Ruchi Gupta

Abstract

India's central risk from artificial intelligence is not mass technological unemployment but increased productivity without labour absorption: firms becoming more efficient while the economy loses the institutions through which young people enter work, acquire judgment, and move into economic citizenship. India may be less exposed than advanced economies but more vulnerable because AI threatens its development model of services exports and entry-level cognitive work. This essay examines three channels of disruption: the contraction of entry-level work, threats to outsourced employment from AI and geopolitics, and the effects of algorithmic management and data extraction on physical and informal workers.  It argues for a predistributive response centred on a national AI and Work Observatory, rebuilt pathways into work and worker protections, and permanent parliamentary oversight. These measures must be accompanied by a growth strategy that creates labour-absorbing demand.

Introduction

Artificial intelligence is neither simply a threat to employment nor a guaranteed engine of prosperity; its consequences depend on how it is developed, deployed, and governed. For India, the central risk is productivity without labour absorption, with firms becoming more efficient while the economy becomes less capable of bringing young people into stable work, teaching them professional judgment, and creating credible routes into the middle class.

The IMF estimates that roughly 60 per cent of employment in advanced economies is exposed to AI. Among the major emerging markets it profiles, India’s high-exposure share, at about 26 per cent, is the lowest. But exposure is different from vulnerability. Advanced economies have older populations, greater fiscal capacity, deeper welfare systems and control over the firms driving the transition. India has a younger workforce in need of jobs, an overwhelmingly informal labour market, thin transition institutions, and a growth path unusually dependent on services exports and entry-level professional employment. The argument is thus that a targeted shock could carry greater impact, because it strikes precisely the pathways on which India depends for labour absorption and social mobility.

The deeper risk is not to particular firms or workers but to the development model itself. India's path into economic citizenship has run through a narrow channel – services exports and entry-level cognitive work – and that is the channel AI compresses first. For a young, labour-surplus democracy, a shock to the pathway of mobility is not a sectoral problem but a systemic one.

The disruption operates through three channels and timelines. The contraction of entry-level cognitive work is underway: generative systems now produce the drafts, code, and analysis previously assigned to juniors, thinning both hiring and the apprenticeship pipeline through which judgment is formed. The export-demand channel is also gathering pace; the work need not move to a cheaper country when it can move in-house to the model. And the channel running through physical and informal work operates on two clocks: algorithmic management of platform work is already pervasive and physical replacement remains distant in India, but the harvesting of the data that will train it has begun on Indian factory floors.

The same capabilities offer gains India must capitalise on: wider access to scarce expertise and lower organisational costs for small firms and individual entrepreneurs. Upside and downside run through the same mechanisms; the task is to shape the terms of adoption.

India is alive to the potential of AI to reshape its economy and its governance, and a range of institutions and initiatives has grown up around it. This is welcome, but two things are missing: sustained attention to what AI is doing to work, and democratic representation in how the transition is governed, anchored in a dedicated parliamentary body.

This essay proposes three foundations for that response: seeing the transition clearly; rebuilding the pathway into work and protecting existing workers; and governing the transition democratically. These would have to be supported by measures to spread adoption and to carry the argument outward through a Global South compact.

Work as Democratic Infrastructure

Debates about AI and employment often centre on the number of jobs that may disappear. That is a central concern, but not the only one. Work distributes income, teaches competence, establishes status, creates relationships, and gives individuals independence from both family and state; for young people in particular, employment is the transition into full economic and civic citizenship.

Entry-level roles occupy a special place in this process. The first three to five years of working life in an occupation – whether as junior researcher and analyst, trainee nurse, legal assistant, or fresher developer – constitute the period in which workers hold positions designed to be supervised. They do not begin with mature judgment; they acquire it by producing work, making errors, receiving corrections, and observing how institutions operate. Routine work can be tedious, but it is where tacit knowledge is formed.

