Understanding continuance intention of artificial intelligence (AI)-enabled mobile banking applications: an extension of AI characteristics to an expectation confirmation model

Mobile banking, as a key FinTech platform, offers convenience to users and cost efficiencies to banks. Recent advances have embedded AI into mobile banking apps, transforming them into intelligent systems that can personalize services and enhance user experience. While prior research on continuance often relies on TAM or UTAUT, the expectation confirmation model (ECM) better explains post-adoption continuance via confirmation, perceived usefulness, and satisfaction. The study addresses the research question: Do perceived intelligence and perceived anthropomorphism of AI-enabled mobile banking apps affect users' continuance intention through the mechanisms specified by ECM, and if so, how? The purpose is to integrate AI characteristics (perceived intelligence, perceived anthropomorphism) into ECM and empirically test their effects on confirmation, perceived usefulness, satisfaction, and continuance intention.

ECM, derived from expectation confirmation theory, posits that confirmation of expectations influences perceived usefulness and satisfaction, which in turn shape continuance intention. In mobile banking, ECM has been validated as a robust framework for explaining continuance. The distinctive features of AI systems—perceived intelligence (capability to understand, adapt, and assist) and perceived anthropomorphism (human-like interactions)—are salient in AI-enabled mobile banking apps and can shape user evaluations. The authors extend ECM by introducing perceived intelligence and perceived anthropomorphism as antecedents to confirmation and perceived usefulness, and by modeling intelligence as a driver of anthropomorphism. Prior studies show intelligence may enhance anthropomorphic perceptions and both can impact perceived performance and emotions. The paper formulates ten hypotheses: H1: Confirmation → Perceived usefulness (positive). H2: Confirmation → Satisfaction (positive). H3: Perceived usefulness → Satisfaction (positive). H4: Perceived usefulness → Continuance intention (positive). H5: Satisfaction → Continuance intention (positive). H6: Perceived intelligence → Perceived anthropomorphism (positive). H7: Perceived intelligence → Confirmation (positive). H8: Perceived intelligence → Perceived usefulness (positive). H9: Perceived anthropomorphism → Confirmation (positive). H10: Perceived anthropomorphism → Perceived usefulness (positive). A literature synthesis (e.g., Yuan et al., 2016; Susanto et al., 2016; Kumar et al., 2018; Hidayat-ur-Rehman et al., 2021; Sinha and Singh, 2022) indicates confirmation, perceived usefulness, trust, and quality-related factors are key drivers of continuance in mobile banking, supporting the integration of AI features into ECM.

Design: Cross-sectional survey targeting Chinese users with experience using AI-enabled mobile banking apps. Sampling and data collection: Random sampling via a professional online survey firm (Sojump). Screening ensured only respondents with AI-enabled mobile banking experience proceeded. Of 400 invitations, 370 responses were received; 5 incomplete were removed, yielding 365 valid responses (response rate 91.25%), exceeding the minimum sample size of 119 estimated using G*Power (effect size 0.15, power 0.95, alpha 0.05). Instrument development: Items adapted from validated scales; English-to-Chinese back-translation applied. Expert review by three domain experts ensured face and content validity; pilot test with 30 users informed revisions. Measurement: 7-point Likert scales. Perceived intelligence and perceived anthropomorphism measured with 5 items each (Lee & Chen, 2022b). Perceived usefulness with 4 items (Venkatesh & Davis, 2000; Bhattacherjee, 2001). Confirmation with 4 items (Bhattacherjee, 2001). Satisfaction with 4 items (Bhattacherjee, 2001). Continuance intention with 2 items (Bhattacherjee, 2001; Yuan et al., 2016). Common method bias controls: Procedural remedies (varied scale anchors, concealed construct names, randomized items) and post hoc tests: Harman’s single-factor test (first factor 35.7% < 50% threshold), full collinearity VIFs < 3.3 for all constructs, and a marker variable (average monthly income) uncorrelated with model variables, indicating CMB not a major concern. Analysis: Partial least squares structural equation modeling (PLS-SEM) using SmartPLS 3. Two-stage evaluation: measurement model and structural model. Reliability and validity: Composite reliability 0.858–0.914 and Cronbach’s alpha 0.881–0.922 exceeded 0.7; AVE 0.588–0.751 exceeded 0.5; item loadings > 0.7 and significant (p < 0.001). Discriminant validity supported by HTMT with all values below 0.85. Structural model: Bootstrapping with 10,000 resamples; assessed path coefficients, f2 effect sizes, and R2 values. Predictive assessment: PLSpredict (10 folds, 1 repeat) indicated positive Q2 values for continuance intention indicators and better prediction (lower RMSE/MAE) than a linear model. Mediation analysis: Zhao et al. (2010) approach used to test indirect effects among AI features and ECM constructs.

