AI adoption in Saudi Arabia is linked to shrinking national employment in routine-heavy sectors and rising national presence in high-skill digital roles; the sectoral pattern suggests a technology-driven segmentation of the labour market under Vision 2030.

Main Finding

• AI adoption in Saudi Arabia is structurally associated with a decline in the share of Saudi nationals in routine‑intensive sectors (fixed‑effects β = −0.41, p < 0.05) and with an increase in employment associated with high‑skill digital competencies (β = +0.67, p < 0.01). The paper interprets these as sectoral displacement in routine work and conditional inclusion/upskilling in high‑skill digital sectors within a state‑led, rentier context.

Key Points

• Scope: Balanced panel of 120 sector‑year observations (10 sectors, 2013–2024): Finance, Health, Education, Retail, Logistics, Public Administration, Construction, Manufacturing, Energy, Telecom.
• Core result: Higher AI adoption correlates with lower national employment shares in routine‑intensive sectors and higher shares where high digital skills are required.
• AI Adoption Index: constructed from three indicators — AI job postings (%), AI investment, and automation level.
  • Internal consistency: Cronbach’s α = 0.87.
  • PCA: PC1 explains 68% of variance; PCA‑weighted index highly correlated with equal‑weighted index (r = 0.94, p < 0.001).
• Robustness & diagnostics:
  • Fixed‑effects preferred (Hausman χ² = 14.73, p = 0.008).
  • Heteroskedasticity present (Modified Wald χ² = 28.4, p < 0.001); cluster‑robust SEs used.
  • No strong first‑order autocorrelation (Wooldridge F = 2.1, p = 0.087).
  • VIF max = 2.8 (no severe multicollinearity).
  • Panel unit‑root tests (LLC, IPS) indicate stationarity (p < 0.01).
• Endogeneity strategy: 2SLS using one‑period lagged AI Adoption Index as instrument.
  • First‑stage F = 18.4 (relevance satisfied).
  • Sargan test p = 0.34 (overidentification not rejected).
  • Authors acknowledge associations (not strict causal claims).
• Supplementary descriptive analysis: K‑means clustering (k = 3) on AI adoption × skill level:
  • Cluster 1 (High AI, Low Nationals): Finance, Logistics, Manufacturing.
  • Cluster 2 (Low AI, High Nationals): Public Administration, Education, Retail.
  • Cluster 3 (High AI, High Nationals): Telecom, Energy, Health.
• Theoretical framing: Uses structural labor process / macro‑sociological lens to interpret sectoral labor segmentation without making micro‑level claims about worker subjectivity.

Data & Methods

• Data sources (public, sector‑level, 2013–2024):
  • AI job postings: LinkedIn Workforce Report (sector aggregates).
  • Automation level: World Bank Digital Adoption Index.
  • AI investment: MoHRE annual reports.
  • National employment share: GASTAT Labor Force Survey.
  • Skill level: UNESCO/World Bank (ISCO‑08 based), normalized 0–1.
  • Controls: skill level, sector GDP growth, productivity index, foreign labour quota (Nitaqat), Saudization rate.
• Index construction:
  • Min–max normalization to [0,1]; equal‑weighted average as baseline; PCA sensitivity check.
• Econometric approach:
  • Baseline: sector fixed‑effects panel regression with cluster‑robust SEs.
  • Endogeneity: 2SLS IV using lagged AI adoption; first‑stage and Sargan reported.
  • Diagnostics reported (Hausman, Modified Wald, Wooldridge, VIF, LLC/IPS).
• Software and reproducibility:
  • Stata 18 (panel regressions, IV), SPSS v28 (PCA, Cronbach’s α), R 4.3.1 (clustering); dataset and syntax provided as supplementary materials.

Implications for AI Economics

• For theory:
  • Reinforces the view that AI adoption drives qualitative restructuring of labor (task fragmentation, polarization), but in a rentier/state‑led context this operates through state industrial strategy (Vision 2030) rather than purely market forces.
  • Illustrates a dual trajectory: technology can produce displacement in routine roles while simultaneously creating opportunities for nationally targeted upskilling in high‑skill digital domains.
• For empirical work:
  • Demonstrates a transparent workflow for composite AI adoption indices (validate with Cronbach’s α and PCA rather than relying on ad hoc weights).
  • Shows feasibility and limits of sector‑level panel analysis in small panels (parsimony recommended; extensive diagnostics necessary).
  • Notes the pragmatic use of lagged adoption as an IV — useful but requires careful scrutiny of the exclusion restriction in future work.
• For policy in rentier economies:
  • State‑directed AI rollout accelerates structural change; without complementary, sector‑specific upskilling and inclusive governance, national labour shares in routine sectors are likely to decline.
  • Policy levers include targeted retraining for displaced workers, aligning Saudization incentives with digital skill pathways, and regulation of AI hiring/adoption to avoid reinforcing segmented labor outcomes.
• Limitations and cautions (relevant to decision‑makers and researchers):
  • Results are associative, not definitive causal proof — IV strategy helps but does not fully eliminate concerns about omitted time‑varying confounders.
  • Aggregated sector‑level outcome (national employment share) cannot capture worker‑level heterogeneity (wages, hours, transitions).
  • Measurement issues: LinkedIn job postings may underrepresent some sectors/work types; investment figures and automation proxies carry reporting biases.
  • Small panel (10 sectors × 12 years) constrains power and the complexity of models that can be credibly estimated.
• Suggestions for next steps in AI economics research:
  • Use firm‑ or worker‑level microdata to estimate wages, transitions, and heterogeneity in displacement/upskilling.
  • Combine quantitative panel evidence with qualitative studies in affected sectors to document lived experiences and institutional mechanisms.
  • Explore policy experiments (training programs, hiring mandates) with quasi‑experimental evaluation to move from association to causal evidence.