Estimates indicate that approximately 3 million private-sector workers in Türkiye occupy roles with a high probability of labor displacement due to artificial intelligence. Another 7 million face substantial reorganization: their employment is unlikely to disappear, but AI will fundamentally alter their task composition and production functions. Conversely, roughly 1 million workers in professional and technical occupations may experience net employment growth, as AI-induced productivity gains lower marginal costs and stimulate elastic demand. For the remaining 16 million workers, near-term structural shifts appear statistically modest.
These findings emerge from a novel application of the Richmond (2026) AI Jobs Transition Framework to Türkiye's 2024 Household Labour Force Survey (HLFS), evaluating 27.7 million private-sector workers across 39 occupational classifications. The full working paper is forthcoming; this summary highlights core microeconomic findings.
Analytical Framework
The classification of each occupational group relies on two primary dimensions of analysis.
AI Exposure quantifies the proportion of occupational tasks susceptible to substitution or automation by AI technologies. This relies on the Complementary AI Occupational Exposure index (C-AIOE), developed by Pizzinelli et al. and empirically extended to the Turkish labor market by Aşık et al. (2026). Currently, approximately 41% of Turkish private-sector employment is concentrated in high-exposure occupations.
Human Necessity (H) captures the degree to which an occupation requires physical presence, in-person caretaking, or licensed/regulatory accountability. These factors act as significant frictions against capital-labor substitution. For instance, long-haul truck drivers, nurses, and licensed pharmacists exhibit high H scores, whereas data entry clerks and call centre agents score low.
These dual metrics, interacted with the price elasticity of demand (|η|) for the augmented service—determining whether productivity-driven price declines generate sufficient demand expansion to offset task displacement—sort workers into four distinct archetypal labor market trajectories.
Figure 1. Classification logic. Each occupational group is assigned sequentially based on microeconomic thresholds: low AI task exposure → Less Change; high exposure + elastic demand → Grow with AI; high exposure + inelastic demand + high human necessity → Reorganize; high exposure + inelastic demand + low human necessity → Automation Risk. Employment shares rely on NACE-weighted private sector data from the 2024 HLFS.
Archetypal Labor Market Trajectories
Comprising agricultural workers, construction tradespeople, drivers, and personal service workers. Defined by low task substitution potential or prohibitively high frictions (human necessity). Structural disruption is minimized in the short-run for these 16.2 million workers.
Dominated by sales workers (~3 million), plant operators, and food preparation personnel. While AI introduces positive productivity shocks, output demand is inherently constrained by physical capacity or human necessity. The marginal product of labor shifts, altering tasks rather than employment volume.
Routine clerical occupations: office clerks, numerical clerks, and customer service staff. Characterized by high task substitutability, minimal physical frictions, and inelastic product demand. Productivity gains are highly unlikely to stimulate sufficient demand to maintain current employment levels for these 3.1 million workers.
ICT professionals, business administrators, and specialized craft workers. AI operates as a highly complementary input. Output demand is sufficiently elastic that AI-induced cost reductions are projected to spur net labor demand. Structurally, this cohort is notably smaller in Türkiye compared to advanced economies.
Empirical Measurement and Model Validation
Figure 2. Left — Bivariate relationship between AI exposure (C-AIOE) and rule-based human necessity (H) across 39 occupational groups; node size proportional to HLFS 2024 employment share; Spearman ρ = −0.61. Centre — Correlation between LLM-elicited vs. O*NET rule-based human necessity proxies (ρ = 0.80). Right — Robustness of demand elasticity estimates across LLM architectures (Claude Sonnet 4.6 vs. GPT-5.4 mini, ρ = 0.72). ISCO 95 and 96 excluded from panels 1–2 due to O*NET crosswalk limitations.
The left panel demonstrates a statistically significant negative correlation between AI exposure and human necessity (ρ = −0.61). This aligns with theoretical expectations regarding capital-labor substitution: occupations with fewer physical or regulatory frictions are inherently more susceptible to automation.
The centre and right panels address methodological robustness regarding the elicitation of variables via Large Language Models. Comparing LLM-derived indices of human necessity against deterministic, rule-based mappings from the U.S. Department of Labor's O*NET database yields a strong correlation (ρ = 0.80). Furthermore, elasticity parameter estimates across heterogeneous LLM architectures remain stable (ρ = 0.72), with variance largely confined to margin-of-error bounds near the unitary elasticity threshold.
Gender Heterogeneity in Displacement Risk
Aggregate macroeconomic metrics mask severe occupational sorting along gender lines.
Among formal-sector workers, 61.9% of female employment is concentrated in occupations classified under Automation Risk or Reorganize, compared to just 38.7% of male employment.
This variance is primarily driven by the disproportionate allocation of formal female labor into routine cognitive tasks—precisely those categorized as Automation Risk (e.g., clerical administration)—as well as their overrepresentation in sales and service occupations facing high reorganization pressure. Conversely, informal labor markets exhibit lower aggregate exposure due to a higher concentration in manual tasks that possess intrinsic human necessity frictions. Consequently, the labor cohort most structurally vulnerable to AI-induced displacement in Türkiye consists of formally employed women.
Cross-Country Benchmarking: Türkiye vs. United States
Contrasting these findings against Richmond's (2026) benchmark estimates for the United States highlights critical differences in structural labor market composition.
| Archetype | Türkiye | United States |
|---|---|---|
| Less Change | 58.6% | 46.0% |
| Reorganize | 26.6% | 24.0% |
| Automation Risk | 11.2% | 18.0% |
| Grow with AI | 3.5% | 12.0% |
Türkiye’s lower Automation Risk share (11.2% vs. 18.0%) is largely a function of differing relative factor endowments—specifically, a lower per capita density of white-collar cognitive occupations. The elevated Reorganize share (26.6% vs. 24.0%) reflects Türkiye's industrial reliance on retail trade and manufacturing assembly. Most critically, the gap in the Grow with AI category—3.5% in Türkiye compared to 12.0% in the U.S.—indicates a structural deficit in complementary, high-skill professional employment. This poses significant long-run macroeconomic challenges regarding skill-biased technological change and technology adoption.
Methodological Limitations
This framework operates largely in partial equilibrium. Demand elasticities rely on occupational profiles synthesized as of 2024 and do not account for dynamic cross-elasticities or general equilibrium effects. Sensitivity analyses reveal significant variance depending on sectoral aggregation. For example, aggregating Türkiye's ~3 million sales workers purely by aggregate occupational elasticity initially sorted them into Grow with AI. However, adjusting for the fact that 71.4% operate in retail—a sector where cost efficiencies exhibit highly inelastic consumer demand—reallocated the entire cohort to Reorganize. Depending on assumptions regarding sector-specific demand curves, the aggregate Grow with AI share fluctuates between 0.6% and 31.5%.
The forthcoming working paper expands on these partial equilibrium limitations, providing regional disaggregations, extensive robustness checks against O*NET variables, and active labor market policy (ALMP) recommendations.
Comments
Post a Comment
Constructive feedback is always welcome. Thank you