Abstract
<title>Abstract</title> <p> <bold>Purpose</bold> : This study aims to investigate the underlying mechanisms through which work stress influences employee work engagement in the era of conversational artificial intelligence (AI). Specifically, it unravels the cognitive “black box” by examining the dual mediating roles of cognitive vigilance and cognitive offloading, driven by a multidimensional conceptualization of AI dependence (instrumental and emotional). Furthermore, it assesses the moderating effect of perceived AI hallucinations on these cognitive responses. <bold>Design/methodology/approach</bold> : Adopting a cross-sectional quantitative design, data were collected from 236 employees working in multinational corporations (MNCs) in Vietnam. To capture causal complexity, a pioneering multi-methodological approach was employed. This integration included Partial Least Squares Structural Equation Modeling (PLS-SEM) to estimate linear relationships, Artificial Neural Networks (ANN) for non-linear predictive modeling, fuzzy-set Qualitative Comparative Analysis (fsQCA) to uncover asymmetric configurations (equifinality), and Necessary Condition Analysis (NCA) to identify indispensable bottleneck factors. <bold>Findings</bold> : The results indicate that work stress strongly fuels both instrumental and emotional AI dependence. Notably, both forms of dependence significantly increase cognitive offloading without undermining cognitive vigilance. In terms of moderation, a profound "automation bias" was uncovered: the awareness of AI hallucinations neither diminishes cognitive offloading nor heightens cognitive vigilance. Crucially, the triangulation of the four analytical methods demonstrates that cognitive offloading is not only the most dominant predictor but also the singular necessary condition for achieving high work engagement. <bold>Originality/value</bold> : This paper makes three major theoretical contributions. First, it reconceptualizes conversational AI dependence as a multidimensional construct encompassing emotional attachment, moving beyond pure instrumental utility . Second, grounded in the Job Demands-Resources (JD-R) theory, it challenges traditional cognitive psychology paradigms by reframing cognitive offloading as a valuable "job resource" that optimizes employee capacity, rather than a manifestation of intellectual laziness . Third, it establishes a new methodological benchmark by combining symmetrical (PLS-SEM, ANN) and asymmetrical (fsQCA, NCA) approaches in human-AI interaction research. </p>