AI compresses the distance between production and review. Organisations once needed many junior workers to produce drafts, research, routine code, and preliminary analysis, and fewer experienced people to assess and correct it. When the production layer becomes cheap and fast, the review layer can operate with fewer people beneath it. The ladder does not merely become shorter; its lower rungs may be removed while the upper rungs remain.

When AI makes it cheaper and faster to produce work without juniors altogether, every firm can rationally thin its junior layer, while the economy loses the institutional capacity through which inexperienced workers become experienced. This logic is well documented: firms under-invest in training because they cannot retain the benefits if the trained worker leaves, and so it can be rational to free-ride on workers trained elsewhere. The problem gains new force from a technology that removes the very junior tasks through which such training was informally financed. No single firm behaves irrationally, but the overall effect is a market failure in worker training. This provides the economic, not merely democratic, justification for the interventions proposed here.

There are democratic stakes too. The democratic bargain is not a formal promise of equality but the expectation that citizenship comes with a credible route to economic agency: citizens tolerate differences in wealth and status when they believe mobility remains possible and effort can improve their circumstances. If large groups come to believe they are economically redundant, the legitimacy of markets and democratic institutions alike weakens.

AI could intensify India’s existing weaknesses – those of inadequate job creation, informality, and gender exclusion – by making it easier for firms to grow without employing proportionately more people. The relevant policy question is not only whether AI raises output, but whether it expands the number of people who can acquire skill, develop judgment, and participate in economic life.

Less Exposed, More Vulnerable

The strongest objection to an India-centred argument is that advanced economies appear more exposed to AI, with their larger shares of professional and cognitive employment. The IMF’s analysis bears this out: roughly 60 per cent of employment in advanced economies is exposed, against 40 per cent in emerging markets. Among the emerging markets it profiles, India’s high-exposure share of about 26 per cent is the lowest. But this framing sidesteps the developmental question. Exposure measures the proportion of tasks technology may affect; vulnerability concerns the consequences. A country can be less exposed in aggregate yet more vulnerable if the disruption concentrates in economically and politically central sectors.

India’s vulnerability has four sources.

First, it is demographically young. A large young population becomes a dividend only when the economy can bring people into productive work; blocked entry converts the same structure into frustration, dependency, and political instability.

Second, India’s post-liberalisation mobility story has depended disproportionately on services. This was structural: late developers now tend to deindustrialise prematurely, which is why India’s escalator into the middle class ran through IT services and professional occupations rather than factories. These sectors do not employ most Indians, but they have played an outsized role in the creation of the Indian middle class, urban mobility, aspiration, and export earnings. An AI shock concentrated here strikes the narrow channel through which structural transformation was occurring in India.

Third, India’s transition institutions are weak, and its labour absorption problem predates AI. The “jobless growth” literature documents two decades in which output expanded far faster than decent employment; roughly nine in ten workers are informally employed. AI threatens to intensify India’s existing absorption problem.

Fourth, India has limited control over the frontier: advanced compute, model development, cloud infrastructure and commercial decision-making remain concentrated elsewhere.

The question, then, is not how many tasks AI will automate but what it does to India’s development model itself. India’s mobility has depended on a narrow set of pathways; a technology that strikes those pathways threatens not a slice of the economy but the mechanism by which the country has turned growth into opportunity. That is the argument the rest of this essay develops.

Three Channels of Labour-Market Disruption

The three channels operate on different timelines. First, the compression of entry-level cognitive work is underway, with the first Indian evidence visible. Second, the export-demand channel is a medium-term risk whose pace depends on how global clients reorganise. Third, the disruption of physical and informal work runs on two clocks – algorithmic management is already the governing structure of platform employment while physical replacement remains the most distant, even as the collection of the data that will enable it has begun.

The Entry-Level Ladder

Emerging US evidence suggests early-career employment has weakened in highly AI-exposed occupations, particularly where systems automate rather than complement: researchers have found a marked relative decline among workers aged 22 to 25 in the most exposed occupations, and the IMF has called reduced entry-level hiring a growing concern. These findings are contested – critics note the difficulty of separating AI effects from the post-2022 tech-hiring correction – but this essay’s argument rests not on estimates but on the mechanism. Junior work has traditionally consisted of tasks easy to specify and review, including initial drafting, document checking, routine coding, standard queries, and summarisation. These are among the tasks on which generative systems have improved most rapidly. 

