The AI–Well-being Paradox in the Workplace: A Dual-Pathway Model of Employee Well-being, Engagement and Sustainable Organizational Performance

Abstract

Artificial intelligence (AI) is changing how employees perform tasks, make decisions, communicate, learn and demonstrate performance. The central problem is that AI can operate simultaneously as a job resource and a job demand: it may improve autonomy, task assistance, learning, information access and creativity while also increasing technostress, insecurity, cognitive overload, monitoring pressure and digital spillover. This article develops and empirically tests the AI–Well-being Paradox, linking AI-enabled job resources and AI-induced job demands to employee well-being, employee engagement and sustainable organizational performance, with digital boundary control as a protective contextual resource. The study analyzes 600 employee records from the supplied dataset. Composite scores were calculated from seven multi-item constructs and analyzed using reliability and validity diagnostics, controlled regression with HC3 robust standard errors, 1,000-resample bootstrap mediation and sequential mediation, and interaction analysis. All seven constructs demonstrated satisfactory measurement quality (Cronbach’s α=.844–.875; composite reliability=.895–.909; AVE=.656–.700), and the maximum HTMT ratio was .727. AI resources positively predicted well-being (β=.441, p<.001), whereas AI demands negatively predicted well-being (β=−.348, p<.001). Well-being predicted engagement (β=.568, p<.001), and engagement predicted sustainable performance (β=.567, p<.001). The indirect and sequential pathways were significant. Digital boundary control significantly moderated the AI-demand pathway (β=.110, p=.002), but the resource × boundary-control interaction was not significant (β=.008, p=.846). Thus, 11 of the 12 hypotheses were supported. Importantly, although the dataset contains Wave1_Date, Wave2_Date and Wave3_Date fields, it does not contain distinct wave-specific item responses; therefore, the present statistical analysis is cross-sectional and does not claim a genuine longitudinal three-wave test.