The Commonplace
Home Papers Evidence Explore Trends Syntheses Digests References Docs 🎲 Workforce Futures
← Papers
Direction, evidence grade, and study type are AI-generated labels (gpt-5-mini), not human-verified. Syntheses are LLM-written. "Tensions" are machine-detected candidates, not confirmed contradictions. A research-acceleration tool, not peer review. How this is built →

AI adoption falls well short of technical potential and is highly selective: firms mainly deploy AI in finance, ICT and professional services and among middle-wage office roles in large companies, while public, education and many high-exposure occupations see little implementation.

AI Exposure vs. AI Adoption: Comparing Technological Potential and Economic Realization
Byung You Cheon, Youngmin Shin · December 31, 2025 · Korean Development Economics Association
openalex descriptive low evidence 7/10 relevance Summary only summary available; pdf_status=not_found DOI Source

Structured author observations

Linked only from stored provider relations; the raw author line above is never matched by name.

OpenAlex

Latest observation:

  1. Byung You Cheon provider ID
  2. Youngmin Shin provider ID

Semantic Scholar

Latest observation:

  1. Byung You Cheon provider ID
  2. Y. Shin provider ID
AI adoption in enterprise services diverges substantially from technical exposure: actual implementation is concentrated in certain industries, middle-wage office roles, large firms, and among male, older, and long-tenured workers, while public, education, small firms, and some high-exposure occupations show limited uptake.

Citation observations

Cumulative provider counts captured on specific dates; providers are never combined.

This study noted that the impact of artificial intelligence (AI) on the labour market is determined not merely by the potential of technological advancement, but by the structure of its actual economic application. Previous research has primarily focused on the technical potential (potential exposure) of AI to replace job-specific tasks. However, this approach failed to explain how technological advances are actually adopted within firms and organisations, and under what socio-economic constraints they are implemented. To address this, this study constructed an AI Application Indicator for Enterprise Services (AIFE). By comparing this with existing AI exposure indicators (AIOE) and GPT-based exposure indicators (AIOE_by_GPT), it analysed the gap between technical potential and economic realisation. The correlation between AIOE and AIFE was not high, at around 0.3 points, confirming a consistent gap exists between technically feasible domains and those actually implemented. Even within high-exposure occupations, the level of AI application is not uniform. While active adoption has occurred in productivity-enhancing roles such as research and development and financial management, implementation is lagging in roles where ethical judgement and social trust are paramount, such as legal, education, and administrative management. By wage bracket, actual implementation is more prevalent among middle-wage office and service workers (e.g., data management, administrative support, customer service, content review). Adoption is concentrated among males, middle-aged and older workers, large enterprises, and long-term employees, whereas female workers and those in small and medium-sized enterprises exhibit relatively lower implementation levels. By industry, high exposure and high adoption patterns emerged in finance, information and communications, and professional services, whereas the public and education sectors showed limited adoption despite high technological exposure. These results demonstrate that the diffusion of AI is not merely a technological innovation but is being realised selectively within a context where industrial structure, institutional environment, skill composition, and socio-economic conditions intersect.

Summary

Main Finding

There is a persistent and systematic gap between where AI is technically capable of replacing or augmenting tasks (technical exposure) and where AI is actually being implemented in firms and organisations (economic realisation). The study’s AI Application Indicator for Enterprise Services (AIFE) correlates only weakly (~0.3) with existing AI exposure measures (AIOE and AIOE_by_GPT), showing adoption is selective and shaped by institutional, industrial, and socio-economic contexts.

Key Points

  • Measurement: The study introduces AIFE, an indicator of actual AI application within enterprise services, and compares it to two technical exposure indicators (AIOE; AIOE_by_GPT).
  • Low correlation: AIFE and AIOE correlate at roughly 0.3, indicating many technically exposed tasks are not being realised in practice.
  • Occupational heterogeneity:
    • High adoption in productivity- and information-focused roles (R&D, financial management, data management, administrative support, customer service, content review).
    • Low adoption in roles requiring ethical judgement, trust, or social interaction (legal, education, administrative management), despite technical exposure.
  • Wage- and worker-group patterns:
    • Actual implementation concentrated among middle-wage office and service workers.
    • Adoption skewed toward male, middle-aged and older workers, employees in large firms, and long-tenured staff.
    • Female workers and employees in small and medium-sized enterprises show relatively lower implementation.
  • Industry patterns:
    • High exposure + high adoption: finance, information & communications, professional services.
    • High exposure + low adoption: public sector and education.
  • Conclusion: Diffusion of AI is not uniform technological diffusion but a selective process driven by industrial structure, institutional rules, skill composition, and broader socio-economic constraints.

Data & Methods

  • New indicator: Constructed an AI Application Indicator for Enterprise Services (AIFE) to capture realised AI use in enterprise settings (focus on services).
  • Comparison framework: Benchmarked AIFE against established technical exposure measures (AIOE) and a GPT-based exposure metric (AIOE_by_GPT).
  • Analytical approach:
    • Correlation analysis to quantify alignment between technical potential and realised application (reported correlation ≈ 0.3).
    • Cross-sectional comparisons across occupations, wage brackets, worker demographics (gender, age), firm characteristics (size, tenure), and industries to map heterogeneity in adoption.
  • Note on scope: The study emphasises realised application rather than purely task-level technical feasibility; it integrates economic and institutional factors in its empirical comparisons.

