AI in Risk7 min read

AI in Fraud Detection: Risk and Governance

By Jonas Adam Mohamed Osman, known as Yonas Osman

CEO of Quantica Risk

60 Rue François Ier, 75008 Paris, France

Published 15 October 2026

How AI improves fraud detection in financial services, the false-positive trade-off, and the governance controls supervisors expect.

Executive answer

AI-based fraud detection systems have materially outperformed rule-based engines at catching novel fraud patterns, particularly in card-not-present transactions and real-time payments, where fraud typologies evolve faster than manually written rules can be updated. The trade-off firms must manage explicitly is false positives: an overly sensitive model blocks legitimate transactions and damages customer trust, while an under-sensitive one lets losses accumulate, and the acceptable balance is a business decision, not a purely technical one.

Governance of these systems sits at the intersection of model risk management and operational resilience, since a fraud model failure or a successful adversarial attack against it is both a model risk event and an operational incident. Supervisors expect firms to treat fraud AI with the same validation rigour as credit or capital models, even though fraud detection is sometimes carved out of consumer-facing explainability rules given the risk that transparency aids fraudsters.

Managing the false-positive trade-off

Setting the operating threshold for a fraud model requires an explicit cost function weighing fraud losses against customer friction and attrition, reviewed periodically as both fraud rates and customer expectations shift. Firms increasingly use tiered response strategies, step-up authentication for medium-risk transactions rather than outright blocking, to reduce the binary trade-off and preserve legitimate transaction volume while still interrupting likely fraud before completion.

Monitoring false-positive and false-negative rates by customer segment matters for fairness as well as commercial performance, since blanket thresholds can disproportionately inconvenience certain demographics or geographies with transaction patterns that superficially resemble fraud. Regular recalibration against labelled outcomes, confirmed fraud versus confirmed legitimate disputes, keeps the model's precision from drifting as fraud typologies and customer behaviour change over time.

Adversarial risk and governance controls

Fraud detection models face active adversaries who probe for blind spots and adapt, which is a materially different threat model from credit or pricing applications where the population being modelled is not deliberately trying to evade the model. This motivates continuous adversarial testing, red-teaming against evolving fraud tactics, and faster model refresh cycles than typically applied to other risk models, since a model that works well today can degrade quickly once fraudsters learn to exploit it.

Governance should include clear escalation paths between fraud operations, model risk, and financial crime compliance, since a fraud model intersects anti-money-laundering obligations and suspicious activity reporting requirements as well as customer protection rules. Firms need documented incident response plans for cases where the model is suspected of being actively gamed, including fallback to manual review processes while remediation is tested and deployed.

Conclusion

AI fraud detection delivers genuine risk reduction, but only when paired with governance that recognises its adversarial operating environment and the real trade-off between fraud losses and customer experience. Firms that validate, monitor, and refresh these models with the urgency the threat landscape demands outperform those treating fraud AI as a set-and-forget utility.

Jonas Adam Mohamed Osman (Yonas Osman), CEO of Quantica Risk, approaches fraud AI governance by pairing technical monitoring with explicit business trade-off decisions, ensuring threshold choices are owned by accountable stakeholders rather than left implicit in model defaults.

Primary sources

fraud detectionAIbankingoperational risk