Half a million jobs vanished. Nobody was fired.

Some time in September 2025, a 32-year-old quality-review analyst named Megha S. was called into a meeting at the Bengaluru software firm where she checked customer-service calls for roughly $10,000 a year. The company had adopted an AI tool that reviews the calls automatically. "I was told I am the first one who has been replaced by AI," she told Reuters. It was a few weeks before the festive season. She had not told her parents.

Hold that sentence, the first one, because it is precise in a way its speakers didn't intend. Megha at least got in. She had a decade inside the industry. She got the meeting, the sentence, the severance conversation. The larger event of this decade is quieter and has no meeting at all. Nobody was fired from the five lakh entry-level jobs that disappeared between FY22 and FY25. Nobody could be. The people who would have held them never got the call. There is no labour complaint you can file about an offer letter that was never printed.

A layoff has a face, a date, a ministry summons. A closed gate has none. It shows up only as statistical silence: placement percentages sliding at a Belagavi engineering college, "batch of 2025" LinkedIn bios that never acquire a company name, 24.76 lakh graduates, MBAs and PhDs applying for 53,749 peon posts in Rajasthan. The previous two dossiers in this series argued that India rents its AI substrate and mis-spends its sovereignty. This one is about the bill arriving. Not in the data centre. On the payroll of the youngest people in the country.

Start with what did not happen, because the honest version is stranger than the doom version. India's tech industry crossed $315 billion in revenue in FY26, growing 6.1%. TCS still onboarded some 42,000 trainees in FY26 and 14,000 freshers in a single quarter of 2026. The machine runs. But look at the joints. NASSCOM's net-addition series reads like a heart monitor: +445,000 in FY22, then +290,000, then +60,000, then +126,000, then +135,000. Headcount now grows at 2.3% while revenue grows nearly three times faster. In FY24, TCS, Infosys and Wipro together shed 63,759 people, the first collective decline in their history. In July 2025 TCS announced the largest layoff it has ever made, and over the following two quarters its headcount fell by more than 30,000, which is more than the announced cut, because hiring quietly slowed underneath it. That same year TCS posted its first-ever annual decline in dollar revenue while reporting $2.3 billion in AI revenue. The company got more profitable per person as it employed fewer people. That is not a slump. That is a new operating model.

The freshers absorbed the adjustment first and worst, because they are the adjustable part. Infosys hired 50,000 freshers in FY23 and 11,900 in FY24, a 76% cut in one year, with campus recruitment halted for three consecutive quarters. Wipro, in 2023, offered thousands of waiting recruits roughly half the salary they had signed for: ₹3.5 lakh against a promised ₹6.5 lakh, after making them wait more than a year. Some 2,000 Infosys offer-holders from 2022 spent two years in onboarding limbo. And for the lucky ones who do get in, the rung itself has not been repriced in a generation.

"What we saw in the last 30 years was a fairly linear scaling of revenues and people… time is already up for that business model." C. VIJAYAKUMAR, CEO, HCLTECH · NASSCOM LEADERSHIP FORUM, FEB 2025

The owners of the model are not hiding the mechanism. Vijayakumar again, in 2026: renewal deals now carry 3–5% "AI deflation," meaning "a $100 million deal would be much lesser today. Maybe 80 million." Clients aim "to double the revenue with half the existing headcount." And NASSCOM's 2026 review describes the industry's shift from "volume hiring" to favouring "job-ready candidates" over the hire-then-train model. Read that phrase twice. Hire-then-train was not a perk of Indian IT. It was Indian IT: the machine that took a lakh of B.Tech graduates of wildly uneven preparation every year and, in six months of training bays in Mysore and Chennai, made them employable. "Job-ready" retires that machine. The gate isn't just narrower. The ramp leading to it has been dismantled.

The cleanest data comes from millions of American payroll records. Stanford's Digital Economy Lab found that employment for 22-to-25-year-olds in the most AI-exposed occupations, software development and customer service, fell 13% relative to older workers in the same jobs. The revised version of the paper moved that to roughly 16%, and young software developers specifically fell nearly 20% from their late-2022 peak. Older workers in the very same occupations stayed stable or grew. By April 2026 the divergence was still widening, not fading.

And it is not one study. Point any instrument at the entry gate and the needle jumps the same way. Indeed's postings data found US software postings recovered 15% after AI coding agents launched, but 71% of the rebound was senior roles. Their name for it is the thesis of this act: "seniority-biased technological change."

Why the bottom rung specifically? The research answers with an uncomfortable elegance. Every major study of assistive AI found it helps novices most: customer-support agents in the lowest skill quintile gained around 34% productivity while top agents gained roughly nothing; junior developers gained 27–39% versus 8–13% for seniors. That sounded like good news for juniors. It was actually the autopsy. What those studies proved is that entry-level skill is the most substitutable component of production. The thing AI replicates most easily is precisely the thing a fresher sells. And as tools went agentic, even the consolation prize expired: Anthropic's analysis of 400,000 Claude Code sessions found experts succeed at twice the rate of novices and extract more than double the work per instruction. PwC now measures entry-level postings in exposed occupations as seven times more likely to demand traditionally senior skills. They call it "seniorization." The junior job isn't disappearing everywhere. It is morphing into a job a junior cannot get.

