The debate on artificial intelligence and employment is increasingly framed in extremes: between predictions of widespread job loss and claims of negligible impact. In practice, labour markets rarely adjust in such abrupt ways. Technological change works more subtly, through shifts in hiring patterns, changes in task allocation, and gradual reallocation across workers.

It is at these margins that the earliest effects of AI are now beginning to appear.

Recent evidence from the US titled Canaries in the Coal Mine?, based on high-frequency payroll data, points to a striking pattern. Employment among early-career workers (particularly those aged 22–25) has declined in occupations most exposed to generative AI, even as employment for more experienced workers has continued to grow.

In some cases, employment for this group has fallen by 6 per cent to 20 per cent since late 2022, while older cohorts have seen steady gains. The implication is not one of aggregate job loss, but of a weakening in entry-level hiring.

The question is whether a similar shift is underway in India.

To examine this, we analyse monthly net payroll data from the monthly Employees’ Provident Fund (EPFO) payroll data (HFI series), covering April 2020-21 to July 2025-26, disaggregated by age group. Net payroll additions (a measure of employment) are defined as new EPF subscribers, minus members exiting the system, plus those who rejoin and resubscribe. In effect, this captures the net flow of workers into formal payroll each month.

We use November 2022 (the public release of generative AI tools) a natural reference point to compare employment dynamics before and after the acceleration in AI adoption.

A first look at the data, in levels, is reassuring.

Formal payroll additions across all age groups have continued to grow, with no visible collapse after late 2022. The series move broadly together, reflecting macroeconomic conditions rather than any clear structural break. At this level of aggregation, there is little to suggest disruption. But this is precisely where headline numbers can mislead.

The relevant question is not whether employment is rising, but whether it is rising evenly across groups.

Reading the data

Instead of focusing on absolute levels, we ask: how have formal payroll additions for each age group changed relative to where it stood when AI began to diffuse rapidly?

This is done by converting employment into an index, where November 2022 is set equal to 100: if the index rises to 120, formal payroll additions are 20 per cent higher than at that point; if it falls to 95, formal payroll additions are 5 per cent lower. In other words, the level 100 in November 2022 represents a certain level of monthly employment.

The resulting picture is more informative. While all groups continue to move together, the pace of growth begins to diverge.

A shift at the margin

We then compare each cohort’s average monthly formal payroll additions levels before and after November 2022, excluding the earlier Covid-disrupted period. This allows us to see how different age groups have fared relative to the onset of AI adoption.

The pattern is clear. While overall formal payroll additions has grown by nearly 14 per cent in the post-AI period, growth for workers aged 22-25 has been almost flat. In contrast, formal payroll additions gains for older cohorts have been substantial.

This is not a story of job loss. Younger workers are not being displaced in absolute terms. But they are clearly not sharing in formal payroll additions growth to the same extent as more experienced workers. The labour market is expanding, but unevenly.

This pattern is directionally consistent with international evidence, though significantly weaker. In the US, employment among 22-25-year-olds in AI-exposed occupations has declined sharply, by as much as 15 log points in relative terms and up to 20 per cent in levels. No such contraction is visible here. Instead, what emerges is a more muted version of the same mechanism: not decline, but relative stagnation.

This is consistent with emerging evidence on how generative AI is affecting labour demand. It substitutes for routine, codifiable tasks (often performed by entry-level workers) while complementing tasks that rely on experience and judgment. Firms need not reduce employment; they adjust hiring at the margin.

These findings should be interpreted with caution. EPFO data captures net additions to formal employment, not total employment. It covers only the formal sector, leaving out informal and gig work where AI effects may differ. Age-wise trends may also reflect policy and compliance changes, not just labour demand.

Finally, the absence of occupation-level AI exposure means aggregate patterns may understate sector-specific impacts. These limitations do not invalidate the findings, but they do suggest that the observed patterns should be interpreted as early signals rather than definitive evidence.

The real signal

Even with these caveats, one pattern stands out: the relative stagnation of the 22-25 cohort. Entry-level hiring is the primary mechanism through which workers enter the labour market and accumulate experience. A slowdown at this margin does not immediately show up as unemployment. It shows up instead as weaker early-career progression and, over time, widening disparities across cohorts.

Employment growth is increasingly tilted toward workers with experience, while those at the entry level face a slowdown.

Technological transitions rarely announce themselves through crisis. They unfold through incremental adjustments that are easy to overlook in real time. The earliest signal is not job loss; it is a change in who benefits from growth.

In India, that shift appears to have begun. It is not yet large enough to dominate aggregate statistics, but it is sufficiently clear to warrant attention.

The policy challenge is to recognise a transition while it is still unfolding. Because if the entry gate to the labour market narrows, the consequences will not be immediate, but they will be lasting.

The writer is Lead Economist and Head of Center of Data for Economic Decision-making (CoDED), Pahle India Foundation

Published on June 26, 2026