Indonesian manufacturers that report adopting AI and automation display roughly 18–23% higher total factor productivity on average, with productivity gains concentrated among medium firms in tech-intensive sectors and larger incumbents in traditional industries; however, the association weakens in within-firm analyses and may partly reflect selection rather than purely causal effects.

Main Finding

AI/automation adoption is positively associated with higher firm-level productivity in Indonesian manufacturing during 2020–2021. AI-adopting firms show productivity premiums on the order of roughly 17–32% depending on specification and productivity measure (preferred industry+year specification ≈ +18%). AI adoption is also linked to stronger post‑COVID productivity recovery (greater resilience).

Key Points

• Data and coverage: Firm-level panel drawn from Statistics Indonesia’s Large and Medium Manufacturing Survey covering roughly 24k medium-to-large manufacturers matched over 2020–2021 (final analysis samples reported in the paper range ~24k–27k firms/observations).
• AI measure: binary indicator of firm-reported use of digitally enabled/algorithmic technologies (automation systems, embedded AI in production). Adoption prevalence is low (~7.6% of firms in the sample).
• Magnitude: baseline pooled specifications show large raw correlations (bivariate OLS coefficient 0.74), but with controls and industry & year fixed effects the estimated TFP premium is stable around 0.18–0.23 log points (≈18–26%). The paper reports an overall range of 17–32% across specifications and alternative productivity measures.
• Heterogeneity:
  • Gains decline with higher labor and capital intensity → diminishing marginal returns for larger, input‑intensive firms.
  • Sectoral pattern: in technologically intensive industries medium-size firms capture larger AI productivity premiums; in traditional sectors productivity gains tend to concentrate among larger incumbents.
• Resilience: AI adopters experienced stronger productivity recovery in the post-pandemic period, consistent with AI/automation acting as a buffer against COVID-19 disruptions.
• Robustness checks: results are robust to alternative productivity measures (LP-based TFP, OLS residual TFP, labor productivity), multicollinearity diagnostics (mean VIF ≈ 1.17), and quasi-experimental adjustments (Propensity Score Matching and IPWRA). A firm-fixed‑effects specification yields an insignificant coefficient, attributed to the short panel and low within-firm changes in AI status.

Data & Methods

• Data source: Indonesia Large and Medium Manufacturing Survey (BPS), matched firm panel 2020–2021.
• Sample construction: cleaned and matched firms with full labor and capital information; final panel ~24,033 firms (two-year observations); alternative reported observation counts appear in tables (~26–48k depending on variable and aggregation).
• Outcome measures:
  • Primary: Total Factor Productivity (TFP) estimated via Levinsohn–Petrin (LP) semi-parametric method using intermediate inputs (fuel/electricity) to control for simultaneity.
  • Alternatives: TFP from OLS residuals, labor productivity.
• Identification / estimation:
  • Baseline regressions: OLS with controls, industry and year fixed effects (preferred cross-sectional identification leverages within-industry cross-sectional variation).
  • Robustness to selection: Propensity Score Matching (Probit-based propensity) and Inverse Probability Weighted Regression Adjustment (IPWRA, doubly robust).
  • Additional diagnostics: VIFs for multicollinearity; attempted firm fixed effects (limited by short panel and persistent treatment).
• Limitations acknowledged by authors: self-reported binary AI measure (no fine-grained app/type information), short panel (2020–2021), lack of plausibly exogenous instrument for adoption → cannot claim full causal identification.

Implications for AI Economics

• Evidence from an emerging market: Adds micro-level evidence that AI/automation can raise productivity in a developing-economy manufacturing context, albeit gains vary by firm size, input intensity and sector.
• Complementarity with human capital and absorptive capacity: Diminishing returns in more input‑intensive firms and larger premiums in tech‑intensive sectors for medium firms suggest returns depend on firm characteristics and capabilities—policy should pair AI diffusion with skills and managerial adoption support.
• Distributional implications: In traditional sectors AI appears to reinforce incumbent large-firm advantages, whereas in tech-intensive sectors AI can be an equalizer for medium firms. Policymakers should be mindful of possible concentration effects and design measures (e.g., support for SME adoption, training, shared digital infrastructure) to broaden benefits.
• Resilience and shock-absorption: AI/automation can act as a resilience mechanism (faster recovery post-COVID), implying public support for digital adoption may also buttress economic stability in crises.
• Research gaps and cautions: Results are associative rather than strictly causal given data constraints (self-reporting, short panel, no instrument). Future work should use richer adoption measures (type/intensity of AI), longer panels, and exogenous variation (policy rollouts, instrumented adoption) to pin down causal channels and generalize across other emerging economies.