Future of work

Researchers have developed the AI Startup Exposure (AISE) index, a new approach to measuring occupational exposure to artificial intelligence (AI) that goes beyond theoretical assessments to examine real-world AI adoption patterns.  

Unlike exposure measures that focus on technical feasibility, AISE analyzes the actual AI applications being developed by venture-backed startups to provide insights into which jobs face (or are exposed to) automation by these.  

The methodology leverages data from 958 AI startups funded by Y Combinator from 2005 to March 2024 (with most receiving funding after 2020). It uses Meta’s Llama 3 language model to assess whether each startup’s AI application could replace humans for essential tasks in more than 1,000 occupations. This approach captures not only the technical feasibility, but also the economic viability and social acceptability of AI adoption.  

The findings reveal significant differences versus previous measures of AI exposure.  

Traditional indices—such as the AI Occupational Exposure (AIOE) index by Felten et al., automation risk assessments by Frey and Osborne, and patent-based measures—typically rely on expert assessments, crowd-sourced evaluations, and/or patent analysis to determine theoretical automation potential rather than examining actual market activity. These indices have tended to suggest that high-skilled professionals like judges, pediatric surgeons, and specialized medical practitioners face high AI exposure based on their cognitive abilities. But AISE finds that these professionals have relatively low exposure.  

For example, database administrators and lawyers both require sophisticated cognitive skills, but they show vastly different startup-based exposure levels, with database administrators facing much higher practical automation pressure. This reflects the reality that high-stakes, ethically sensitive roles present substantial barriers to AI integration—beyond mere technical feasibility.  

In contrast, occupations that involve routine organizational tasks have the highest AISE scores. These positions typically involve information processing, data analysis, and administrative tasks that startups are targeting for automation.  

The research also mapped AI exposure geographically and by sector. In the U.S., metropolitan areas with robust tech ecosystems (such as San José and San Francisco) are significantly more exposed to AI than regions focused on manufacturing and agriculture, reflecting its economic reliance on industries that have been slower to adopt AI technologies.  

At the sectoral level, service-oriented industries requiring high levels of information processing show greater exposure compared to traditional sectors like construction and agriculture. However, industries like educational services and health care exhibit lower exposure (despite requiring workers to have high education levels), indicating that specialized knowledge and skills are barriers to practical AI adoption.  

The study’s approach addresses the limitations of existing exposure measures that rely on expert assessments or patent data, which can be subjective or lag real developments. When compared with the established AIOE index across 873 overlapping occupations, AISE showed strong correlations while also revealing distinct patterns. By focusing on startup activity, AISE captures market-driven innovations and reflects investor confidence in specific AI applications.  

The researchers also explored robotic automation exposure using data from 103 robotics-focused startups and found that the combination of AI and robotics could affect an even broader range of occupations. Many jobs with low AI exposure showed high exposure to AI-powered robotics, particularly roles involving manual tasks. This suggests that the combination of AI and robotics may create more widespread workplace disruption than either technology alone.  

These findings have important implications for workforce planning and policy development. Rather than the widespread displacement of high-skilled workers that some analyses have suggested, this research points to a more nuanced pattern where routine cognitive tasks face greater immediate risk than complex professional roles.  

The geographic concentration of AI exposure also highlights the need for region-specific approaches to managing technological transition.  

The AISE methodology offers advantages for ongoing analysis. It can be updated continuously with new startup data, providing real-time insights into evolving AI impacts. It could also inform more responsive policy development while helping stakeholders anticipate changes in the labour market as AI technologies continue to develop.