Nor is the evidence confined to the United States. U.S. companies attributed about 55,000 announced job cuts to AI in 2025, roughly fourteen times the 2023 figure. India’s top five IT firms shed nearly 70,000 positions in a single year. Tata Consultancy Services alone cut over 12,000 jobs in FY26 even while publicising large-scale AI deployment contracts. Fresher hiring, which peaked at around 600,000 in FY22, fell by 80 per cent to about 120,000 by FY25. This shows that displacement operates through both slower hiring and separations. The damage therefore concentrates on those trying to enter, which explains why, as the essay later argues, a transition system triggered by formal retrenchment will miss it. Whether these cuts are “caused by AI” in a strict sense is examined in the next section; what is not in dispute is that the firms that built India’s largest professional apprenticeship system are shrinking its intake.

The optimistic outlook is that AI will enable less-experienced workers to perform at a higher level: it has been argued that well-designed AI could widen access to expertise and thus help rebuild the middle class. That possibility should be taken seriously, with policy implications adapted to the Indian context.

Individual capability and aggregate hiring are two separate issues, however. If one junior equipped with AI performs the work of three, the one who retains the position gains while two lose the opportunity to enter the profession. The distributional question is not whether AI makes individual workers more productive, but how many people retain access to the institutional pipeline in which capability is developed and honed. The answer is not to protect yesterday’s routine tasks but to reconstruct the training function they performed: supervised responsibility, real clients and real-world problems, room to make and correct mistakes, and a route from assisted production to independent judgment.

The Export-Demand Channel

India’s labour-arbitrage model rests on foreign demand for Indian cognitive labour. A bank or retailer that once outsourced coding, testing, support or document processing to India may, as AI improves, perform that work internally with a smaller team. The work need not move to a cheaper country. It can move to the model.

What happens next depends on how global clients reorganise and whether Indian firms move from execution to higher-value capability. NITI Aayog’s scenarios illustrate the stakes: under adverse conditions India’s technology-services workforce could contract from roughly 7.5–8 million today to about 6 million by 2031; under a favourable scenario it could expand to about ten million.

India’s more than 1,700 Global Capability Centres, through which foreign firms employ Indian workers directly, may mitigate some of this impact. However, there are two vulnerabilities here too. First, GCCs retain Indian employment largely as an execution layer while capability, intellectual property, and decision rights concentrate in the foreign parent, and an execution layer is precisely what AI compresses first. Second, a political coalition is hardening in the United States around discouraging offshored labour and services, including models like the GCCs.

The same logic extends to the export of workers themselves. India has relied on sending its knowledge workers abroad. AI-driven contraction in demand for imported cognitive labour, compounded by hardening visa politics, points to a fall in foreign demand for Indian labour.

The way forward thus requires redirecting expertise toward domestic developmental priorities through appropriate public investment and policy.

Physical and informal work

The third channel concerns the large majority outside formal professional occupations. Agriculture still employs roughly 43 per cent of India’s workforce, and the question AI poses is whether the pathways out of agriculture will remain open and for how long.

Platform workers are already subject to algorithmic management: software assigns work, calculates pay, sets metrics, penalises and deactivates. In theory, such systems can reduce favouritism and improve allocation, but they have often made employment more arbitrary through opaque rules and the absence of workers’ right to human review. The EU’s Platform Work Directive offers a precedent India should incorporate: consequential employment decisions made through software must be reviewable.

Manual workers are also entering the AI economy as producers of training data. Multiple news reports from Indian factories have documented workers wearing cameras while sewing, folding, sorting and assembling; The resulting first-person footage – egocentric data – teaches robotics systems how human hands perform physical tasks. In some cases, workers were neither separately compensated nor fully informed.

