AI Fraud Detection Under Institutional Constraints: An Integrative Review

This review integrates two bodies of literature that seldom cite each other: the technical literature on machine learning fraud detection, and the organizational literature on the institutional pressures that shape how those systems are operated commercially. Reading the technical evidence through the lens of coercive isomorphism (DiMaggio & Powell, 1983), the review examines four themes: the labor constraints that bound alert volume; the precision–recall trade-off in imbalanced fraud data; cost-sensitive learning as an economic, yet partial, remedy; and the ratio thresholds that payment networks enforce through monitoring programs such as Visa's Acquirer Monitoring Program (VAMP). The synthesis shows that each strand counts or prices classification errors while omitting the institutional penalty term: the loss of processing privileges that follows when dispute ratios breach network thresholds. Cost-sensitive models minimize immediate dollar losses but not termination risk; industry cost studies scale the price of fraud without modeling network sanctions; and behavioral evidence on information security compliance shows that permanent monitoring produces a standing tension between the cost of complying and the risk of being measured. The review identifies a practical knowledge gap—the absence of empirical work connecting a threshold decision inside a model to a dispute ratio, and a dispute ratio to a firm's institutional standing—and sets out the agenda such work must follow.