The Reskilling Mirage: Where Scale Does Not Become Jobs

A decade of scale, counted plainly

The base numbers are not in dispute and they are large by any comparison. Across ten years to 2025, the National Skill Development Corporation and the Pradhan Mantri Kaushal Vikas Yojana together trained about 3.3 crore people, 33 million. PMKVY on its own trained or oriented around 1.6 crore people and certified about 1.29 crore of them. The physical footprint behind that number is real too: India runs 14,953 Industrial Training Institutes, a network built specifically to carry vocational training into districts a college system does not reach.

The funding line matches the ambition. The Cabinet approved a Restructured Skill India Programme with an outlay of Rs 8,800 crore, running from FY23 through FY26, to carry PMKVY and its adjacent schemes forward. A larger figure, Rs 12,000 crore, circulates in press coverage of a "PMKVY 4.0," but that number has not been independently confirmed against a Cabinet or ministry document; the verified, Cabinet-approved outlay for the restructured programme is Rs 8,800 crore, and that is the figure that should anchor any read of the programme's size.

None of this is a small effort. It is the scale a country attempts when it takes workforce transition seriously, and it is the correct starting point for the harder question the rest of this piece asks: what happened to the people who went through it.

The conversion problem, now audited

The Comptroller and Auditor General, India's constitutional audit body, examined PMKVY's placement record and tabled its findings in Parliament in December 2025 as Report No. 20 of 2025. The picture it drew was of conversion getting worse over time, not better. Placement rates ran at 18.4% under the scheme's first phase, 23.4% under its second, and fell to 10.1% under its third, a phase in which just 30,599 people were placed out of 3,99,860 who were certified. The CAG attributed the collapse to a mismatch between what the scheme trained people for and what the labour market was actually hiring, alongside weak planning and, in some cases, instances of fraudulent certification it flagged directly.

A separate Parliamentary panel review added a spending problem on top of the placement problem: PMKVY 4.0 had spent only about 14% of its allocation, and a related scheme, PM-SETU, had spent under 1%. A programme can under-spend and under-place in the same year, and that is what the panel found.

The headline placement rate for PMKVY, taken across its first three phases combined, is where the numbers stop agreeing with each other, and that disagreement is itself the finding worth reporting plainly. The skill ministry has cited a combined placement figure of around 41%. An independent analysis of the same period put the number closer to 22%. A separately circulated official figure has put it at about 54%. Three credible-sounding numbers for what should be one fact, spanning more than double, is not a rounding difference; it is the kind of gap that shows up when a programme is measured differently depending on who is doing the measuring, and it is why the CAG's own audited phase-by-phase figures, 18.4%, 23.4%, 10.1%, are the numbers this piece treats as the load-bearing ones.

The floor beneath AI skilling is moving, unevenly

Any AI-skilling claim sits on top of a more basic question: can the workforce use a computer at all. The National Sample Survey Office's Multiple Indicator Survey for 2020-21 found that only about 27% of Indian youth could send an email with an attachment, meaning roughly 73% could not perform what is, in most workplaces, a baseline task. That is a severe floor for any AI-adoption programme to build on, and it was the most recent government-collected national figure for years.

A more recent survey, run by the Observer Research Foundation in August 2025, found the opposite picture among youth: about 85% could send an email with an attachment. Read together, these are not two competing claims about the same present moment; they are two readings taken five years apart, using different survey instruments and different sample designs, of what may be a real and fast improvement, helped along by cheaper smartphones, expanded broadband, and five more years of school and college cohorts passing through a more digital curriculum. The floor the 2020-21 survey found should not be treated as still current in 2026. Equally, one 2025 survey should not be treated as a settled, cross-validated national baseline on its own. What can be said with confidence is that the gap narrowed materially over those five years; how far it narrowed, and where it sits today, is not yet pinned down by a single trusted instrument.