Automation is already replacing physical labour across the developed world and in China, from automated checkouts and factory robots to self-driving cars. Low wages may postpone substitution in India for some years. However, workers can be simultaneously protected from replacement by their low wages and selected as low-cost sources of the data required to build future replacement systems. The worker receives wages for present labour, and the firm acquires long-term option value. Consent here cannot be reduced to a signed form. A worker may agree to recording without understanding what the data will train or whether refusal would affect employment. It is therefore crucial that data governance accounts for bargaining power, not merely formal permission. Whether, and on what terms, such replacement-training data may be collected at all is a question we must debate collectively.

Gender Across all Three Channels

AI’s effects are likely to hit women workers more. Clerical and administrative occupations have the highest generative-AI exposure of any category, and women’s employment is concentrated in exactly these roles. Women hold a substantial share of India’s business-process and support employment – among the few jobs offering educated women in smaller cities formal wages, transport and social legitimacy as workers. A demand collapse removes one of the few sanctioned routes into the labour force. Consequently, any transition policy must disaggregate by gender at every step.

Productivity Without Absorption

The central economic danger is that productivity gains may be captured without increasing employment, wages, or domestic demand. For a firm, producing more with fewer workers can be rational. For a labour-surplus economy, the same outcome may be destabilising: if output rises while labour income stagnates, gains accrue to capital owners and specialised workers, and since richer households spend a smaller share of additional income, aggregate demand may weaken even as measured productivity rises. This is productivity without absorption, wherein an economy becomes technically more efficient while becoming less effective at incorporating people into productive work.

The dynamic is supported by a developed body of literature. Acemoglu and Restrepo distinguish automation that creates new tasks for labour from “so-so automation” – good enough to displace workers, but not good enough to generate the productivity that would create new demand for them – and Acemoglu and Johnson argue that technology raises broad living standards only when institutions force a direction of innovation that complements rather than replaces labour. The stakes of that directional choice are highest in a labour-surplus democracy, where the surplus is young and the legitimacy of the political order rests on absorbing it.

Evidence on AI productivity remains mixed. A 2025 METR trial even found experienced developers slower with then-current tools, and benchmark performance cannot be treated as proof of organisational value. In early 2026, the same researchers revisited the question and concluded that AI tools probably accelerated developers more than they did a year earlier, though the size of the gain remains uncertain.

This evidentiary gap produces an apparent contradiction: firms announce AI-driven workforce reductions faster than rigorous studies can find the productivity gains that would justify them. A meta-analysis pooling 371 estimates found no robust link between AI adoption and aggregate labour-market outcomes, and Oxford Economics has called AI’s role in recent cuts overstated. Part of the resolution is AI-washing: “AI” has become a face-saving label under which companies execute corrections for pandemic-era over-hiring and ordinary margin defence. This reads to investors as transformation rather than distress, and one that capital markets reward, so the narrative itself reshapes labour markets whether or not the capability exists. The implication cuts both ways. AI-attributed displacement statistics overstate the technology’s current power, but firms’ willingness to invoke AI when shedding workers reveals what they expect and intend, and normalises labour-light growth before the technology has earned it.

The strongest rejoinder is the J-curve: general-purpose technologies from electricity to computing depressed measured productivity during adoption before yielding large gains. The argument deserves engagement, but it smuggles in an assumption that the trough will be managed. Historical J-curves unfolded within nation-states with labour institutions, welfare systems, and decades of adjustment time; this transition is faster, transnational, and governed by fewer actors. In fact, concerns about an economic transformation “larger than the Industrial Revolution but unfolding over a vastly shorter time frame [leading to] large-scale job displacement” have prompted more than 200 economists, AI researchers, and Nobel laureates to sign a joint statement to “act now.” The worry is that without deliberate intervention the trough may not be a trough at all but a permanent reallocation of returns from labour to capital. For a labour-surplus democracy, a decade-long trough may mean a wasted generation, not just lower growth. 

Individual acceleration and social benefit are distinct: policy should favour technological uses that raise worker capability or expand access and be wary of subsidising adoption whose principal effect is to transfer income from workers to capital.

