Fraud, Identity & AMLResearch Lab
Resistant AI: what a document-fraud flag can tell a lender
Document forensics examines how a file was made or changed. A warning still needs to be connected to reliable facts and a fair, timely financial decision.
THE RESEARCH LIBRARY
Deep-dive library
766 deep dives in 24 categories · last updated
50 matching deep dives
Research Lab: these deep dives are AI working drafts unless tagged Reviewed research. Read how review works.
AI & Tech
Financial AI, data and technology, including tools, operating benefits and their limitations.
Fraud, Identity & AMLResearch Lab
Sardine: what transaction-history research can show about fraud
Sardine reports gains from adding sequence-model features to card-fraud scoring. Experimental detection metrics, live deployment and the cost of declined purchases remain different questions.
Fraud, Identity & AMLResearch Lab
Nasdaq Verafin: from research summaries to automated investigations
Verafin’s public materials show a move from gathering evidence to clearing or escalating alerts. That changes what faster work means and who controls the decision.
Fraud, Identity & AMLResearch Lab
Oscilar: from a fraud score to an investigated decision
Oscilar markets both risk-decisioning models and agents that help investigate alerts. Their effects become clearer when detection, evidence, human decisions and legitimate customer access are measured separately.
AI, Data & Technology GovernanceResearch Lab
Passkeys in banking: a stronger login still leaves a recovery problem
Passkeys resist credential phishing through public-key authentication tied to the genuine service. Enrollment, lost devices, fallback access and payment authorization remain distinct parts of the customer’s experience.
AI, Data & Technology GovernanceResearch Lab
Microsoft Copilot Studio: how an answer becomes an action
Copilot Studio connects information, agent behavior and external services. Authentication, connector permissions and execution records determine what an agent can actually read or do; conversation volume alone says little about the result.
Decisioning & Model ToolsResearch Lab
When a loan changes an outcome: the problem of proving cause
Borrowers can improve after taking a loan without the loan causing the improvement. Credit-builder loans, randomized second looks and medical-debt relief show how a study’s treatment, comparison group and take-up define what its results can establish.
Company & Platform ProfilesResearch Lab
Antithesis: making hard-to-find software failures repeatable
Built by former FoundationDB developers, Antithesis puts software inside a controlled simulation to find and replay failures. Its funding and customer examples show growing interest, while the tests remain bounded by what they represent.
Company & Platform ProfilesResearch Lab
Cursor and Anysphere: AI coding after the SpaceX acquisition
Cursor grew from code assistance into a platform whose agents can work across a software project. Its completed SpaceX acquisition, financial-sector customer accounts and cloud execution model show how the product and its data boundaries have expanded.
AI, Data & Technology GovernanceResearch Lab
Spring Labs: AI compliance software for banks adds a regulatory library
Spring Labs sells AI software that reviews bank customers' calls, emails and chats to find complaints and check service quality. On October 7, 2026, it acquired Winnow and launched SpringGRC, a compliance-management platform now in limited beta.
Decisioning & Model ToolsResearch Lab
Affirm’s hybrid underwriting: what the checkout experiment shows
Affirm reported more completed purchases after adding transformer-derived credit information to its underwriting. The result concerns a defined experiment; repayment, explanations and the economics of those extra purchases remain separate parts of the evidence.
Fraud, Identity & AMLResearch Lab
Alloy: how identity checks become an account-opening decision
Alloy connects identity data, fraud scores and review tools. Its customer examples show how that can change account opening, while leaving important questions about losses, cost and customer experience.
AI, Data & Technology GovernanceResearch Lab
Amazon Bedrock Guardrails: what an AI safeguard actually checks
AWS offers configurable filters around AI interactions. The important boundaries are which content reaches them, what a grounding score means and how many legitimate requests are blocked alongside harmful ones.
Decisioning & Model ToolsResearch Lab
Automated mortgage validation: what a verified input does and does not settle
Fannie Mae’s validation service can reduce paperwork and defined loan-sale exposure. Its latest employment-verification update shows why the evidence, closing date and scope of contractual relief still matter.
Fraud, Identity & AMLResearch Lab
BioCatch: what digital behavior can reveal about a scam
BioCatch analyzes how people interact with financial services. Visa’s announced acquisition highlights the technology’s role, but behavioral signals, research prototypes and proven customer outcomes remain different things.
Decisioning & Model ToolsResearch Lab
Conformal prediction: what a 90% forecast range really promises
Conformal prediction adds calibrated ranges to model forecasts. A worked cash-flow example shows how those ranges are built, why average coverage can conceal weak subgroups, and why changing conditions and delayed outcomes limit the promise.
Decisioning & Model ToolsResearch Lab
Credit-model fairness: why equal error rates can produce unequal outcomes
Two groups can have identical error rates and different approval rates. A worked example separates calibration, mistaken rejection and selection, and shows why statistical equality cannot decide a legal or ethical question.
Fraud, Identity & AMLResearch Lab
Featurespace: detecting fraud without stopping genuine payments
Visa completed its purchase of Featurespace in 2024. A NatWest case shows how behavioral analytics can influence payment decisions, and why detected fraud value is different from losses prevented.
