Somewhere in the last eighteen months, a slide started appearing in board decks that nobody quite challenges: cost per employee falling, revenue per head rising, margins expanding, all of it attributed, correctly, to generative AI absorbing work once done by the newest, cheapest people in the building. It is a genuinely good slide. It is also, for most companies presenting it, an incomplete one, because nobody has added the line that matters most: who, in twelve years, will be qualified to run this division.
That question will not show up on any dashboard a CEO currently watches. It will not show up for a decade. By the time it does, the decision that caused it will have been made by people long since moved on, which is precisely why almost no board is asking it now.
The data is no longer anecdotal
Erik Brynjolfsson and Bharat Chandar at Stanford’s Digital Economy Lab spent the past three years tracking millions of workers across tens of thousands of firms using ADP payroll data, in what they describe as the largest real-time effort yet to measure AI’s effect on employment. Their finding, published this year, is specific and uncomfortable: workers aged 22 to 25 in occupations most exposed to AI, software engineering, marketing, customer service, have seen employment fall 16 percent relative to trend since ChatGPT’s late-2022 debut. Workers over 30 in the same exposed occupations grew employment 6 to 12 percent over the identical period. This is not a story about AI destroying jobs in aggregate. It is a story about AI destroying a very particular rung of the ladder while leaving the rungs above it intact, or even reinforced.
The Burning Glass Institute’s analysis of job postings from 2018 to 2024, released under the title “No Country for Young Grads,” quantifies the same pattern from a different angle. Entry-level software development postings requiring three years of experience or less fell from 43 percent of listings to 28 percent. Entry-level data analysis roles fell from 35 percent to 22 percent. Entry-level consulting roles fell from 41 percent to 26 percent. As one career strategist quoted in that research put it, employers used to hire potential; now they hire productivity.
Wall Street offers the sharpest version of this because its junior tier was always the most standardised training system in corporate life. Debasish Patnaik, a senior partner at McKinsey’s QuantumBlack, has described banks cutting junior analyst classes by as much as two-thirds industry-wide. Jamie Dimon has said plainly that AI will eliminate jobs. Jane Fraser has said some roles will simply no longer be required. Goldman Sachs president John Waldron called parts of the analyst function a “human assembly line,” an accurate description of the work and an unintentionally precise one of what is now being switched off.
What junior work actually taught
The mistake in most commentary on this shift is treating it as a jobs story. It is really a training story. Junior roles were never primarily about the memo, the model, or the dataset. They were repetition under supervision with real consequences attached: draft the memo badly enough and a partner tells you why; clean the data wrong and the forecast breaks in front of a client. Carina Cortez, chief people officer at Cornerstone OnDemand, has made the point directly: entry-level roles function as structured learning environments, not simply as output generators. When AI absorbs the output, it does not preserve the learning environment. It deletes it, quietly, as a side effect of doing the task better.
Harvard Business Review contributors Julia Shin and Sandra Sucher have documented the consequence one layer up: middle managers are increasingly absorbing the work of validating AI output, often without any adjustment to their workload or expectations. The judgment that used to be built slowly, by junior staff learning where the numbers usually lie, is now being demanded instantly of managers who never had the chance to build it themselves, because they came up through a version of the pipeline that had already started thinning.
This is compounding with a second, largely unconnected trend: the deliberate flattening of management layers. Gartner predicted in 2024 that a fifth of organisations would eliminate more than half their middle-management roles by 2026. Live Data Technologies data, reported by the Wall Street Journal, shows manager headcount down 6.1 percent between May 2022 and May 2025, and Google has reported 35 percent fewer managers overseeing small teams year over year. Gallup finds the average manager’s span of control rose from 10.9 direct reports in 2024 to 12.1 in 2025. DDI’s 2025 Global Leadership Forecast, drawing on more than 10,000 leaders across 50 countries, puts leadership bench strength at just 20 percent, its lowest recorded level. Korn Ferry’s Lesley Uren summarised the mechanism bluntly: a leaner organisation today can produce a leadership crisis tomorrow.