Now the honest brake, stated at full strength as this series requires. Yale's Budget Lab finds no discernible AI disruption in aggregate US labour data. The sharpest critique of the Stanford result notes that exposed-occupation postings peaked in March 2022, six months before ChatGPT, implicating interest rates and post-COVID overhiring. The NY Fed attributes much of young-grad unemployment to remote work. Corporate AI-layoff claims keep getting walked back: Klarna rehired humans, IBM's "7,800 jobs" became "a couple hundred," Amazon de-attributed its own layoffs within 72 hours. All true. But notice what the counter-case defends and what it concedes. The aggregate is contested; the entry gate is not. Even the sceptics' own data shows the youngest cohort in the most exposed jobs doing worst, and every rival explanation merely changes who started the fire that AI is now visibly feeding.

On paper, India looks safe: the IMF puts AI exposure at around 26% of Indian employment versus 60% in advanced economies, because most Indians work on farms and construction sites AI cannot touch. But composition is destiny. The exposed slice is not a random 26%. It is the slice: the $315 billion IT-BPM sector, 5.95 million direct jobs, over half of India's record $387.5 billion in services exports, and, by a NASSCOM-CRISIL estimate, four further jobs in the wider economy for each one inside it. This is the machine that manufactured the Indian middle class: the EMI, the flight, the Whitefield flat. When Bernstein's analysts wrote an open letter to the Prime Minister in April 2026, this was the warning: 10 to 15 million IT-anchored livelihoods hold up the "aspirational middle class," and the window to protect the model is "narrow."

Worse, offshore work is not average white-collar work. It is white-collar work that was specifically selected, over thirty years, for being codifiable: decomposable into rulebooks, scripts, tickets and SLAs that could be shipped across an ocean. Codified is exactly what AI eats first. The IMF's study of the Philippines, the nearest comparator economy, found 89% of BPO workers highly exposed. The ILO warns that generative AI lets firms automate "tasks previously outsourced to lower-wage economies", a sentence that describes the outsourcing development path itself being switched off. Jefferies models Indian call centres losing half their revenue to AI within five years. NITI Aayog's own worst case has tech-services employment falling from roughly 8 million to 6 million by 2031. The market has already voted: the top five IT firms' combined value fell 46% between August 2024 and July 2026, and IT's weight in the Nifty 50 hit its lowest level since 2002.

And here is the cruellest line in the ledger: the crowd is accelerating toward the closing gate. B.Tech seats filled hit an eight-year high in 2024-25, with computer science the largest branch. Families are still pricing in a ladder that no longer exists at the advertised width, while graduates are already India's unemployment problem: youth make up around 83% of the unemployed, and the graduate unemployment rate runs near nine times that of workers who cannot read or write.

The government's own Economic Survey said the quiet part in print: much of the IT workforce sits in low-value-added services "particularly vulnerable to automation," and if worst-case projections materialise it could "set the country's economic growth trajectory off course."

Listen to the state think out loud. India's Chief Economic Adviser, in June 2026: the era when software and MBA education gave India its advantage "is over"; the future is "trade skills… As a chef, you cannot be replaced by AI." He is not wrong about the exposure. But when the chief economist of the country that built a quarter-trillion-dollar industry on code advises its graduates toward the kitchen, the message underneath is that the state has no plan for the ladder itself. The economists who study India most closely are blunter about what to fix. Raghuram Rajan and Rohit Lamba: "Coding sweatshops don't stand a chance against AI." The era of routine training is ending, and the answer is remaking higher education around judgment, not syntax. Rajan's optimistic corollary deserves equal billing: the $50,000 Indian engineer versus the $250,000 American one, both equipped with AI, is a differential that survives. The arbitrage endures if the skill is real. That conditional is the whole game.

Because the deep problem is not that AI destroys work. It is that AI destroyed the apprenticeship: the thirty-year-old deal in which a fresher's first two years of low-value work paid for their training. PwC's "seniorization" data and Anthropic's expertise curves say the same thing from opposite ends. The market now demands judgment at the gate, and judgment was the thing you used to acquire inside the gate. No individual firm has an incentive to fix this: training juniors is a public good the "job-ready candidate" era lets everyone free-ride on. Which is precisely the definition of a policy problem. The fixes follow from the diagnosis: make first jobs a measured policy object rather than a hoped-for by-product; pay for placements, not certificates (the CAG has shown what certificate-counting produces); use the state's own procurement, the same DPI stack this series has argued India should build on, to fund AI-era apprenticeships at GCC and startup scale, in the tier-2 cities where 93% of GCCs are not; and hold NASSCOM's members to publishing entry-level hiring the way they publish attrition, because a gate that closes in the dark closes faster.

And the series' own thesis closes the loop. The previous dossiers argued India should own the deflating, labour-touching layer of AI: cheap inference, Indic voice, the consumer surface. Rent the frontier, lose the margin; buy shares in champions, distort the market. That argument was about sovereignty. It turns out to be the jobs argument too. The services pyramid monetised Indian labour by the hour, which is exactly why AI guts it. A domestic consumer-AI economy, voice agents in twenty-two languages, AI-native games, creator tools, the "services-as-software" that Nilekani sizes at $400 billion, monetises what Indians build and own. And it hires at the bottom, because application layers always do. The country does not need its graduates to become chefs. It needs them to stop being the world's cheapest hour and start being the owners of the layer the previous two dossiers described. The gate to the old ladder is closed. The honest project is not prying it open. It is building the next ladder somewhere the rent isn't due in San Francisco.