All ten hypotheses (H1–H10) were supported. Key path coefficients (all p < 0.01): H1 CONF → PU: β = 0.367 (T = 4.559, f2 = 0.225, medium). H2 CONF → SAT: β = 0.466 (T = 7.891, f2 = 0.337, medium). H3 PU → SAT: β = 0.492 (T = 8.963, f2 = 0.355, large). H4 PU → CI: β = 0.411 (T = 5.855, f2 = 0.267, medium). H5 SAT → CI: β = 0.388 (T = 6.121, f2 = 0.272, medium). H6 PI → PA: β = 0.555 (T = 12.548, f2 = 0.371, large). H7 PI → CONF: β = 0.399 (T = 6.897, f2 = 0.291, medium). H8 PI → PU: β = 0.412 (T = 6.532, f2 = 0.285, medium). H9 PA → CONF: β = 0.487 (T = 7.331, f2 = 0.326, medium). H10 PA → PU: β = 0.198 (T = 2.671, f2 = 0.193, medium). Model explanatory power: R2 values—PA = 0.305; CONF = 0.530; PU = 0.571; SAT = 0.735; CI = 0.652—indicating substantial explained variance, especially for satisfaction and continuance intention. Predictive performance: PLSpredict indicated positive Q2 and lower RMSE/MAE than a linear benchmark, evidencing out-of-sample predictive power. Mediation: Complementary mediations were observed across theorized indirect paths, including PI → PA → CONF; PI/PA → CONF → PU/SAT; PI/PA → PU → SAT/CI, indicating that AI characteristics influence continuance intention both directly (via ECM antecedents) and indirectly through sequential mechanisms of confirmation, perceived usefulness, and satisfaction.

The study demonstrates that integrating AI characteristics into ECM clarifies how users form continuance intentions for AI-enabled mobile banking. Perceived intelligence enhances perceived anthropomorphism, and both AI attributes elevate confirmation of expectations and perceived usefulness. These, in turn, raise satisfaction, which, along with perceived usefulness, drives continuance intention. Thus, the findings answer the research question by showing that AI features shape continuance via ECM’s core mechanisms. The significant R2 values for satisfaction and continuance intention underscore the model’s explanatory strength. Practically, intelligent, anthropomorphic design features can create human-like, efficient interactions that align with or exceed expectations, building utility perceptions and satisfaction, thereby sustaining usage. The complementary mediation patterns suggest that improving AI capabilities can cascade through multiple ECM pathways, offering multiple levers for designers and managers.

By extending ECM with perceived intelligence and perceived anthropomorphism, the study advances understanding of continuance intention for AI-enabled mobile banking apps. Empirical evidence from 365 experienced users shows intelligence fosters anthropomorphism; both AI features boost confirmation and perceived usefulness, which increase satisfaction and ultimately continuance. The integrated model exhibits strong explanatory and predictive power. Contributions include illuminating mechanisms by which AI features affect post-adoption evaluations and behavior and enhancing ECM’s applicability in AI-rich financial services. Future research should replicate across countries, incorporate potential moderators (e.g., personal innovativeness, trust), examine personality trait interactions, adopt longitudinal designs to assess causality and dynamics, and integrate AI service quality frameworks (e.g., AISAQUAL) with ECM to further explain continuance.

• Sample restricted to China limits generalizability; cross-cultural replication is needed. • Modest sample size (n = 365) may not fully represent the population. • Cross-sectional design precludes strong causal inference; longitudinal studies are recommended. • No moderators were modeled; potential moderators (e.g., personal innovativeness, trust) may improve explanatory power. • Personality traits were not considered; trait–AI feature interactions may influence continuance. • Focus on AI features did not explicitly incorporate AI service quality dimensions (e.g., AISAQUAL), which may further shape continuance.