A rising, two-sided cost
record U.S. consumer fraud losses reported in 2025, up from $12.5B in 2024 (FTC, via The Spokesman-Review).
Source: FTC / Spokesman-Review
estimated annual cost of false declines to online merchants — far above actual card fraud (~$40.8B/yr), showing why the false-positive trade-off matters (Ravelin, citing Ekata).
Source: Ravelin / Ekata
of global banks already use AI/ML for fraud prevention and detection (562 financial-services respondents, Feedzai).
Source: Feedzai
generative-AI fraud-detection market from 2026 to 2035 (~26.9% CAGR) — heavy investment on both sides of the fight (Next Move Strategy Consulting).
Source: Next Move Strategy
transaction-direction error rate measured on CONE RED’s production bank-statement automation across 21 banks — an example of measured (not guaranteed) accuracy in financial document intelligence (first-party).
Source: CONE RED (first-party)
How AI fraud detection works — the key terms
- Anomaly / outlier detection
- An unsupervised technique that learns what normal behavior looks like for an account, device, or population and flags activity that deviates from it. Because it does not need labeled examples of fraud, it is the main defense against novel or fast-changing attacks that supervised models have never seen.
- False-positive rate
- The share of legitimate transactions or customers a system wrongly flags as fraudulent. High false-positive rates create customer friction, abandoned carts, and lost lifetime value; they are the counterweight to the false-negative rate (missed fraud). Tuning the threshold between them is a business decision about risk appetite, not a purely technical one.
- KYC / AML
- Know Your Customer and Anti-Money-Laundering: the regulated processes for verifying customer identity, screening against sanctions and watchlists, and monitoring transactions for suspicious patterns. AI is increasingly used to reduce alert backlogs and triage suspicious-activity reports, but the underlying obligations and audit requirements remain the institution’s responsibility.
- Behavioral biometrics
- Passive signals about how a user interacts with a device — typing rhythm, mouse or touch dynamics, how they hold a phone — used to distinguish a genuine account holder from an impostor or bot. Valuable for detecting account takeover because it does not rely on credentials that may already be compromised.
- Model-risk management (MRM)
- The governance discipline — codified for U.S. banks in guidance such as SR 11-7 — for developing, validating, documenting, and monitoring models used in decisions. For fraud and credit models it covers independent validation, ongoing performance monitoring, bias testing, and clear ownership, so a model’s decisions can be explained and defended to regulators.
How to evaluate a fraud-AI builder — a checklist
Measured on your data, explainable to a regulator, monitored for drift, and human-in-the-loop.
- Ask for measured performance on your own data, expressed as a full confusion picture — detection rate AND false-positive rate at a stated threshold — not a single “accuracy” or “detection” headline.
- Confirm explainability: can the system produce a human-readable reason for every decline or alert, sufficient for adverse-action notices, analyst review, and a regulator’s audit?
- Require a model-risk and monitoring story: independent validation, drift and performance monitoring, retraining cadence, bias/fairness testing, and documented ownership per guidance like SR 11-7.
- Probe the adaptive-adversary plan: how are new fraud patterns detected and fed back into the models, and how fast can the system adapt when attackers change tactics?
- Check real-time constraints: can it score transactions within your latency budget (often tens of milliseconds) at peak volume, and how does it degrade under load?
- Map data needs honestly: what labeled fraud, historical, and third-party signals are required, who owns the data, and how is PII handled, retained, and secured?
- Clarify build vs. buy and integration: how it fits your existing decisioning, case management, and core systems — and whether you get a black box or a system your team can inspect and extend.
- Walk through the human-in-the-loop design: where analysts review, override, and improve the model, so automation augments the fraud team rather than silently replacing judgment.
Frequently asked questions
How does AI fraud detection actually work?
Most systems combine several techniques. Supervised machine learning trains on large volumes of labeled past transactions (fraud vs. legitimate) to score new ones. Anomaly/outlier detection learns normal behavior and flags deviations, which catches novel fraud that has no labeled examples yet. Real-time scoring applies these models in milliseconds at the point of a transaction. Graph and network analysis links accounts, devices, and payees to expose fraud rings and mule networks. Behavioral biometrics adds passive signals like typing and navigation patterns. In practice the strongest programs layer these rather than relying on any single model.
What are the main use cases?
Four dominate. Payments and card fraud — scoring transactions to block stolen-card and card-not-present abuse. Account takeover — spotting when a legitimate account is being operated by an impostor, often via device and behavioral signals. KYC/AML and sanctions — verifying identity, screening watchlists, and monitoring for money laundering while reducing false alerts. Application and synthetic-identity fraud — catching fabricated or stitched-together identities at onboarding. Each has different data, latency, and regulatory characteristics, so a system strong at one is not automatically strong at another.
Why is the false-positive versus false-negative trade-off so important?
Every fraud model sits on a threshold. Loosen it and you miss fraud (false negatives), taking direct losses and chargebacks. Tighten it and you decline good customers (false positives), creating friction, abandoned transactions, and churn — and false declines are estimated to cost merchants far more in aggregate than the fraud itself. There is no universally “correct” setting; the right threshold depends on your margins, customer base, and risk appetite. Any serious evaluation looks at both error types together, on your data, not a single detection percentage.
What do regulators and model-risk teams expect?
Models that influence financial decisions are expected to be explainable, validated, and monitored. That means a human-readable reason for each decision (needed for adverse-action notices and analyst review), independent validation, ongoing performance and drift monitoring, bias/fairness testing, and clear documentation and ownership — in the U.S., aligned with guidance such as SR 11-7. A pure black box that cannot explain a decline is a compliance and reputational liability, however good its raw accuracy looks.
Where do large language models fit in?
LLMs are strong on unstructured and semi-structured data rather than raw transaction scoring. They help summarize and triage alerts and cases, draft suspicious-activity narratives, extract entities and intent from documents and messages, and reason over scam text. They complement — not replace — the numeric models that score transactions in real time. Because they can also be wrong or fabricate, LLM output touching fraud or compliance decisions needs guardrails, source-grounding, and human review rather than blind automation.
How do we choose a builder, and what are the red flags?
Favor builders who measure performance on your own data, explain their models, and can articulate a model-risk, monitoring, and retraining story — and who design humans into the loop. Weigh custom builds against off-the-shelf platforms based on how specific your fraud patterns and integrations are. Red flags: a guaranteed detection rate (fraud is adversarial and changes — no honest builder guarantees it), a black box with no explainability, and no plan for monitoring, drift, or model risk. As a reference point, CONE RED builds custom AI for financial-services operators — including data and document-intelligence systems — typically running a feasibility assessment in about six weeks and reaching production near 90 days, and it measures accuracy rather than guaranteeing it. Whoever you choose should be comfortable being held to that standard.
CONE RED’s 0.24%/21-bank figure is a first-party result from financial document intelligence, cited as an example of measured accuracy — not a fraud-detection guarantee. Fraud is adversarial and changes; no honest builder guarantees a detection rate, and every consequential decision should be explainable and reviewable.