Hartford Patents an Outlier Engine for Underwriting and Fraud

Hartford Fire Insurance Company's US Patent 12,700,040 B2, granted August 4, 2026, scores any data point by how far its attribute grouping sits from the norm across correlated data sources. It then feeds that outlier back into the model as a fresh input, and routes the result to an underwriting decision, a fraud investigation, or a claims outcome. Trade coverage defaults to the fraud frame because that is the plainest use Hartford names. The consequential reading sits one layer down.

Key Takeaways

• An outlier flag that influences acceptance is risk selection, the same function a filed rating variable performs, without the actuarial justification and disparate-impact testing a named variable clears before it touches price.
• "Typical" is whatever the model currently computes it to be. A feedback loop that reinjects flagged outliers means that definition shifts over time, so there is no fixed variable to hand a regulator.
• More than 20 US jurisdictions had adopted the NAIC AI bulletin by mid-2026, including Connecticut, Hartford's own domiciliary state, which issued Bulletin No. MC-25 in February 2024.
• A changing SIU referral rate changes loss emergence. Claims held for investigation pay later and carry an added expense layer, and that shift is indistinguishable from adverse development inside a standard triangle.

Patent Details

Patent Number	U.S. 12,700,040 B2
Granted	August 4, 2026 (Official Gazette week 31)
Assignee	Hartford Fire Insurance Company
Inventors	Arthur Paul Drennan III; Tracey Ellen Steger
Classification	CPC G06Q 40/08 (insurance subclass of finance and business methods)
Predecessor	U.S. 11,244,401 B2, "Outlier System for Grouping of Characteristics" (same inventors)
Downstream routing	Underwriting decision, special-investigations referral, or claims outcome

Grouping, Separation, and a Feedback Loop

The lineage runs back to an earlier grant from the same inventor pair, Arthur Paul Drennan III and Tracey Ellen Steger. Their prior US Patent 11,244,401 B2 describes a system that receives data from multiple external and internal sources, correlates selected parameters including derived characteristics, and models the data by identifying attributes associated with particular groups based on the groupings' relative separation from typical groupings when spatially graphed.

The August 2026 grant extends that with a loop. Once the system flags a grouping as an outlier, it reinjects the flagged data as a new input source, so an anomaly surfaced in one pass becomes part of what the next pass correlates against. The output then routes to one of three downstream paths: an underwriting decision, a special-investigations referral, or a claims outcome.

Nothing in that mechanism requires a fraud signal. Both patents sit under CPC G06Q 40/08, the insurance subclass of finance and business methods, rather than a fraud-specific classification, and the placement matches what the claim language does. Separation from a typical grouping is a general statistical property; the routing step decides what a given outlier means for a given policy or claim.

A homeowners application whose attribute grouping sits far from the norm can be routed to underwriting the same way a claim with an unusual combination of loss-type, provider and timing characteristics can be routed to an investigator. The patent is a segmentation engine wearing an anomaly-detection label, and segmentation is exactly the tool a pricing actuary already knows how to regulate, just not usually one that discovers its own groups without being told which attributes to watch.

Risk Selection Without a Filed Variable

New York's Department of Financial Services drew the clearest line. Circular Letter No. 7 requires insurers using external consumer data or AI in underwriting to evaluate the extent to which those sources correlate with protected-class status, and to run a three-step test before deployment and on a cadence afterward: check for disproportionate adverse effects, determine whether a legitimate business reason explains any disparity, and search for a less discriminatory alternative.

It also requires insurers to explain at all times how the system operates, and entitles a denied consumer to a detailed explanation within 15 days, with no exception for proprietary vendor logic.

That framework was built for a named variable: a credit-based insurance score, a territory factor, a motor-vehicle-report tier, something a regulator can point to and test. Hartford's outlier score is not that. It is the output of an unsupervised process that decides for itself which attributes, in whatever combination the current data pull happens to correlate, define a typical grouping, then measures distance from that self-defined center.

There is no single variable to hand a regulator, because the variable is whatever the model currently treats as normal, and the feedback loop means that definition shifts as flagged outliers get reinjected. A carrier can test the final routing decision, but testing an output is a weaker check than testing the input variable, and the input level is where the NAIC's AI model bulletin expects governance to operate.

The exposure is concrete. More than 20 US jurisdictions had adopted the bulletin by mid-2026, Connecticut among them, so the documentation standard this design resists applies first to the department that knows Hartford's filings best. The industry has already seen the cost of an algorithmic factor without a clean paper trail: State Farm is defending a race-discrimination suit over claims-handling algorithms plaintiffs argue functioned as an unfiled proxy (Bloomberg Law, 2026). An engine never filed as a rating variable does not avoid that exposure; it moves the fight from a rate hearing to a market-conduct exam.

Reserving carries a separate, quieter version of the same problem. A routing decision that sends a claim to special investigations rather than ordinary adjustment changes when paid losses and allocated expense emerge, because a claim held for review pays out later and the investigation adds an expense layer.

Development factors assume the mix flowing through an accident period develops in a reasonably stable pattern. If the flag rate moves, because the model retrains or the separation threshold gets tuned, the SIU referral rate becomes a variable the reserving actuary has to monitor. Development that looks like adverse emergence could be a change in how many claims got pulled onto the slower track before reaching the same disposition, and that is invisible in the triangle.

The Tension With Hartford's Own Transparency Record

Hartford published a voluntary Algorithmic Impact Assessment in February 2026, the first among top-20 P&C carriers, disclosing bias audits for named variables including ZIP code, age and property type across its production models (site coverage, May 2026).

That disclosure works precisely because ZIP code, age and property type are named variables with fixed definitions an auditor can test for protected-class correlation year over year. An unsupervised outlier score has no equivalent fixed definition to publish. The carrier that set the industry's transparency bar for named variables now holds a patent on a mechanism that, by design, does not have one to disclose.

The same claim spans two uses whose governance obligations are not close to equivalent. Using an outlier flag to catch a malformed record before it enters an approved rating algorithm is data hygiene, close to validation a data warehouse already runs. Using the same flag as an input that changes whether a policy is accepted, at what tier, or whether a claim gets investigated is a decisioning function producing exactly the adverse consumer outcome the NAIC bulletin defines.

Drafting a claim broad enough to cover both is standard practice, a pattern the site found across the Section 101 guidance shift. The compliance question is whether anyone tracks which decisions the score is touching once it leaves the data-cleaning pipeline.

The volume running through that seam is growing. Business Insurance posted a 5% premium increase and an 89.3 underlying combined ratio in the second quarter, while Personal Insurance improved to 86.3 despite a 7% decline in written premium, results management attributed partly to automation and AI-enabled underwriting. Growth built on AI-assisted underwriting raises the number of decisions running through that governance process at the same pace it raises the number of policyholders affected by it.