The Commonplace
Home Dashboard Papers Evidence Syntheses Digests 🎲
← Papers

AI exposure reshapes firm skill portfolios: displacement cuts routine cognitive roles while augmentation raises demand for nonroutine analytical skills, and both produce shallower, more dispersed skill mixes concentrated in small, low-threshold firms; policy should favor broad, portable reskilling over deep single-track specialization.

Toward Sustainable Workforce Development: How AI Reshapes Skill Demand Structure—Evidence from 67 Million Job Postings in China
Lingzhe Zhang, Chenglei Zhang · Fetched May 23, 2026 · Sustainability
semantic_scholar quasi_experimental medium evidence 8/10 relevance DOI Source
Using 67 million Chinese job postings and patent-task matching, the paper finds that AI exposure linked to displacement reduces routine cognitive skill shares while AI exposure linked to augmentation raises nonroutine analytical skill demand, and both lead firms to 'de-core' skill portfolios — effects concentrated in small, low entry-threshold firms.

How artificial intelligence (AI) reshapes the internal structure of firm-level skill demand remains largely uncharted. Using approximately 67 million online job postings from two major Chinese recruitment platforms (2019–2024), we construct firm-by-year potential AI exposure via semantic matching between AI patent texts and detailed occupation task descriptions, decompose exposure into displacement and augmentation components based on task routineness, and measure four skill-category demand shares and their within-category importance from job-description text, with identification from within-firm variation under firm and city-by-year fixed effects. Displacement and augmentation exposure exhibit opposing relationships with skill demand: displacement is negatively associated with the routine cognitive share, while augmentation is positively associated with the nonroutine analytical share. Both forms of exposure are associated with a de-coring pattern, a shallower and more dispersed skill portfolio with within-category importance diverging from share movements, concentrated among low entry-threshold, small firms. Reskilling policy should therefore emphasize portfolio breadth and portable competency frameworks rather than deeper single-track specialization, particularly for workers in small, lower-threshold firms.

Summary

Main Finding

Firm-level potential AI exposure—measured through semantic overlap between AI patent texts and occupation task descriptions—alters how firms demand skills in opposing ways. AI-related displacement reduces routine cognitive skill shares, while AI-related augmentation increases nonroutine analytical skill shares. Both forces produce a "de-coring" of firms' skill portfolios (shallower, more dispersed skill requirements and divergence between skill-category shares and the importance of particular skills), an effect concentrated in small, low entry‑threshold firms. Reskilling policy should prioritize breadth and portability of competencies over deeper single-track specialization.

Key Points

  • Data scale: ~67 million online job postings from two major Chinese recruitment platforms, 2019–2024.
  • AI exposure measure: firm-by-year "potential AI exposure" constructed by semantic matching between AI patent texts and detailed occupation task descriptions.
  • Decomposition: exposure split into displacement vs. augmentation using task routineness (routine tasks → potential displacement; nonroutine analytical tasks → potential augmentation).
  • Skill measurement: four skill-category demand shares derived from job-description text plus within-category importance (which skills are emphasized within each category).
  • Identification: exploited within-firm variation with firm fixed effects and city-by-year fixed effects.
  • Empirical patterns:
    • Displacement exposure correlates with a decline in the routine cognitive skill share.
    • Augmentation exposure correlates with an increase in the nonroutine analytical skill share.
    • Both exposures are linked to "de-coring": skill portfolios become shallower and more dispersed, and the relative emphasis on particular skills within categories diverges from overall category share movements.
    • These portfolio changes are concentrated among small firms and firms operating in low entry-threshold occupations.

Data & Methods

  • Data: ~67M online job postings (2019–2024) from two large Chinese job platforms; firm-level aggregation by year.
  • AI exposure construction:
    • Textual semantic matching between AI patent documents and occupation-level task descriptions to estimate how much a firm's tasks overlap with AI capabilities.
    • Generate a firm-by-year exposure index; decompose it into displacement vs. augmentation components according to task routineness.
  • Skill demand measurement:
    • Extract skill mentions from job-description text and classify into four skill categories (including routine cognitive and nonroutine analytical).
    • Compute two margins: (a) category demand share (the share of postings or mentions in each broad category) and (b) within-category importance (the emphasis placed on specific skills inside each category).
  • Identification strategy:
    • Panel regression using within-firm variation, controlling for firm fixed effects and city-by-year fixed effects to net out time- and location-driven confounders.
    • Robustness checks and heterogeneity analyses focus on firm size and occupational entry thresholds (results show concentration of effects in small, low-threshold firms).