AI, Data & Technology GovernanceResearch Lab
Federated learning: what shared financial models still reveal
Research shows how institutions can train together while keeping original records in place. Privacy depends on the exchanged updates, record matching and model outputs, while usefulness depends on whose data and outcomes the model represents.
AI, Data & Technology GovernanceResearch Lab
Fiddler: what AI monitoring can reveal, and what it cannot prove
Fiddler describes tools for observing predictive models, tracing AI agents and filtering requests. The distinction between an early warning, a confirmed error and a successful correction explains both their usefulness and their limits.
AI, Data & Technology GovernanceResearch Lab
Financial AI research: when a citation supports the answer and when it does not
Retrieval lets an AI system bring filings and policies into an answer. Accuracy still depends on which version it finds, which exceptions survive extraction, and whether each claim follows from the evidence.
AI, Data & Technology GovernanceResearch Lab
Financial cloud resilience: keeping services running and records correct
Cloud services offer flexible capacity while linking financial firms to shared infrastructure. A second region, a backup and a route to another provider address different failures, with different costs and limits.
Decisioning & Model ToolsResearch Lab
Mortgage credit scores: new choices, different numbers and unchanged questions about risk
VantageScore 4.0 is available for eligible Enterprise loans and Fannie Mae has aligned pricing with Classic FICO. FICO 10T remains a future delivery option. Model choice, report coverage and the meaning of a cutoff are separate parts of the transition.
Fraud, Identity & AMLResearch Lab
NICE Actimize SAM: what happens after a transaction raises suspicion
NICE Actimize combines rules, customer comparisons and alert ranking in its monitoring products. Following an alert into an investigation shows why fewer alerts can coexist with unchanged work or weaker detection.
Decisioning & Model ToolsResearch Lab
Nova Credit: what cash-flow data adds before a lending decision
Cash Atlas turns permissioned account data into reports and attributes; NovaScore adds a proprietary risk score. Customer announcements show where those tools are being applied, while seasonal income and connection costs explain the limits of a richer data feed.
Payment Networks & SettlementResearch Lab
Payment tokens: what changes when a card moves into a wallet
A wallet can use a restricted substitute for a card number. Its protection depends on enrollment, evidence about each payment and the ability to update or suspend the credential later.
AI, Data & Technology GovernanceResearch Lab
Post-quantum finance: the long transition behind stronger cryptography
Central-bank experiments show how post-quantum signatures can work in payment systems. Moving from a successful test to durable protection involves data lifetimes, software, counterparties and the capacity to process payments.
Decisioning & Model ToolsResearch Lab
Provenir: connecting the decisions behind credit, fraud checks and customer service
Provenir combines data, rules, models and AI-assisted workflows across the customer relationship. Its broad product claims become meaningful only when the task, customer experience and economic comparison are kept distinct.
Fraud, Identity & AMLResearch Lab
Quantexa: why connecting financial records changes an investigation
Entity resolution brings scattered records together; network analysis shows their relationships. Useful context depends on keeping unrelated people and businesses apart.
Decisioning & Model ToolsResearch Lab
SAS Model Manager: how a financial model reaches the decision it supports
SAS links model deployment with records of ownership, validation and permitted use. The distinction matters when a technically successful release delivers the wrong version, stale data or an answer that arrives too late.
Fraud, Identity & AMLResearch Lab
SentiLink: why an identity-risk score is only part of account opening
SentiLink’s materials distinguish identity verification, fraud signals and credit evaluation. A complete result also depends on false positives, usable explanations and a way to resolve mistakes.
Fraud, Identity & AMLResearch Lab
Socure Sigma Synthetic: what an identity-risk score can tell you
Socure’s synthetic-identity tools combine scores, reason codes and verification options. Their value depends on what happens after an identity is flagged, and on keeping a relative risk ranking distinct from proof of fraud.
AI, Data & Technology GovernanceResearch Lab
Synthetic financial data: useful imitation, conditional privacy
Generated records can help test financial software and study relationships without distributing original files. A realistic-looking dataset still needs separate evidence about privacy, statistical usefulness and performance on real outcomes.
Decisioning & Model ToolsResearch Lab
Taktile: what an automated financial decision contains
Taktile combines rules, data, case work and AI agents. Its financial-spreading benchmark shows one kind of measured result; completed applications, consequential errors and time released tell different parts of the story.
Fraud, Identity & AMLResearch Lab
ThetaRay: why an unusual payment still needs an explanation
ThetaRay describes monitoring that uses behavioral and network patterns to identify unusual activity. A payment example shows the difference between detecting a change, establishing suspicion and completing an investigation.
AI, Data & Technology GovernanceResearch Lab
UAD 3.6: a new appraisal format and a conditional transition
Fannie Mae and Freddie Mac retain the November 2, 2026 mandate while allowing approved Sellers extra time. The redesign makes appraisal information more structured, but comparable data still requires judgment about comparable homes.
Page 1 of 2 · 1–36 of 50 articles