Startup-based AI index predicts which jobs AI will change first

A lot of the anxiety around artificial intelligence has centered on a blunt question: which jobs can the technology do? A new analysis by Enrico Maria Fenoaltea, and his team at Centro Studi e Ricerche Enrico Fermi, suggests that may be the wrong place to start. The better question may be which jobs companies and investors actually want AI to enter first, and the answer is narrower, messier, and more social than many forecasts imply.

Instead of treating AI as a force that will sweep through the labor market evenly, the study tracks where venture-backed startups are building products right now. By matching those startup products to job descriptions in the federal O*NET occupation database, the authors created what they call the Occupational AI Startup Exposure index, or AISE. The idea is simple: if funded startups are pursuing a use case, that use case has already cleared at least some tests of commercial appeal, investor confidence, and social acceptability.

That changes the picture.

The jobs that rose to the top were not a random cross-section of the economy. General office clerks ranked highest, followed by data scientists, computer and information systems managers, and market research analysts and marketing specialists. These are jobs packed with information processing, organization, planning, writing, sorting, and software-related work, the kinds of activities AI companies are already trying to productize.

The study’s basic argument is that technical capability alone does not decide where AI lands. A language model may be able to perform parts of many jobs in theory, but that does not mean employers, workers, regulators, or the public will welcome it into the heart of those roles.

That helps explain why some occupations came in low on the new index even when they look vulnerable on paper. Athletes, chefs, and construction workers were among the least exposed because so much of their work is physical. But the more revealing cases were jobs such as judges and pediatric surgeons. Both involve skills that AI might support in limited ways, yet few people seem ready to hand over the core of those roles to a machine.

The authors argue that this gap matters. Existing measures of AI exposure often focus on whether a system could perform a task. Their index tries to capture something closer to near-future pressure by looking at funded startups, especially those backed by Y Combinator, and using a version of Meta’s Llama 3 model to compare startup descriptions with occupational descriptions. They also checked the pattern against a sample of funded European startups and found similar results.

In that sense, the index does not measure actual adoption, and it does not predict with certainty which jobs will shrink or grow. It also stays neutral on whether AI complements a worker or replaces one. What it does offer is a view of where the startup market is aiming.

That is where the study departs from some of the grimmer narratives about a coming white-collar wipeout.

The authors compared their startup-based index with a well-known benchmark that estimates exposure from the underlying abilities a job requires. Those ability-based measures often rank highly educated occupations as especially exposed, because many of those jobs depend on reasoning, writing, classification, and other cognitive skills that AI systems can imitate.

The new index agrees up to a point. High-skilled office work is still very much in the frame. But once jobs involve a wider bundle of crucial skills, high stakes, or ethically loaded decisions, startup interest tends to thin out.

Lawyers and database administrators offer a clear contrast. On an abstract ability-based measure, the two can look similarly exposed because both depend on reasoning and information ordering. In the startup-based index, they diverge sharply. Database administration fits neatly into a world of optimization, workflow, and software tools. Law carries a heavier load of judgment, trust, accountability, and social consequence.

The same pattern shows up in teaching, counseling, and medicine. Large language models may be able to handle parts of those jobs, especially routine or secondary tasks, but the study suggests that is different from seeing the professions themselves become prime targets for AI startups. In many of these roles, the issue is not simply whether AI can produce an answer. It is whether people will trust it to make consequential decisions, whether errors are tolerable, and whether society wants those decisions handed off at all.

The researchers also projected the index onto places and industries in the United States, using employment data to estimate where exposure might be concentrated.

The highest levels appeared in metro areas with strong digital economies and deep tech ecosystems, including Silicon Valley, the San Francisco Bay Area, Boston, Washington, Austin, Denver-Boulder, Salt Lake City-Provo, and Miami. Regions with stronger ties to manufacturing and agriculture, especially in parts of the Midwest, looked less exposed.

By sector, the most exposed industries were information, management, professional and technical services, finance and insurance, wholesale trade, and administrative support. These are fields where text, data, forecasting, compliance, customer support, documentation, and planning dominate the workday. Construction and agriculture sat at the other end of the scale. Education and health care fell in the middle, reflecting their mix of paperwork-heavy tasks and jobs where human judgment still carries unusual weight.

That split could matter for how AI changes work. The study suggests that jobs with high theoretical exposure but lower startup exposure may be more likely to see AI as an assistant rather than a replacement. By contrast, occupations that score high on both measures could face more direct substitution pressure.

There is also an important caveat. When the authors narrowed their lens to startups combining AI with robotics, the field widened. Manual occupations that looked relatively sheltered in the software-focused index became more exposed once machines could act in the physical world. If that trend accelerates, the current boundary between cognitive and manual work could shift fast.

The study points toward a more selective view of AI’s labor impact. In the near term, the strongest pressure may fall on work that is routine, digital, and easy to package into a product. That does not mean every exposed job disappears, but it does suggest that clerical and data-heavy occupations may feel change first.

For policymakers, the findings argue for a narrower and more targeted response than broad claims about economy-wide disruption. Training, worker protections, and transition planning may be most urgent in office-based service work and in metro areas with dense tech and finance activity.

For employers and workers, the message is that AI adoption will likely depend not only on what systems can do, but also on trust, risk, regulation, and whether people accept the trade-offs involved.

The spread of AI may be powerful, but the study suggests it will move through the labor market along the paths society is willing to open.

Research findings are available online in the journal PNAS Nexus.