A third, quieter pressure feeds the same problem: fewer young employees appear to want the job that is disappearing. Deloitte’s 2025-26 survey of more than 23,000 Gen Z and millennial respondents across 44 countries found only 6 percent cite reaching a leadership position as their primary career ambition. Even where a supervised path to management still exists, demand for walking it is thinning from the other direction.
India’s version of the problem is structurally larger
India’s technology services sector has functioned for three decades as the country’s largest formal-sector on-ramp into the salaried middle class, absorbing a meaningful share of the roughly 1.5 million engineering graduates it produces every year. That model is now visibly straining, with a twist that makes India’s exposure more severe than the headline global numbers suggest.
Net fresher additions across India’s top IT services firms fell from over 50,000 in a quarter to under 5,000 in the June 2026 quarter, a decline of roughly 72 percent, even as several of the same firms continued adding experienced, specialised headcount. TCS alone reduced its workforce by more than 25,000 people in the first nine months of FY26. HCLTech’s chief executive summarised the new logic candidly: revenue grew 4 to 5 percent while headcount did not move at all. NASSCOM’s public framing remains that AI will ultimately expand India’s technology workforce, and the country’s AI talent pool is genuinely projected to reach 1.25 million professionals by 2027. Both things are true simultaneously: the AI-specialist segment is expanding while the mass-hiring, learn-on-the-job segment that historically absorbed India’s engineering graduates is contracting in real time.
The World Bank’s 2026 World Development Report frames this as a structural distinction, not a paradox: developing economies face lower aggregate automation risk than high-income ones, roughly 4.5 percent of jobs versus 14.2 percent, but the jobs at highest risk within developing economies sit disproportionately in the information technology, finance, and business-services roles that India built its graduate-absorption model around. South Korea’s youth labour market previews what redistribution without net creation looks like: youth employment fell by 285,000 between mid-2022 and mid-2026, with 94 percent of that decline concentrated in AI-exposed sectors, while employment among workers in their fifties in the same sectors rose by 230,000. The jobs did not vanish from the economy. They moved to people who had already banked their experience, leaving nothing for the next cohort to bank theirs.
Layered onto this is a pre-existing employability gap unrelated to AI: various estimates put the share of Indian engineering graduates unable to secure relevant employment as high as 72 to 83 percent, driven by a curriculum that has lagged industry need for years. AI did not create India’s entry-level crisis. It arrived on top of one, and is now closing the door that used to compensate for it: “we will train you if the college did not.”
The decision no board is actually pricing
None of this is an argument against using AI to automate junior-level work. The productivity case is real and, for most firms, competitively unavoidable. The argument is that cutting entry-level hiring is currently being evaluated as a cost decision when it is functionally a capital allocation decision with a payback period measured in a decade, and almost nobody is running it through that lens. A company that stops training analysts, associates, or engineers today is not saving money. It is borrowing leadership capacity from 2036 and spending it now, at a discount rate nobody has bothered to calculate, because the invoice will land on someone else’s tenure.
The practical response is specific, not a vague appeal to invest in people. It means treating leadership bench strength as an audited metric the board reviews with the same seriousness as customer concentration or key-person risk, since DDI’s data suggests most companies currently cannot answer the question even if asked. It means redesigning the junior role so AI absorbs volume while a deliberately preserved sliver of supervised, consequential decision-making remains, because that sliver, not the drafting or the data-cleaning, was always the actual training mechanism. And it means making the make-or-buy leadership choice explicit: a company that has decided, in effect, to buy its future leaders from the market rather than grow them should say so, price that dependence honestly, and stop being surprised when the bench turns out empty.
The CEOs benefiting most from this wave of AI-driven efficiency are, almost by definition, the ones who came up through an apprenticeship model that no longer exists for the people now entering their organisations. The uncomfortable question for the next decade is not whether AI is good at the entry-level job. It plainly is. It is whether anyone is being given the chance to become good enough to hold the job that used to come after it.