None of this is an argument for preserving inefficient employment. It is an argument for measuring efficiency correctly. A technology that lowers one firm’s wage bill while weakening skills formation and transferring transition costs to households and the state may be privately efficient and socially costly.

Where the Unabsorbed Go

An absorption failure does not appear as mass unemployment – India’s workers cannot afford to be unemployed – but as displacement into lower-quality work. The destinations are already visible.

Over the past eight years the shift out of agriculture has essentially stalled. The farm’s share of employment sits almost exactly where it was in 2017–18, around 44 per cent, having risen during the pandemic rather than continuing its long decline. Additionally, the work the non-farm economy added was not, for the most part, regular salaried employment. The salaried share has, if anything, slipped. While a growing share of workers has been absorbed into self-employment, much of it is low-earning own-account work and unpaid labour in family enterprises. Growth, in other words, has not been translating into better jobs.

The second destination is the gig economy, the de facto shock absorber for (semi-)educated young men in Indian cities, where work is available immediately, algorithmically managed, and disconnected from any ladder into skill or career progression. Official estimates put the gig workforce at 77 lakh (7.7 million) in 2020–21 and project it to reach 2.35 crore (23.5 million) by 2029–30, indicating an increase from about 1.5 per cent of the workforce to roughly 4 per cent.

The third is the government-examination economy, where blocked absorption becomes politically legible. The aspiration for a government job is linked to the scarcity of adequate well-paying private jobs – lakhs of applicants for a few thousand posts, sustained by a coaching industry that absorbs years of young lives and family savings in a lottery with dwindling odds. The recurring protests over examination leaks and recruitment processes are a reaction to rationed entry-level opportunity. This is the democratic-bargain argument operating in real time. If AI narrows private pathways further, pressure on this valve will grow.

These destinations are the counterfactuals policy must weigh. The question is never that of “AI adoption versus the status quo” but AI adoption governed well versus a labour surplus draining into disguised underemployment, algorithmic precarity and examination queues.

The Upside India Cannot Afford to Miss: Labour-Complementing AI

A purely defensive response to AI would be a mistake. India has too many shortages to ignore AI’s productive potential. In a country short of teachers, doctors, nurses, lawyers, extension workers and local administrators, AI can augment scarce human capacity; it can help a nurse triage, a teacher personalise, a farmer access advice, a panchayat official understand government schemes, and a citizen understand a government form in her own language. AI’s first democratic promise for India lies in expanding access to expertise that has long been scarce, expensive, urban, or English-mediated.

AI also lowers the cost of organisational infrastructure – a small team can now do what once required separate support functions. This matters where most Indians actually work: MSMEs, civil-society organisations, local governments, and first-generation entrepreneurs. New forms of work will emerge around deployment, evaluation and data stewardship; but new work does not automatically reach those displaced from old work, and the test is whether it arrives before existing ladders collapse.

Access to easy generation, however, is not access to value. AI lowers the cost of producing output; it does not comparably lower the cost of identifying worthwhile problems, earning trust, reaching users or obtaining finance. A 2026 study of more than 160,000 start-up launches found that generative AI produced a sharp rise in solo entrepreneurship, but solo founders did not become more prevalent among the highest-performing ventures. AI is so far increasing the number of people who can attempt a business more clearly than the number who build exceptional ones. The missing public infrastructure is not just compute but judgment. Policy should democratise the full chain from capability to value – tools, knowledge, validation, mentoring, finance, users, and procurement.

Who Governs the Transition: Power, Sovereignty, and India’s Institutional Landscape 

Three Concentrations of Power

The labour-market problem is intensified by the concentration of power over AI at three levels.

The first is corporate. A handful of companies determine the capabilities, prices, access conditions and safety practices of the most advanced systems, and their product decisions can restructure work across many countries before national institutions have accepted or even debated the consequences.

