Every government in the India–Gulf corridor is running a skills program, and most of them are counting the wrong thing. Learners enrolled, courses completed, certificates issued: these numbers rise reliably, cost little to produce and say nothing about whether anyone was hired, paid more or started a business. Meanwhile the target itself is moving — generative AI has, within three years, changed what an entry-level worker is expected to do — and programs designed for the last decade’s job descriptions are being scaled up for the next decade’s. This article sets out what the evidence says about skills programs that change outcomes, how AI alters the design brief, and what a ministry should specify when it commissions a future-skills program for a young, entrepreneurial economy.
The idea in brief. The evaluation literature on training and active labour-market programs is large and its lessons are consistent: short, generic, supply-driven training produces little; programs that build real human capital, involve employers directly, include work-based learning and are judged on placement and earnings over two to three years produce measurable, lasting gains — especially for those who start furthest behind. AI does not change those lessons; it changes the content and raises the value of two things — foundational skills that do not decay and the personal-initiative, entrepreneurial habits that let a worker use new tools before the curriculum catches up. A future-skills program should therefore be employer-anchored, behaviour-based, AI-native in delivery, and measured at 6, 12 and 24 months on the outcomes that matter.
The demand side: what the labour market is saying
The World Economic Forum’s Future of Jobs Report 2025, drawing on a survey of large employers worldwide, projects that roughly 170 million jobs will be created and about 92 million displaced by 2030 — a net gain of around 78 million — while close to two-fifths of existing skill sets are expected to be transformed or become outdated over the same period, and a majority of the workforce will need some form of training. The exact figures are employer projections rather than forecasts, but the structure of the message is what matters: net job growth alongside enormous churn in what jobs consist of.
The corridor’s two sides face that churn with very different demographics. India’s median age is in the late twenties and roughly two-thirds of its population is under thirty-five; its challenge is absorbing an enormous annual cohort of young workers into productive, formal employment, and its skills programs operate at a scale — tens of millions of trainees over a decade under national missions — that no other democracy attempts. The Gulf states have small national populations, historically absorbed into the public sector, and large expatriate workforces; their challenge is moving nationals into private-sector and entrepreneurial roles that their diversification visions require, which is why their skills agencies — Bahrain’s Tamkeen, Saudi Arabia’s Human Resources Development Fund and their peers — pair training with wage support and employer incentives rather than running training alone.
What both share is a demand-side signal that has been stable for a decade and is now sharpening: employers report shortages not primarily of credentials but of the ability to learn, communicate, solve problems and take initiative — the skills that David Deming’s 2017 analysis in the Quarterly Journal of Economics showed have been growing in labour-market value for a generation, and that AI is now making more valuable still.
What works: the evidence on training and labour-market programs
Skills policy is unusually well studied, because active labour-market programs have been evaluated with experimental and quasi-experimental designs across dozens of countries for decades.
The most comprehensive synthesis is David Card, Jochen Kluve and Andrea Weber’s 2018 meta-analysis in the Journal of the European Economic Association, covering hundreds of program estimates. Its findings are worth stating carefully. Average effects of active labour-market programs are small in the short run and larger two to three years after the program. Programs that build human capital — genuine training — have larger and longer-lasting effects than job-search assistance alone, but they take longer to show them. Effects are larger for women and for the long-term unemployed, and larger in recessions. Public-sector job-creation schemes show the weakest results. The message for a ministry is double-edged: training works, but not on a quarterly reporting cycle, and judging a program at six months will systematically under-count its value.
Developing-country evidence sharpens the picture. Orazio Attanasio, Adriana Kugler and Costas Meghir’s randomised evaluation of a Colombian vocational-training program for disadvantaged youth, published in the American Economic Journal: Applied Economics, found substantial gains in employment and earnings — particularly for women — from a program that combined classroom training with an on-the-job component in firms. Livia Alfonsi and colleagues’ 2020 experiment in Uganda, in Econometrica, compared vocational training with firm-provided apprenticeships and found both raised employment and earnings, with vocational training producing more portable, certifiable skills and firm-based training producing faster initial placement — a trade-off that program designers should choose deliberately rather than stumble into.
The design features that recur across the programs with lasting effects:
- Employer involvement in the curriculum and in delivery — the program teaches what a named set of employers will hire for.
- A work-based component — internships, apprenticeships or live projects, not simulations.
- Sufficient duration and intensity to build real capability, typically months rather than days.
- Soft skills taught explicitly — communication, teamwork, initiative — rather than assumed.
- Placement services and follow-up as part of the program, not an afterthought.
- Outcome measurement at 12 to 36 months on employment, earnings and retention.
Entrepreneurial skills as employability
The corridor’s governments increasingly frame skills programs around entrepreneurship, and the evidence supports a specific version of that framing. As article 31 discussed, the Togo experiment reported in Science in 2017 found that a psychology-based personal-initiative curriculum — proactive, self-starting, persistent behaviour — outperformed traditional business training substantially in raising firm profits. The same trait profile is what employers describe when they say they cannot find graduates who take ownership. Personal initiative is not a founder-only skill; it is the employability skill most in demand, and it responds to training.
That reframes the entrepreneurship component of a national skills program. Its purpose is not primarily to produce founders — most participants will be employees — but to build the habits of experimentation, customer thinking, ownership and resilience that make a worker valuable in a firm and, for a minority, capable of starting one. A skills program that includes a serious action-learning component (of the kind article 34 describes for campuses) produces a workforce that employers notice and a founder pipeline as a by-product. HexGn’s Future Proof program was designed on exactly this logic: six modules built around high-growth industries, microlearning that fits around study or work, and a career toolkit that treats entrepreneurial habits as employability rather than as an alternative to employment.
AI changes the target, not the method
Generative AI arrived in the middle of this policy conversation and has produced two reactions — that everything about skills must change, and that nothing has. The early evidence supports neither.
Three studies from 2023 set the terms. Shakked Noy and Whitney Zhang’s experiment in Science found that access to a large language model made professionals substantially faster at writing tasks and raised the quality of their output, with the largest gains for the weakest performers. Erik Brynjolfsson, Danielle Li and Lindsey Raymond’s field study of customer-support agents, published as an NBER working paper, found productivity gains averaging around 14 per cent from an AI assistant — concentrated among novice and lower-skilled workers, who effectively absorbed the tacit knowledge of experienced colleagues through the tool. Fabrizio Dell’Acqua and colleagues’ study with consultants, published through Harvard Business School, found large gains in speed and quality on tasks inside the tool’s capability frontier and worse performance on tasks just outside it — the “jagged frontier” that makes judgement about when to trust the tool a skill in itself.
For skills policy, three implications follow, and they are consistent with the older evidence rather than contradicting it. First, the novice uplift compresses onboarding: AI tools narrow the gap between new and experienced workers in many tasks, which means entry-level programs can target higher-value work sooner — and that the value of getting people into work quickly rises. Second, foundational skills matter more, not less: judgement, domain knowledge, communication and the ability to evaluate a tool’s output are what separate a worker who is amplified by AI from one who is replaced by it. Third, tool-specific skills decay faster than ever, which shifts the curriculum from mastering a particular tool to learning how to learn tools — the personal-initiative habit again, in a new costume. Article 30 in this series examines the same shift from the employer’s side.
