Follow the money: a startup-based measure of AI exposure across occupations, industries and regions

Technical Summary: Follow the Money: A Startup-Based Measure of AI Exposure Across Occupations, Industries and Regions

Problem Statement
The rapid integration of Artificial Intelligence (AI) into the workplace necessitates robust metrics to evaluate its tangible impact on the labor market. Existing measures of AI occupational exposure, such as the AI Occupational Exposure (AIOE) index proposed by Felten et al. (2019, 2021), largely rely on the theoretical potential for AI to substitute or complement human labor based on technical feasibility. These approaches often utilize expert judgments, crowd-sourced evaluations, or patent analyses to map AI capabilities to occupational tasks. However, these metrics suffer from significant limitations: they are often subjective, non-reproducible, or lag behind current developments due to patent filing delays. Crucially, they measure potential exposure rather than actual adoption, failing to account for economic viability, social desirability, and the specific market choices of AI developers. Consequently, there is a lack of insight into which occupations are genuinely being targeted by current AI innovations, limiting the utility of these indices for policymakers.

Methodology
To address these gaps, the authors introduce the AI Startup Exposure (AISE) index, a novel metric grounded in actual investment activity rather than theoretical capability. The methodology proceeds as follows:

1. Data Sources:

   • Occupational Descriptions: The study utilizes the O*NET database, specifically the short summary descriptions of essential tasks for 1,016 Standard Occupational Classification (SOC) jobs. These summaries were chosen over full task lists to reduce noise and focus on salient job characteristics.
   • AI Innovation Data: The authors scraped descriptions of startups funded by Y Combinator (YC), a leading US-based venture capital firm and accelerator. They filtered for startups with specific AI-related tags (e.g., "generative AI," "machine learning," "computer vision"), resulting in a dataset of 958 AI startups funded between 2005 and March 2024.
   • Geographic and Sectoral Data: Employment shares for occupations were drawn from the US Bureau of Labor Statistics' Quarterly Census of Employment and Wages (QCEW) to calculate weighted averages for Metropolitan Statistical Areas (MSAs) and NAICS sectors.

2. LLM-Based Analysis:

   • The authors employ Meta's Llama 3 (an open-weight, state-of-the-art Large Language Model) to assess the similarity between startup product descriptions and occupational task summaries.
   • Prompting Strategy: For each pair of (Startup, Occupation), the LLM is prompted to determine if the startup's product or service can directly replace humans in performing one or more of the essential tasks described in the O*NET summary. The model is instructed to reply only "yes" or "no" based strictly on the provided text.
   • Scoring: The Occupational AISE for a specific job is calculated as the normalized number of YC-funded AI startups identified by the LLM as capable of substituting essential tasks for that job.

3. Validation and Comparison:

   • The AISE index is compared against the standard AIOE index (Felten et al., 2021). While AIOE is an ability-based measure derived from crowd-sourced surveys linking AI applications to 52 O*NET abilities, AISE is a task-based measure derived from actual startup offerings.
   • Robotic Startup Exposure (RSE): A parallel methodology is applied to 103 robotics-focused startups to create an RSE index, analyzing the intersection of AI and physical automation.

4. Geographic and Sectoral Aggregation:

   • Geographic exposure is computed as the employment-weighted average of Occupational AISE across MSAs.
   • Sectoral exposure is computed as the employment-weighted average of Occupational AISE across 19 two-digit NAICS sectors.

Key Results

• Occupational Exposure Patterns:

  • High Exposure: The occupations with the highest AISE scores are those involving routine organizational and information processing tasks. Top examples include General office clerks, Data scientists, Computer and information systems managers, and Market research analysts. These roles are increasingly targeted by startups developing LLMs, multi-agent systems, and automation tools.
  • Low Exposure: Occupations with low AISE scores include Athletes, Magistrate judges, and Pediatric surgeons. Despite having high theoretical exposure according to AIOE (due to cognitive complexity), these roles are less targeted by startups. The authors attribute this to ethical constraints, high-stakes decision-making, and the necessity of human accountability, which create barriers to adoption beyond mere technical feasibility.
  • Divergence from AIOE: A significant finding is the heterogeneity between AISE and AIOE. While AIOE suggests high-skilled, high-education jobs are uniformly at risk, AISE reveals that many high-skill roles (e.g., judges, doctors) have low actual exposure. This "reverse skill-bias" suggests that the need for advanced skills and the high cost of errors in these roles deter current AI investment, even when the tasks are theoretically automatable.

• Geographic Distribution:

  • AI exposure is highly concentrated in knowledge-intensive metropolitan areas with expanding digital economies, such as San Jose, San Francisco, Seattle, Austin, and Boulder.
  • Conversely, the Midwest exhibits the lowest exposure, correlating with its reliance on manufacturing and agriculture, sectors that have been slower to integrate AI.

• Sectoral Distribution:

  • Service-oriented sectors requiring high levels of information processing and education show the greatest exposure.
  • Sectors like Educational Services and Health Care, while education-intensive, show relatively lower exposure due to the high-stakes nature of their core tasks.
  • Traditional sectors like Construction and Agriculture are less frequently targeted by AI startups.

• AI and Robotics Convergence:

  • The preliminary RSE analysis indicates that while many low-AISE jobs (manual tasks) have high RSE, many high-AISE jobs also show high RSE. This suggests that the joint action of AI and robotics may lead to more pervasive disruption across both white-collar and blue-collar occupations.

Significance and Claims
The paper claims that the AISE index offers a more realistic, forward-looking, and economically grounded assessment of AI's impact on the labor market compared to existing theoretical metrics. By focusing on venture-backed innovations, the index captures the "societal desirability" and "market-oriented choices" of AI players, which are critical determinants of actual adoption.

Key contributions include:

1. Challenging the "High-Skills Catastrophe": The findings suggest that fears of widespread, immediate displacement of high-skilled professionals may be overstated. Instead, AI adoption appears to be gradual, prioritizing routine, economically viable tasks while leaving expert-driven, high-stakes roles relatively shielded in the short term.
2. Policy Utility: The AISE provides a dynamic, reproducible tool for policymakers to monitor the evolving landscape of AI adoption. Because it relies on open-weight LLMs and publicly available startup data, the index can be updated in near real-time as new funding rounds occur.
3. Methodological Innovation: The study demonstrates how combining occupational descriptions with startup data via LLMs can disentangle theoretical potential from actual market interest, offering a nuanced view of how AI is reshaping the labor landscape.

The authors remain modest regarding the net labor market effects (substitution vs. complementarity), noting that AISE measures exposure to innovation rather than definitive job loss. They acknowledge limitations, including the potential bias of Y Combinator's funding preferences and the assumption of occupational homogeneity across firms and regions, suggesting future work should expand data sources and incorporate firm-level dynamics.