The second is jurisdictional. In June 2026, the US government directed Anthropic to suspend access to advanced models for foreign nationals; Anthropic temporarily disabled the models for all customers, restoring access to a substantially attenuated model subsequently. Similarly OpenAI, at the US government’s request, released its most capable models only to a small group of “trusted partners” vetted by Washington, staggering their wider release under a new executive framework that lets the government review frontier models before they reach the public. This demonstrated that access outside the United States can be interrupted by an American political decision unrelated to the priorities of countries relying on the systems. Alongside continuing chip export controls, it establishes a broader reality: compute and frontier models are an instrument of national-security and competition policy.

The third is domestic. Workers, students and small firms have little influence over decisions to introduce monitoring, automate services, or collect data. Those most exposed to the consequences are least likely to participate in deployment decisions.

Sovereignty and India’s Institutional Landscape

The diffusion of open-weight models and falling inference costs reduces some dependence, and India should welcome it, but it comes with its own caveats. The open weights Indian firms increasingly adopt are, to a substantial degree, Chinese; the DeepSeek and Qwen families and their successors. While reliance on American frontier models carries export-control and pricing risk, reliance on Chinese open weights carries different questions of security review, data practice, and geopolitical exposure.

Sovereignty requires the practical capacity to choose among suppliers, evaluate claims independently, maintain critical public systems during external disruptions, and set domestic conditions for use.

India is alive to the potential of AI to reshape its economy and its governance, and a range of institutions and initiatives has grown up around it. The IndiaAI Mission, approved in 2024 with an outlay of over INR 10,000 crore, supports compute, foundation models, datasets, applications and skilling. The AI Governance Guidelines of November 2025 set out a principles-based framework for its use. An inter-ministerial AI Governance and Economic Group, constituted in 2026, coordinates policy across departments and is tasked, among other things, with examining AI’s effect on jobs. A parliamentary standing committee has reported on the impact of AI. Alongside these sit the Digital Personal Data Protection Act, 2023, whose rules were notified in November 2025, and the labour codes, in force since November 2025, which recognise gig and platform workers for the first time and create a worker re-skilling fund. Behind them all stands the Digital Public Infrastructure model – the government’s central narrative of technology-led inclusion, and a set of rails on which several of the proposals below could run. The proposals in this essay are designed to work through these institutions rather than around them; each has a natural home in one or more of them.

At the same time, two central gaps remain: legislative and democratic.

First, the data protection framework has a workplace lacuna. It leaves unresolved whether consent within an employment relationship, where refusal carries consequences, can be meaningful; whether the egocentric video of a worker’s hands recorded to train the very systems that may one day replace her is her personal data; and whether workers hold any claim on the value of the training data derived from them. Similarly, the labour codes are silent on the mechanism by which platform work is actually controlled – algorithmic allocation, rating-based discipline, automated deactivation. Whether copyrighted work may be used to train a model at all is contested before the courts. It is also the subject of a proposed compulsory-royalty licence, which, notably, accepts that those whose work trains a model have a claim on the value it creates but does not give them the option to opt out and determine that value independently.

The second gap is democratic. Because AI strikes the development model itself, the questions it raises are not technical settings for ministries and firms to adjust but choices about the country’s economic direction. These questions belong in Parliament. The deliberation so far has been framed around industry and technology rather than youth and labour; the parliamentary report on AI drew heavily on the sector’s own optimistic employment projections.

Recommendations: Three Foundations

India must judge its AI strategy by how many people the economy absorbs into increasingly productive work and not by the productivity gains of its largest firms.

The objective of democratic policy should be to ensure that productivity gains from AI are widely shared, to preserve credible routes into capability and responsibility for young workers, to protect existing workers against opaque decisions and abrupt losses, and to give Indian institutions the capacity to evaluate and govern the systems reshaping the economy.

The scale of the transformation demands a coherent response rather than fragmented interventions. And given India’s status as the world’s largest democracy with the world’s largest youth population, the response must show that innovation and equitable development can proceed together.

