Abstract
Payment
banks process enormous volumes of small-value digital transactions, making them
attractive targets for increasingly sophisticated payment fraud. This study
investigates artificial intelligence (AI)-based fraud detection in payment
banks through a dual-method design. First, a machine learning benchmark
compares four supervised classifiers—logistic regression, decision tree, random
forest, and gradient boosting—on a class-imbalanced transaction dataset (N =
20,000; 3.5% fraud incidence), with the random forest ensemble achieving the
best performance (accuracy = 98.0%; AUC = 0.932), consistent with the strengths
of ensemble learning (Breiman, 2001; Chen & Guestrin, 2016) and
anomaly-detection principles (Chandola et al., 2009). Second, drawing on the
Technology Acceptance Model (Davis, 1989), UTAUT2 (Venkatesh et al., 2012),
expectation–confirmation theory (Bhattacherjee, 2001), and trust theory (Gefen
et al., 2003; McKnight et al., 2002), the study tests an integrated behavioural
model in which AI detection accuracy, real-time responsiveness, and system
transparency shape customers’ continuance usage intention through perceived
fraud-detection effectiveness and trust in the payment bank. Survey data from
428 payment bank customers were analysed using EFA, CFA, and PLS-SEM with 5,000
bootstrap subsamples. The measurement model demonstrated strong psychometric
properties (α = 0.819–0.869; CR = 0.867–0.904; AVE = 0.620–0.701; all HTMT <
0.85), and all nine hypothesised paths were supported. Trust in the payment
bank was the strongest driver of continuance intention (β = 0.419, p <
.001), and the model explained 47.0% of intention variance. The findings
demonstrate that the commercial value of AI fraud detection lies not only in
classifier performance but in its capacity to build customer confidence in
digital payment ecosystems.