Jakarta MSMEs use AI but don’t see performance gains: intensity of AI use shows no significant effect on business outcomes, while firm size strongly predicts performance; policymakers should pair technology access with training and complementary capabilities rather than only subsidizing licenses.

Main Finding

AI usage intensity among Jakarta MSMEs is not significantly associated with better business performance (β = 0.085, p = 0.113). Firm size (small and medium vs. micro) has a large, significant direct positive effect on performance (small: β = 0.446, p < 0.001; medium: β = 0.548, p < 0.001). Although small firms tended to report higher AI usage intensity (β = 0.269, p < 0.001), AI usage did not mediate the effect of firm size on performance (indirect effects non‑significant). The authors label this an “AI adoption paradox”: adoption/intensive use alone does not automatically produce measurable business value in a resource‑constrained environment.

Key Points

• Sample/context: 300 MSME owners/managers in Jakarta who had used ≥1 AI tool in the prior 3 months; stratified by region and sector (trade, food, creative services).
• Main outcome: self‑reported business performance (turnover, profit, customer count vs. prior year).
• Main explanatory variable: AI usage intensity (frequency, duration, variety; 3 items, α = 0.81).
• Model quality: excellent fit (χ2/df = 1.13; CFI = 0.993; TLI = 0.996; RMSEA = 0.021; SRMR = 0.018).
• Explained variance: R2(AI usage) = 0.074; R2(Business performance) = 0.469 (firm size + AI usage).
• Mediation/bootstrapping: bias‑corrected bootstrap (5,000 reps) — indirect paths Small→AI→Performance and Medium→AI→Performance not significant.
• Measurement caveats: cross‑sectional design, self‑reported performance, aggregate AI usage construct (mixes generative, predictive, automation tools), potential omitted complementary assets (data governance, skills, infrastructure).
• Practical takeaways from authors: access/subsidized licenses alone are insufficient; complementary resources (dynamic capabilities, data governance, HR skills, infrastructure) are required to convert AI use into performance gains.

Data & Methods

• Design: cross‑sectional survey (June–July 2025), N = 300 valid responses; stratified proportionate sampling across 5 Jakarta municipalities and sectors.
• Constructs:
  • AI Usage Intensity: 3 items, 5‑point Likert, Cronbach’s α = 0.81.
  • Business Performance: 3 items (turnover, profit, customers), 5‑point Likert, Cronbach’s α = 0.85.
  • Firm size categories: micro (≤4 employees), small (5–19), medium (20–99) per Indonesian law.
  • Controls: business age, sector.
• Estimation:
  • Covariance‑Based SEM (CB‑SEM) with Maximum Likelihood estimation in JASP 0.18.
  • Missing data: Full Information Maximum Likelihood (FIML).
  • Fit indices and thresholds reported; measurement model tested for convergent validity (AVE > 0.50) and reliability.
  • Mediation: bootstrap (5,000) bias‑corrected CIs.
  • Post‑hoc power: >0.80 for effects ≈ β = 0.15.
• Limitations noted by authors: cross‑sectional mediation bias, self‑report measures, aggregation of heterogeneous AI tool types.

Implications for AI Economics

• Heterogeneous returns to AI: This paper provides empirical evidence that intensive AI use does not uniformly raise firm performance in a developing‑country MSME population. Productivity or revenue gains from AI are conditional on complementary assets; simple adoption metrics will overstate welfare or productivity gains unless context/capabilities are measured.
• Role of firm size and absorptive capacity: Firm size explains a large share of performance variation and is associated with capacity to realize technology benefits. Models of AI diffusion and productivity should include firm‑level complementarities (data pipelines, skilled labor, governance, finance) and not treat AI as a standalone capital input.
• Policy design implications: Subsidizing access or licenses is unlikely to be sufficient. Effective policy mixes should pair technology access with (a) managerial/digital training, (b) infrastructure financing (connectivity, cloud), (c) data governance/analytics assistance, and (d) incentives or support to build dynamic capabilities. Cost–benefit analyses of AI promotion programs must account for these additional investments.
• Measurement and evaluation guidance for researchers and practitioners:
  • Disaggregate AI by function (generative, predictive, automation) — aggregated usage can mask opposing effects.
  • Prefer panel or quasi‑experimental designs (or RCTs) to identify lagged impacts and causal channels (cross‑sectional mediation is fragile).
  • Use objective performance indicators (accounting revenue, productivity metrics) where possible to avoid self‑report bias.
  • Incorporate measures of absorptive capacity, data quality, and organizational processes to model complementarities.
• Macroeconomic/productivity accounting: Estimates of AI’s contribution to aggregate productivity in emerging economies may be upwardly biased if based solely on adoption counts. Heterogeneity across firm size and capability suggests slower, uneven productivity diffusion.
• Research priorities: causal identification of complementary interventions (e.g., training + AI vs. AI only), disaggregation by AI tool type and sector, and exploration of external support mechanisms (mentoring, subsidized infrastructure) that can unlock value in micro and small firms.

Suggested next steps for scholars and policymakers: run longitudinal or experimental evaluations of bundled interventions (AI tool + capability building), collect objective firm performance data, and develop targeted policy packages differentiating between AI tool types and firm size/sector.