from the document. It is derived from the document as interpreted by the carrier’s own underwriters over thousands of real submissions. The manual was the seed. The underwriter feedback is the ongoing training signal that shapes what the model actually does with the next submission that arrives.

A carrier that has operated this way for three years has a model that knows, for example, not just that “we do not write food manufacturing” but that when a food manufacturing application shows a retail sales mix consistent with a quick-service restaurant account, the senior underwriters have accepted it under the food service class. That learned distinction exists only inside the vendor’s model. It is not documented anywhere in the carrier’s manual. It cannot be extracted as a rules document. It cannot be transferred to a replacement system. The carrier’s own appetite, refined through its own team’s decisions, is now inseparable from a vendor’s proprietary model architecture and training data.

Appetite Portability and Vendor Switching Costs

The vendor lock-in literature for SaaS software focuses on data portability, integration complexity, and retraining costs. For AI underwriting systems, there is a dimension that does not apply to conventional software: model knowledge portability.

When a carrier contracts with an underwriting AI vendor and that vendor’s system learns from the carrier’s underwriters over time, the knowledge created by that learning process has an ambiguous ownership structure. The carrier owns the original underwriting manual. The carrier’s underwriters generated the feedback that refined the extracted rules. But the model that performed the learning, the architecture that encoded the feedback, and the trained weights that represent the result all belong to the vendor.

If the carrier wants to replace the system, it faces the standard data migration challenge (moving submission history, document repositories, and workflow integrations) plus a problem with no established solution: how to transfer the model’s learned understanding of carrier appetite to a new vendor’s architecture. The new vendor can ingest the manual. It cannot ingest the learned refinements accumulated over three years of production use. The new system starts from the manual alone, which is exactly where the current system started when it was deployed. The carrier has not lost data; it has lost model maturity, and model maturity in this context is indistinguishable from underwriting institutional knowledge.

This framing makes the January 2026 strategic investment from Guidewire in Sixfold’s Series B more legible. Guidewire serves roughly 500 P&C carriers globally through its core policy, billing, and claims systems. An AI underwriting layer that sits inside Guidewire’s ecosystem and learns from carrier-specific feedback creates deep integration that compounds over time. Carriers that deploy Sixfold through the Guidewire environment are building appetite knowledge into a model that is integrated with both the core system and the AI vendor simultaneously. The switching cost is not a one-time implementation effort; it is a continuously growing gap between the maturity of the current system and the starting state of any replacement.

Vendor AI discussions in the insurance industry have stayed at the level of submission throughput, straight-through bind rates, and per-underwriter premium productivity. The model knowledge portability question has not yet surfaced as a standard procurement diligence item. The Sixfold patent, by making the feedback-learning mechanism explicit in the claims, gives carrier risk officers and actuarial teams a concrete document to assess.

Referral Rules and Underwriting Authority

Sixfold’s product description includes the ability to “flag submissions outside underwriter authority based on referral rules.” In a traditional carrier underwriting governance framework, referral rules are among the most carefully controlled documents in the underwriting manual. They define which risk attributes or account characteristics require senior underwriter review, which require management approval, and which require reinsurance clearance before binding. Referral rules interact directly with the carrier’s underwriting authority matrix, which in turn affects reinsurance treaty compliance, regulatory filings, and the appointed actuary’s opinion on reserve adequacy for lines subject to schedule rating or judgment pricing.

In several states, changes to underwriting guidelines that affect which risks are eligible or ineligible require prior approval or at minimum regulatory notice. A referral rule that emerges from the model’s extraction of a decades-old manual, refined by three years of underwriter feedback, is a different kind of artifact than a referral rule that a senior underwriting officer wrote, reviewed with counsel, and filed with the department. Both affect the same downstream outcomes. The documentation trail behind them is entirely different.

The practical question for carriers is straightforward: does the governance process that applies to manually written referral rule changes also apply to referral rule changes that emerge from model refinement? If the feedback loop leads the model to tighten a referral threshold in a specific NAICS segment, because senior underwriters have been consistently referring those submissions over the past year, is that a change to the underwriting guideline? Does it require the same documentation and approval as a formal manual revision?

No current NAIC guidance, ASOP, or state regulatory bulletin has answered that question for AI-driven underwriting systems. The Sixfold patent makes the question impossible to ignore, because it describes the exact mechanism by which referral thresholds can shift through model training without any explicit revision to the underlying manual document.

The AAA Brief and the Profession’s June 2026 Framing

The American Academy of Actuaries’ AI, Data Science, and Analytics Committee published its issue brief on AI use cases in insurance and pension in early June 2026, the same week Sixfold’s patent drew attention in vendor AI discussions. The Academy’s brief identified underwriting efficiency as a primary use case for AI alongside pricing support, claims triage, and fraud detection, and it noted the expanding role of actuaries in assessing model accuracy, ensuring fairness, and evaluating data quality for AI systems that affect underwriting and rating decisions.

