AI adoption has a double-edged effect on employees: confidence and engagement raise self-reported performance while job insecurity and AI-induced stress depress it; both positive and negative mediation pathways are statistically significant in analysis of 280 Indonesian workers.

Main Finding

AI adoption in Indonesian workplaces operates through two simultaneous psychological pathways. A positive (gain-oriented) route—AI adoption → technological self-efficacy → job engagement → higher employee performance—is statistically significant. At the same time, a negative (loss-oriented) route—AI adoption → perceived job insecurity → AI‑induced stress → lower employee performance—is also significant. Thus AI both raises performance (via empowerment and engagement) and reduces it (via insecurity and stress); both effects coexist in the same workforce.

Key Points

• The paper proposes and tests a dual-process framework integrating four theories: Job Demands–Resources (JD–R), Social Cognitive Theory (SCT), Conservation of Resources (COR), and technostress theory.
• Hypothesized gain pathway:
  • H1–H5: AI adoption increases technological self-efficacy and job engagement; self-efficacy and engagement positively affect performance.
  • H6: AI → self-efficacy → engagement → performance is a sequential mediation.
• Hypothesized loss pathway:
  • H7–H9: AI adoption increases perceived job insecurity and AI-induced stress; insecurity elevates stress, which reduces performance.
  • Sequential mediation through insecurity and stress is supported.
• Empirical result: both sequential mediations (gain and loss) are statistically significant — confirming the coexistence of opportunity and risk in AI-driven transformations.
• Normative emphasis: organisational investments must both reduce anxiety/strain and build digital skills to capture AI’s benefits sustainably.

Data & Methods

• Sample: 280 Indonesian employees across multiple business sectors (diverse industries; Indonesia-specific context highlighted).
• Design: Cross-sectional survey (self-reported measures).
• Analysis: Structural equation modeling using SmartPLS (PLS-SEM) to test direct effects and serial mediation paths.
• Main constructs measured: AI adoption (employee perceptions/use), technological self-efficacy, job engagement, perceived job insecurity, AI-induced stress (technostress), and employee performance.
• Limitations noted (implicit/typical for this design):
  • Cross-sectional data limits causal inference and dynamics over time.
  • Reliance on self-reported performance and perceptions may introduce common-method bias.
  • Sample size (n=280) adequate for PLS-SEM but limits fine-grained subgroup analysis; generalizability to other countries or informal-sector workers is limited.
  • Measurement of AI adoption is perceptual; objective usage/intensity or algorithmic characteristics were not detailed.

Implications for AI Economics

• Modeling productivity: Microfoundations of AI-driven productivity should include psychological channels. AI’s net effect on worker productivity is the result of offsetting gains (skill complementarities, increased engagement) and losses (stress, insecurity). Macroeconomic or firm-level productivity models that ignore stress/insecurity risks overestimating net gains.
• Labor-market impacts and welfare accounting:
  • Short-run transitions: Increased AI diffusion can produce transient productivity gains but also impose mental-health and job-insecurity costs that reduce effective labor input and welfare.
  • Distributional effects: Heterogeneous access to upskilling and institutional protections (urban/rural, formal/informal sectors) will amplify inequality; developing-country contexts like Indonesia face larger downside risks.
• Policy and firm strategy:
  • Active labor-market policies (retraining, subsidized reskilling) and firm-level investments in digital literacy increase the positive path and mitigate the negative path.
  • Organizational practices to reduce technostress: algorithmic transparency, participatory implementation, workload redesign, psychological support, and clear career pathways reduce insecurity and stress.
  • Regulatory design: Balance incentives for adoption with worker protections (retraining mandates, disclosure requirements for algorithmic evaluation, monitoring of wellbeing metrics).
• Empirical research recommendations for AI economics:
  • Use longitudinal/panel data and objective performance metrics to identify dynamic causal effects and recovery or persistence of stress impacts.
  • Quantify transition costs (productivity loss from stress, healthcare, turnover) to include in cost–benefit analyses of AI diffusion.
  • Model heterogeneity explicitly (skill levels, sectors, formal vs informal employment, regional infrastructure) to predict where net gains are likely vs where safeguards are essential.
  • Incorporate psychological frictions (self-efficacy, perceived insecurity, technostress) into structural labor-market models to capture complementarities and adjustment frictions.
• Measurement implications:
  • Collect both perceptual and objective indicators of AI intensity, algorithmic opacity, and performance evaluation practices.
  • Include wellbeing and stress metrics in firm productivity datasets; these are necessary to estimate the full social returns and costs of AI adoption.

Overall, the paper argues that economists and policymakers should treat AI diffusion not only as a technological shock but as a socio-psychological process whose net effect on productivity and welfare depends on complementary investments in human capital, institutional safeguards, and organizational design.