Wide but shallow: what "AI-skilled" actually measures

One step up from basic digital literacy sits the question the AI-jobs conversation actually cares about: how many working professionals can use, build, or ship AI systems, and at what depth. A 2026 industry study from NASSCOM found that more than 90% of early-career technology professionals now use AI tools in their work, a figure that on its own would suggest the workforce has already made the shift. The same study drew a sharper line one level down: only 23% qualified as "AI-native," meaning they could build and deploy AI systems independently, while 68% fell into a broader "AI-proficient" tier, defined as familiar with AI tools across a range of tasks but without that independent build-and-deploy capability.

Three other, separately run measurements land in the same neighbourhood. A figure cited by the Ministry of Electronics and Information Technology put the share of AI-skilled IT professionals at close to 16%. Scaler, an independent skilling platform, found about 19% of the professionals it surveyed were "AI-ready" by its own criteria. And a LinkedIn survey found 84% of professionals said they felt unprepared for the hiring shift AI is already causing in their industries, a self-reported anxiety figure rather than a skills test, but one that points the same direction as the harder numbers around it.

Four instruments, four different organisations, four different methodologies, and they converge on the same reading: using an AI tool has become close to universal among people whose jobs touch technology, but the deeper capability, building and deploying AI systems independently, sits somewhere between one-sixth and one-quarter of that same population. Wide adoption and shallow depth are not two separate stories. They are the same statistic, described from opposite ends.

The flagship programme built for exactly this gap, under-executed

India's most direct policy answer to the AI-native shortfall is the IndiaAI Mission, approved in 2024 with a five-year budget of Rs 10,371.92 crore. One of its named pillars, FutureSkills, exists specifically to build the deep AI talent the surveys above show is scarce: scholarships, fellowships, and structured pathways meant to grow the AI-native tier rather than the AI-proficient one.

A Parliamentary Standing Committee review found the pillar had, by its own government's figures, reached roughly 3% of its first-year fellowship targets. The gap is concrete rather than abstract: against a goal of 5,000 undergraduate scholars, 150 were actually selected; against a goal of 100 PhD fellowships, 3 were awarded. Of the mission's full Rs 10,372 crore allocation, about Rs 400 crore had been released, and around 32% of released funds had actually been spent, before the mission's budget was cut by roughly half.

This is the same conversion gap that runs through the placement figures earlier in this piece, showing up one rung higher on the skill ladder. A programme was designed and funded to close the AI-native shortfall the industry surveys measured. On its own government's audit, it missed its first-year targets by more than an order of magnitude, and its budget was then reduced rather than corrected. Building the pathway and walking people down it are, again, two different achievements, and the mission's first year delivered the first without yet delivering much of the second.

Training is not a severance policy

The private sector is where the reskilling numbers look genuinely strong, and that should be stated plainly rather than folded into the pattern above. NASSCOM reported that more than 2 million professionals were upskilled in AI during FY26, with 2 to 3 lakh of them trained in advanced AI specifically, a scale of corporate investment that is real and that the government schemes above have not matched. Tata Consultancy Services reskilled more than 300,000 of its own employees in AI and machine learning over the same period, a figure that stands on its own as a large, credible corporate training effort.

The same company, in the same year, cut about 12,200 jobs, roughly 2% of its workforce. This is not a contradiction in the data; it is the finding. A company can run one of the country's largest AI-reskilling programmes and still reduce headcount in the same period, because reskilling a workforce and protecting that workforce's jobs are two different variables that do not move together automatically. Treating training volume as a proxy for job security is exactly where the reskilling story becomes a mirage: the number that gets published is the training figure, because it is the flattering one, and the number that gets buried is the placement or retention figure that would show whether the training actually held a job in place or moved a person into a new one.

What sits ahead of this is a demand projection, and it should be read as a scenario rather than a forecast stated as fact: NASSCOM and Deloitte project India's AI-talent demand rising from roughly 600,000 to 650,000 professionals today to about 1.25 million by 2027, implying a shortfall of more than 600,000 people against that demand curve. Set beside the numbers earlier in this piece, that shortfall sits inside the same 23% AI-native tier the NASSCOM study measured and the IndiaAI Mission was built to grow. India's training volume, on paper, already dwarfs 600,000. What the evidence assembled here suggests is that volume and the specific, deep capability the 2027 demand curve requires are not the same thing, and the gap between them is where the country's reskilling effort is currently being tested.