US consumers reported $15.9 billion in fraud losses to the FTC in 2025, and the costliest schemes now persuade customers to send the money themselves, so the login is genuine and classic fraud controls see nothing wrong.
AI in fintech is how risk teams are responding, from injection attack detection at onboarding to scam scoring before an instant payment is released. The rules moved in 2026 as well: SR 26-2 superseded SR 11-7 and excluded generative and agentic AI, and FinCEN proposed a rewrite of AML program requirements.
The evidence here is reported and estimated losses from US agencies and regulatory texts current to November 2026. It does not measure any single institution's exposure, and true losses are likely higher than the reported figures.
None of this requires a risk committee to choose a vendor first. It requires the committee to know what each model is allowed to decide, how its threshold was priced and who governs the parts that fall outside SR 26-2. Three questions surface all of it, each answerable with one named artifact.
Ask whether each use case automates, augments or only advises. Identity checks can auto-approve and send failures to review, sanctions screening keeps a deterministic floor no model can override, and a SAR is always decided and filed by a person. The mode sets the explainability, testing and audit evidence needed later, so it comes before model choice. The artifact is a use case register with a decision mode against each line.
Ask what the alert threshold costs in total, not how many alerts it removes. Every false positive carries review cost and customer friction, and every miss carries a fraud loss or regulatory exposure. A model that halves alerts but misses one mule network can cost more than it saves. The artifact is a total cost calculation with precision reported at the investigator capacity line.
Ask where the copilot that drafts alert summaries and SAR narratives sits in the governance framework. SR 26-2 excludes generative and agentic AI, so the copilot needs its own controls: grounding in retrieved case facts, citations to evidence, logged prompts and outputs, and an investigator who edits and approves. The artifact is a copilot control sheet with an audit trail behind it.
Decide whether the system automates, augments or advises before choosing a model, because the mode sets the explainability, testing and audit evidence you will need under BSA, OFAC and consumer protection law.
Gartner advises pairing presentation attack detection with injection attack detection and image inspection. Add device intelligence and document forensics, then tag deepfake cases with FinCEN's key term.
Use device, session, payee and behavioral signals such as an active call during a session, and scale the response from a warning to a confirmation question to a hold with an agent call.
Rate materiality, document validation, fairness tests and reason codes for scoring models, and put copilots under broader risk governance with grounding, citations and logged prompts and outputs.
AI in fintech verifies identities, scores payments and logins for fraud, prioritizes AML alerts, screens sanctions lists and detects scams in progress. Gradient-boosted models handle real-time scoring, graph models find mule networks and synthetic identity clusters, and LLMs draft alert summaries and SAR narratives.
Humans stay accountable for SAR filings, account closures and adverse decisions, and the program must still satisfy BSA/AML, OFAC and consumer protection law.
Yes, when the checks are layered. Presentation attack detection catches masks and screen replays, and injection attack detection catches virtual cameras and tampered video streams, which Gartner reported rose 200% in 2023.
Document forensics and device signals add further checks. FinCEN's 2024 deepfake alert lists red flags and asks institutions to tag related SARs. No single check is reliable alone.
Banks score each instant payment before release using device, session, payee and behavioral signals. Scam models look for signs of coaching, such as an active phone call during a session or a first payment to a crypto exchange.
Responses are tiered, from a warning to a confirmation question or delay to a hold with an agent call. Inbound monitoring catches mule accounts receiving the proceeds.
Yes. On April 17, 2026, the Federal Reserve, OCC and FDIC issued revised model risk management guidance as SR 26-2, OCC Bulletin 2026-13 and FDIC FIL-15-2026. It supersedes SR 11-7 and the 2021 BSA/AML model statement, narrows the definition of a model, carves out deterministic rules and excludes generative and agentic AI. The agencies plan a separate request for information on AI.
Supervised models re-rank rule-generated alerts using past dispositions and SAR outcomes, so investigators work the riskiest alerts first, and graph features add network context single-account rules lack. The cost pressure is real: LexisNexis Risk Solutions estimated financial crime compliance in the US and Canada at $61 billion a year, with labor the largest component.
Measure success as precision at investigator capacity and total cost, not raw alert count.
A 17-minute guide covering the 2025 and 2026 threat data, seven use cases from KYC to SAR copilots, a comparison of model approaches, the regulatory position as of November 2026, explainability and fairness testing, a total cost formula for alert thresholds, a reference architecture, KPIs, a 12-month roadmap and a seven-dimension readiness scorecard for financial crime programs.
It covers reported and estimated losses from US agencies, the Federal Reserve System, Nacha, The Clearing House and the UK Payment Systems Regulator, plus regulatory texts checked through mid-2026. Actual losses are likely higher than the reported figures.
It does not measure any single institution's exposure, vendor names are examples rather than endorsements, and the FinCEN program rule it describes is still a proposal.
Drop your details and we'll send AI in FinTech: KYC, Fraud and Scam Detection straight to your inbox - no spam, unsubscribe anytime.
Bring your current onboarding, monitoring and payment fraud stack and the model inventory behind it. We will work through where each control sits in the payment flow, what it is allowed to decide and how it maps to SR 26-2. SECTION 7 - FAQ - 5 to 8 questions
Book a fraud stack review