Implications for AI Economics

  • Structural labor demand: AI shifts the internal composition of skill demand within firms, not only aggregate occupational employment. Models of AI impact should incorporate within-firm task reallocation and changes in skill‑mix breadth.
  • Complementarity vs. substitution: Displacement and augmentation operate simultaneously and in opposite directions across task types—policy and firm strategy must account for both forces rather than assuming a single uniform effect.
  • Human capital policy design:
    • Prioritize broad, portable competency frameworks (skills that transfer across occupations and firms) and policies that support multi-skill portfolios.
    • Emphasize reskilling that increases breadth (cross-training, foundational analytical skills) rather than deeper single-track specialization, especially for workers in small, low entry‑threshold firms who face the largest portfolio destabilization.
  • Firm strategy and labor markets:
    • Small firms and low entry‑threshold occupations are most exposed to destabilization of skill cores; targeted support (training subsidies, portable credentialing) can mitigate adverse adjustment costs.
    • Hiring and HR practices may shift toward more dispersed, lower-depth skill requirements—implications for wages, career progression, and labor frictions that merit further empirical and theoretical work.

Assessment

Paper Typequasi_experimental Evidence Strengthmedium — Very large, detailed dataset (≈67 million job postings) and thoughtful measurement of AI exposure and task-level mechanisms provide credible descriptive and associational evidence; however, identification rests on observational within-firm variation and may be vulnerable to endogenous AI adoption, time-varying omitted factors, and measurement error from semantic matching, so causal claims are plausible but not ironclad. Methods Rigormedium — Methods combine state-of-the-art NLP (semantic matching) with a clear decomposition of mechanisms and panel fixed-effects, which is rigorous for observational work; nevertheless, potential issues remain (reverse causality, measurement error in mapping patents to tasks, selection into job-posting behavior, and lack of a quasi-random shock or instrument) that limit causal inference. SampleApproximately 67 million online job postings from two major Chinese recruitment platforms covering 2019–2024, linked to firm identifiers to create firm-by-year panels; AI exposure constructed from Chinese AI patent texts matched to detailed occupation task descriptions, with skill-category demand shares and within-category importance derived from job-description text. Themesskills_training human_ai_collab labor_markets adoption IdentificationConstruct firm-by-year AI exposure by semantically matching AI patent texts to occupation task descriptions in job postings, decompose exposure into displacement (routine tasks) and augmentation (nonroutine tasks) components, and identify effects from within-firm variation using firm fixed effects and city-by-year fixed effects to control for time-invariant firm attributes and locality-specific time trends. GeneralizabilitySingle-country (China) context — institutional, regulatory, and sectoral patterns may differ from other countries, Online job-posting data — may under-represent informal employment, internal hires, and occupations less likely to be posted online, Two recruitment platforms — platform user mix may bias sector, firm-size, or skill composition, Patent-to-task semantic matching — measurement errors could vary across industries and languages, Study period (2019–2024) — early-to-mid AI diffusion dynamics may not generalize to later stages

Claims (10)

ClaimDirectionConfidenceOutcomeDetails
The study uses approximately 67 million online job postings from two major Chinese recruitment platforms (2019–2024). Other null_result high dataset size and coverage (number of job postings, platforms, years)
n=67000000
0.8
The authors construct firm-by-year potential AI exposure via semantic matching between AI patent texts and detailed occupation task descriptions. Other null_result high firm-by-year potential AI exposure (constructed measure)
n=67000000
0.8
AI exposure is decomposed into displacement and augmentation components based on task routineness. Other null_result high decomposed AI exposure measures (displacement vs augmentation)
n=67000000
0.8
The study measures four skill-category demand shares and their within-category importance from job-description text. Other null_result high skill-category demand shares and within-category importance
n=67000000
0.8
Identification of effects uses within-firm variation with firm and city-by-year fixed effects. Other null_result high identification approach / econometric controls
n=67000000
0.8
Displacement exposure is negatively associated with the routine cognitive skill share. Skill Obsolescence negative high routine cognitive skill share (share of demand for routine cognitive tasks/skills)
n=67000000
0.48
Augmentation exposure is positively associated with the nonroutine analytical skill share. Skill Acquisition positive high nonroutine analytical skill share (share of demand for nonroutine analytical tasks/skills)
n=67000000
0.48
Both displacement and augmentation exposure are associated with a de-coring pattern: a shallower and more dispersed skill portfolio with within-category importance diverging from share movements. Organizational Efficiency negative high skill portfolio depth and dispersion; divergence between within-category importance and category shares
n=67000000
0.48
The de-coring and skill-demand changes are concentrated among low entry-threshold, small firms. Firm Productivity negative high heterogeneity of skill-demand changes by firm size and entry-threshold (concentration among small, low entry-threshold firms)
n=67000000
0.48
Reskilling policy should emphasize portfolio breadth and portable competency frameworks rather than deeper single-track specialization, particularly for workers in small, lower-threshold firms. Social Protection positive high policy emphasis (recommended focus of reskilling programs)
n=67000000
0.08

Notes