AI adoption patterns differ by technology: firms with STEM-rich workforces are more likely to adopt core AI tools, whereas generative AI is associated with firms employing more non‑STEM university graduates. Firm size and digital maturity predict both adoption and how broadly technologies are deployed, while commonly used occupational exposure measures vary widely in their ability to detect real-world adoption.

Main Finding

AI adoption is highly technology-specific: different types of AI and related advanced technologies diffuse through firms via distinct channels. Firm-level predictors (size, digital infrastructure) broadly forecast adoption, but workforce composition matters in technology-specific ways (STEM for core AI; non‑STEM university education for generative AI). Adoption predictors differ from predictors of deployment breadth, and commonly used occupational exposure measures vary in how well they capture actual adoption.

Key Points

• Adoption is not monolithic:
  • Core AI (e.g., machine learning) correlates with STEM-educated workforces.
  • Generative AI correlates with non‑STEM university-educated workforces, implying different human-capital complementarities.
• Predictors:
  • Firm size and digital maturity/infrastructure predict both adoption and the breadth of deployment.
  • Workforce composition mainly predicts whether a technology is adopted, not how widely it is deployed internally.
• Deployment patterns:
  • Machine learning and natural language processing are used across multiple business functions.
  • Other advanced technologies remain concentrated in specific operational domains.
• Individual-level patterns:
  • Awareness of AI use at work is concentrated among managers and high-skilled employees.
  • Self-reported AI knowledge is higher for younger and more educated individuals.
• Measurement:
  • Occupational AI exposure measures differ substantially in predictive power.
  • Benchmark-based exposure measures outperform patent-based and LLM-focused alternatives in predicting observed adoption.

Data & Methods

• Data sources:
  • Two waves of nationally representative Danish firm surveys covering AI and related advanced technology adoption.
  • Linked employer-employee administrative registers for detailed firm and worker characteristics.
• Empirical approach:
  • Cross-tabulation and regression analyses linking firm adoption of specific technologies to firm characteristics (size, digital infrastructure) and workforce composition (education, STEM vs non‑STEM).
  • Analysis of deployment breadth across business functions by technology type.
  • Individual-level analyses of awareness and self-reported AI knowledge by occupation, age, education, and managerial status.
  • Comparative evaluation of different occupational AI exposure measures (benchmark-based, patent-based, LLM-focused) against observed adoption.
• Identification strategy:
  • Use of rich administrative controls and representative sampling to assess correlates of adoption and deployment; technology-specific analyses to isolate heterogeneity.

Implications for AI Economics

• Measurement and research design:
  • Treat AI types separately in empirical work—aggregating under “AI” can mask distinct diffusion patterns and lead to misleading inferences about labor market impacts.
  • Use benchmark-based occupational exposure measures when possible, as they better predict actual adoption than patent- or LLM-focused metrics.
• Human capital and labor markets:
  • Policy responses should be targeted by technology: training for STEM skills may support diffusion of core AI, while broader university-level upskilling may facilitate generative-AI adoption.
  • Concentrated awareness among managers and high-skilled workers suggests information frictions; policy or firm-level interventions to raise awareness and practical knowledge could shift adoption.
• Firm strategy and policy:
  • Investments in digital infrastructure and scaling (size effects) are central to both adoption and deployment breadth—policies that lower fixed costs of digital maturity can broaden diffusion.
  • Since deployment breadth is distinct from adoption, studies of productivity and labor reallocation should consider both whether a firm adopts and how widely it uses a technology internally.
• Inequality and reallocation:
  • Different complementarities between technologies and worker skills imply heterogeneous effects across occupations and educational groups, with potential implications for wage inequality and job tasks.
• Future research directions:
  • Study causal effects of technology-specific adoption on productivity, wages, and occupational task content.
  • Extend analysis to other countries to assess generalizability.
  • Develop and validate improved occupational exposure metrics tailored to particular AI technologies.