Implications for AI Economics

  • Measurement: Technical exposure metrics (AIOE, GPT-based) are insufficient alone for predicting labour-market impacts; realised-adoption indicators (like AIFE) are essential to evaluate economic effects.
  • Policy targeting: Because adoption is selective, policies (training, adjustment assistance, regulation, procurement incentives) must be targeted by industry, firm size, occupation, and demographic group rather than assuming uniform displacement risk.
  • Inequality dynamics: Selective diffusion—concentrated in certain wage brackets, firm sizes, and worker demographics—may reshape wage and employment distributions, potentially exacerbating existing inequalities.
  • Public-sector and trust-sensitive domains: High technical potential but low adoption in education, legal, and public sectors suggests regulatory, ethical, or institutional barriers that slow realisation; reforms or guidance could alter social-value trade-offs and adoption rates.
  • Research agenda:
    • Incorporate realised-adoption measures in models forecasting labour-market impacts.
    • Study firm-level adoption determinants (costs, complementarities with worker skills, institutional constraints).
    • Longitudinal and causal analyses to trace how adoption patterns evolve and affect wages, employment, and productivity across groups and sectors.

Assessment

Paper Typedescriptive Evidence Strengthlow — The study documents cross-sectional patterns and correlations between a newly constructed AI application indicator (AIFE) and exposure indices, but it does not employ an identification strategy for causal inference; results are descriptive and subject to measurement error and selection biases. Methods Rigormedium — The paper develops a novel enterprise-level application indicator and compares it systematically to existing technical-exposure metrics (including a GPT-based measure), which is a useful and transparent approach; however, the work appears to rely on observational, likely self-reported or administratively aggregated measures without causal controls, and it is vulnerable to omitted-variable bias, sample selection, and measurement validity concerns. SampleUses an enterprise-level dataset constructing an AI Application Indicator for Enterprise Services (AIFE) and compares it to existing AI exposure indices (AIOE and a GPT-based AIOE); analysis is presented across occupations, wage brackets, worker demographics (sex, age), firm size/tenure, and industries (finance, ICT, professional services, public sector, education). The exact country, time period, and sample selection criteria are not specified in the summary. Themesadoption labor_markets inequality productivity GeneralizabilityLikely limited to enterprise services and surveyed/recorded firms — may not represent informal sectors or non-service industries, Possible overrepresentation of large firms and adopters (self-selection), so prevalence estimates may be upward biased, Unclear geographic/national coverage — results may not generalize across countries with different industrial or regulatory contexts, Cross-sectional snapshot: does not capture dynamic diffusion over time or causal effects on wages/productivity, Occupational aggregation may mask within-occupation heterogeneity in tasks and adoption

Claims (10)

ClaimDirectionOutcomeConfidence & EvidenceDetails
The correlation between the technical AI exposure indicator (AIOE) and the AI Application Indicator for Enterprise Services (AIFE) is not high, at around 0.3 points. Adoption Rate negative correlation between technical AI exposure (AIOE) and observed AI application (AIFE)
Reading fidelity high
Study strength medium
around 0.3 points
0.18
There is a consistent gap between technically feasible AI domains and those actually implemented in firms (technical potential does not directly translate into economic realisation). Adoption Rate negative degree to which technical potential translates into actual AI implementation
Reading fidelity high
Study strength medium
not reported
0.18
Even within occupations classified as high-exposure to AI, the level of actual AI application is uneven across job types. Adoption Rate mixed within-occupation variation in AI application
Reading fidelity high
Study strength medium
not reported
0.18
Active AI adoption has occurred in productivity-enhancing roles such as research and development and financial management. Adoption Rate positive level of AI application in R&D and financial management occupations
Reading fidelity high
Study strength medium
not reported
0.18
AI implementation is lagging in occupations where ethical judgement and social trust are paramount, such as legal, education, and administrative management. Adoption Rate negative level of AI application in legal, education, and administrative management occupations
Reading fidelity high
Study strength medium
not reported
0.18
By wage bracket, actual AI implementation is more prevalent among middle-wage office and service workers (e.g., data management, administrative support, customer service, content review). Adoption Rate positive AI application prevalence by wage bracket (middle-wage office/service workers)
Reading fidelity high
Study strength medium
not reported
0.18
Adoption is concentrated among males, middle-aged and older workers, large enterprises, and long-term employees; female workers and those in small and medium-sized enterprises exhibit relatively lower implementation levels. Adoption Rate mixed AI application prevalence by gender, age, firm size, and tenure
Reading fidelity high
Study strength medium
not reported
0.18
By industry, high exposure and high adoption patterns emerged in finance, information and communications, and professional services. Adoption Rate positive industry-level AI application (AIFE) in finance, ICT, and professional services
Reading fidelity high
Study strength medium
not reported
0.18
The public and education sectors showed limited AI adoption despite exhibiting high technological exposure. Adoption Rate negative AI application level in public and education sectors relative to technical exposure
Reading fidelity high
Study strength medium
not reported
0.18
The diffusion of AI is selective and is realised within a context shaped by industrial structure, institutional environment, skill composition, and socio-economic conditions rather than being a uniform technological innovation. Adoption Rate mixed pattern/nature of AI diffusion across socio-economic and institutional contexts
Reading fidelity high
Study strength medium
not reported
0.18

Notes