Maksims Volkovs, the Layer 6 co-founder TD Bank installed as its chief AI scientist, is named among eleven inventors on US Patent 12,711,554 B2, granted August 18, 2026 (USPTO Official Gazette). The claim covers an insurance fraud model that fuses averaged word vectors, a five-layer walk through a claim network, and tabular exposure data, then denies follow-on transactions from every party attached to the flagged claim.

Key Takeaways

  • Five layers of breadth-first search define the feature window. The specification works an example in which the traversal from one claim reaches roughly 200 nodes, of which 70 are claims and 8 are individuals, before any feature is computed.
  • Sixty-five months from filing to grant against roughly thirty for Verisk's forgery detector three weeks earlier. Same insurance CPC code, different art unit, and the routing explains most of the gap.
  • The trigger is a denial, not a referral. Claim 1 recites a computerized action "for denying subsequent transactions from parties associated with the target claim," plural.
  • 57% of insurers name excessive false-positive and false-negative rates as a deployment barrier, and while 54% report more referrals from fraud technology, only 46% report better ones.

Patent Details

FieldValue
PatentUS 12,711,554 B2
TitleSystem and method for dynamically predicting fraud using machine learning
AssigneeThe Toronto-Dominion Bank
FiledMarch 19, 2021 (application 17/207,221)
GrantedAugust 18, 2026 (Official Gazette Vol. 1549 No. 3)
First-named inventorXiao Shi Huang; eleven inventors total, including Maksims Volkovs and Tomi Poutanen
Art unit / class3696; US class 705/4
ClassificationG06Q 40/08 (insurance), G06N 20/20 (ensemble learning), G06F 18/2148
Claims10 total, 2 independent (system and method)
FamilyContinuation 19/572,944 filed March 20, 2026, still pending

What US 12,711,554 Claims

Three feature families feed one model. Text features come from descriptions of communications between the claimant's device and the claims centre, converted so that "a vector of the array is created for each word used in the description" and then averaged into a single representation. That is averaged word embeddings, a technique already mature when the application was filed in March 2021 and published the following year, and the claim language locks it in.

Graph features come from a network in which each node is a claim, a policy or an entity, and each edge records a shared attribute with a timestamp: a shared phone number, email address or vehicle identification number. Tabular features come from claim and exposure records in a current policy database and a legacy database of past fraud investigations.

All three vectors concatenate and enter a gradient boosted decision tree ensemble; the specification names XGBoost. Claim 1 recites training "with each new decision tree in the gradient boosted decision trees trained on residuals of previous decision trees," then recites the consequence: a computerized action "for denying subsequent transactions from parties associated with the target claim."

The background section explains why a bank built this rather than licensing it: existing models, it argues, are "provided to an entity by an external vendor without any customization available," and "the black box that shrouds the inner mechanisms of the fraud model makes it nearly impossible to update the model in accordance with growing fraud practices." TD had bought Layer 6 outright in January 2018, which put it in the minority of carriers building predictive models in-house rather than buying them.

The prosecution history explains the five-year wait. TD issued in sixty-five months through art unit 3696, US class 705/4, insurance; Verisk's forgery detector issued in about thirty through art unit 2494, computer security. Both carry the same G06Q 40/08 classification, so framing an invention as a security mechanism rather than insurance data processing changed the examining corps, and with it the eligibility posture governing insurance AI patents since the Section 101 reset.

How the Neighborhood Features Encode Collusion

The graph half splits into two categories, and the split is where the actuarial content sits. Count features describe the target node itself: how many policies, phone numbers and individuals attach directly to this claim. A tabular fraud model already approximates most of that.

Neighbourhood features describe everything the traversal reaches: the share of nodes that are claims, the largest node degree found, the count of phone numbers created within the past year, and whether any node "linked to fraud investigations with a high suspicion rating." In the worked example, 31 of the phone numbers inside a 200-node neighbourhood were created in the past twelve months.

The hop count is the design decision. A staged-accident ring does not look anomalous one claim at a time, because each participant files a plausible soft-tissue claim with a plausible repair invoice. The ring becomes visible only where participants reuse infrastructure, and reuse shows up two or three edges out rather than at the target node. Setting depth at five layers reaches past the pairs a duplicate-detection rule already catches, and 34% of carriers named organized fraud among the areas of greatest detection impact.

The reserving consequence comes from the trigger rather than the score. Undetected fraud does not stay in the claims department; it settles, enters paid losses, ages into the experience period, and gets trended into the next rate indication as though it were legitimate severity. The Coalition puts US insurance fraud at $308.6 billion a year, of which $45 billion is property and casualty. A pricing actuary cannot strip that from the triangle, because the fraudulent claim carries no flag.

What an automatic denial changes is timing. A conventional referral runs an investigation while the claim stays open, carrying a case reserve that develops normally. A model that denies subsequent transactions truncates the claim at the transaction, so the case reserve comes down on a model score rather than on an adjuster's disposition. Reversals then reopen and re-establish that reserve in a later accident-year evaluation, which means the model's error rate lands in reported development rather than in any fraud statistic anyone tracks.

The Hub Node Problem

A shared phone number is not evidence of anything. Body shops, tow operators, rental desks and plaintiff firms all appear on hundreds of unrelated claims, and their contact details are exactly the attributes the specification uses to draw edges. Every claimant who used one busy collision centre inherits that centre's entire claim history as a two-hop neighbour.

The claim language offers no correction. The traversal is bounded by layer count, not by node degree, and "the largest degree of connection" inside the neighbourhood is itself listed as a feature, so a high-traffic hub raises the score of everyone attached to it rather than being discounted as uninformative. The high-suspicion-investigation feature works the same way: a claimant is scored partly on whether someone several edges away was once investigated, and a past investigation is not a past adjudication.

The wording of the triggered action makes the exposure wider than one claimant. Claim 1 denies subsequent transactions from parties associated with the target claim, plural, and generates alerts on the devices processing them. A legitimate policyholder whose only connection to the network is a repair shop can have downstream payments stopped, and so can the shop.

Carriers already see the shape of this. Respondents put excessive false-positive and false-negative rates third among deployment barriers at 57%, behind limited IT resources at 86% and data quality at 60%. The benefits table is sharper still: 54% said fraud technology produced more referrals, while only 46% said it produced higher-quality ones. Volume is easier to add than precision.

The regulatory frame has moved to meet exactly this decision. The NAIC's AI model bulletin defines an Adverse Consumer Outcome as a decision that "adversely impacts the consumer in a manner that violates those standards," and names the Unfair Claims Settlement Practices Model Act among the standards in scope. Twenty-five jurisdictions had adopted it as of April 2026, with California, Colorado, New York and Texas running their own guidance, a convergence the site has tracked on claims handling specifically.

None of those regimes ask a carrier to justify a score. They ask it to justify a denial, to the person denied. A model whose most discriminating input is the investigation history of strangers five edges away produces a reason code that reads, in substance, as guilt by adjacency, and the party holding the burden of explaining it is the carrier, not the bank that patented the traversal.

Further Reading

Sources