Experience matters: Decoding first signs of Al's impact on job market

A new study leveraging massive real-time US payroll data has found that generative AI has begun having significant and disproportionate impact on entry-level workers in certain jobs, including software developers and customer service agents.

- Binayak Dasgupta

The study by researchers from Stanford University's Digital Economy Lab, noted workers in Al-exposed occupations experienced a 13% decline in employment between late 2022 and July 2025 while their colleagues over 35 in identical roles maintained stable or growing job prospects. For these older workers, job opportunities grew 6-9%.

For highly Al-exposed jobs like software developers and customer service agents, people aged 22 to 25 saw a very striking decline in employment in the last few years in the US,” said Bharat Chandar, one of the Stanford University researchers who authored the paper, in an interaction over email.

When determining which jobs are AI exposed, the research zeroed in on an important distinction: can these AI tools automate these jobs, or do they work toaugment an employee's work?

Chandar said the disruptions are more evident in occupations where LLMs (large language models the type of technology behind ChatGPT, or Claude) “automate work than for ones where it augments work”. For instance, “for less-exposed jobs such as health aides we do not see these declines”.

The insight might be intuitive, but it is now backed by empirical evidence ~ at least from the US market. The study - titled Canaries in the Coal Mine ~ is based on millions of records from ADP, America’s largest payroll processing company, allowing researchers to track employment and earning changes with granularity typically not possible with official labour statistics.

The analysis found that employment for software developers aged 22-25, for instance, fell nearly 20% from its peak in late 2022, while workers over 35 in the same field saw employment grow by 6-9%. Customer service representatives, another Al-automated role, showed similar age-based divergences.