The brief’s implicit tension, which the Sixfold patent makes concrete, is the difference between AI as a tool that makes underwriters faster and AI as a rule engine that operationalizes underwriting judgment at scale. These are not the same function, and they do not require the same governance structures.

A tool that makes underwriters faster is subject to underwriter oversight by design: the underwriter sees every submission, exercises judgment on each one, and the AI output is one input among several. A rule engine that operationalizes appetite extracts the underwriter’s judgment from the workflow and encodes it as a model parameter. The underwriter’s role shifts from decision-maker to exception handler, reviewing only the submissions that fall outside the model’s extracted rules. The total volume of decisions the model makes without direct underwriter review grows with each submission processed.

The Academy’s brief establishes that actuaries should be involved in assessing AI model accuracy and evaluating data sources. For appetite extraction systems, that involvement needs to extend to the training data question: specifically, whether the underwriter feedback that is retraining the model represents sound underwriting judgment, and whether any systematic patterns in that feedback (for example, consistent leniency on a specific class that later shows adverse loss development) are being monitored and flagged before they affect the book.

NAIC Regulatory Implications: The AI System Program Requirement

The NAIC Big Data and Artificial Intelligence (H) Working Group has been developing its AI Systems Evaluation Tool since 2025, with a 12-state pilot underway as of March 2026 and adoption expected at the Fall National Meeting. The tool requires carriers to document what AI systems they use, what data feeds those systems, what governance and risk mitigation practices are in place, and which models meet the threshold for classification as high-risk.

An underwriting appetite extraction system that generates referral rules from a combination of manual text and learned user feedback is a strong candidate for high-risk AI classification under the framework the Working Group is building. The Working Group’s model bulletin, already adopted by several states, requires carriers using AI to maintain a written AI System Program covering the system’s purpose, data sources, testing procedures, and ongoing monitoring. For a system whose rule set evolves continuously through underwriter feedback, “ongoing monitoring” needs to mean something more specific than it does for a static pricing model. The monitoring program needs to track rule drift: whether the model’s active rules have diverged from the carrier’s current written guidelines, and by how much, since the last manual update.

The 12-state pilot includes several of the largest P&C premium states. When the evaluation tool is adopted formally at the Fall National Meeting, carriers in those states will face direct examination questions about AI systems that affect underwriting outcomes, and the documentation requirements will demand the kind of traceability between manual language and active model rules that most vendors do not yet provide in a format ready for state regulatory review. Carriers that have been treating underwriting AI as a workflow tool rather than a regulated model will need to reclassify it.

What Actuaries Need to Address in Carrier AI Programs

The actuarial implications of appetite codification systems run through three practice areas, and none of them appear adequately in the current vendor conversation about underwriting AI productivity.

Triage rate monitoring for pricing adequacy. When an AI system with learned appetite rules handles first-pass triage, the mix of submissions that reaches human underwriters changes relative to a purely manual workflow. If the model is more restrictive than the old manual process at certain NAICS codes, the bound book narrows in those segments. If it is more permissive because the learned rules capture favorable risks that manual review was declining, the book expands in specific classes. Either shift changes the data generating process underlying the loss history that pricing actuaries use to set rates. A triage system that changes the selection profile of written policies requires a change-in-classification indicator tracked from the point of AI system deployment forward, not just monitoring of overall quote-to-bind ratios.

Reserve development under changed selection criteria. The feedback loop in the Sixfold patent changes the carrier’s effective selection criteria in ways that are not visible in the underwriting manual. If the model learns over time to accept a class of risk that the manual nominally restricts, and that learned permissiveness is not captured in any governance document, the reserving actuary has no signal in the traditional paper trail that the selection criteria changed. Loss development triangles for lines where AI triage is active should carry a deployment flag as a segmentation variable, so that pre-deployment and post-deployment development experience can be analyzed separately when LDF selections are made.

ASOP No. 56 validation for a continuously learning model. The Actuarial Standard of Practice on Modeling applies to any model an actuary uses or relies on for actuarial opinions or communications. An underwriting appetite extraction system that generates referral rules used to triage submissions is, depending on how actuarial outputs incorporate those triage decisions, potentially within ASOP No. 56 scope for the appointed actuary. The standard requires understanding of the model’s purpose, design, and limitations, and documentation of whether the model is appropriate for its intended use. For a model whose rule set is updated continuously through user feedback, validation is not a point-in-time event tied to deployment. It is a recurring process tied to the model’s learning cycle, and the validation documentation needs to address how the carrier monitors whether the evolving rules remain consistent with its current reinsurance treaty language and its current state rate filings.

The carrier’s AI System Program needs to treat each of these three areas explicitly when underwriting appetite extraction is in scope. The manual that goes in can be audited against the filed rate manual and the treaty. The model that comes out after three years of underwriter feedback training requires a different audit, one that no current regulatory template or actuarial standard has fully specified. Sixfold’s patent, by protecting the feedback-learning mechanism explicitly, gives the actuarial and risk community a concrete technical document to use in closing that specification gap.

Sources