The frontier firms’ own answers to this problem do not translate well in India. In the advanced economies the distributional question is starting to be asked as one of sharing the gains directly: proposals to hand the public a slice of the leading AI firms – a few per cent of the equity – or to fund a universal basic income out of the wealth they generate. Neither is available to India, since the companies whose value those schemes would tap are American and Chinese, and a basic income at any meaningful level is beyond India’s fiscal capacity.

Consequently, India’s answer has to be structural. Where the frontier can contemplate sharing the proceeds after the fact, India has to shape how growth is produced and spread in the first place, so that it is dispersed across regions and firms rather than concentrated, and complements labour rather than replacing it. It also has to expand and remake the fragmented schemes it already runs into a social security floor strong enough to hold the economy steady as it changes. For India, distribution cannot be a dividend added at the end; it has to be built into how growth is generated.

This is what Jacob Hacker called predistribution, and what Acemoglu and Johnson describe as steering the direction of technology. For India, however, it is an imperative rather than a mere preference: the redistributive instruments available to the frontier are structurally unavailable to it.

One objection to this agenda is state capacity. But Indian state capacity is uneven rather than uniformly weak. The question is not whether the Indian state can execute, but whether it will point that capacity at the labour-market transition. This is why India’s response to AI is a political and democratic question, not merely a developmental one.

What follows rests on three foundations, in order of priority. Without the first, which is the capacity to see what AI is doing to work, everything else is guesswork. The second rebuilds the pathway into work that AI most directly threatens and protects those already on it. The third gives the transition a democratic home, so that these become choices a polity makes rather than outcomes a market imposes. The remaining measures – spreading adoption and carrying the argument to the Global South – support these three.

Foundation 1: Observe the Transition and Make Adoption Answerable

India cannot govern a labour-market transition that it does not track. Existing statistics are poorly suited to detecting the forms of disruption discussed in this essay: fewer entry-level vacancies, slower fresher recruitment, declining hours or wages, movement into lower-quality work, and the growing use of algorithmic systems to allocate, monitor, and terminate work.

The first requirement is therefore a national AI and Work Observatory. It should integrate official labour-force and payroll data with sectoral hiring information, employer disclosures, apprenticeship outcomes, and carefully validated indicators from private recruitment platforms. Its first task should be a baseline of entry-level hiring, wages, retention, and progression across sectors and states followed by regular indicators disaggregated by age, sex, region, and employment status, including informal and platform work. Its purpose is not to declare whether AI caused every observed change, but to identify where conditions are deteriorating enough to require investigation or response.

A White Paper on AI and Work should be completed within the first six months and tabled in Parliament. It should map exposure and vulnerability, identify legislative gaps, cost of the principal interventions, and establish the thresholds that would trigger enhanced monitoring or transition support. It should present alternative scenarios rather than a single forecast. Developed through public consultation and tabled in Parliament, it should provide the shared evidence base and language needed for coordinated national action.

Firms above a certain employee threshold should be subject to mandatory disclosures regarding tasks changed, the expected and actual effects on recruitment and existing workers, and the evidence behind claimed productivity gains. The purpose is not a veto over reorganisation but to make large employment claims answerable to workers, investors and the public.

Foundation 2: Rebuild Pathways into Work and Protect Existing Workers

India should establish sectoral AI-era apprenticeship compacts, beginning where entry-level work is already being reorganised and employers can jointly define credible routes to competence: IT services, financial and business operations, legal services, health administration, and selected public services are plausible starting points.

Apprentices should work with AI rather than be trained for a workplace in which it does not exist. But competence should be measured by domain knowledge, verification, judgment and the ability to recognise when automated output is unreliable, and developed through real work and progressive responsibility rather than simulated exercises or generic certification. Vocational training must be redesigned in partnership with employers to focus on durable capabilities rather than narrow technical tasks likely to be automated.

The collective-action problem identified earlier in this essay requires pooled rather than purely firm-level solutions. Sectoral financing arrangements, employer contributions, and public support can spread the cost. Major public contracts and incentives should reward firms that maintain credible training and progression pipelines.

Safeguards are needed to prevent apprenticeship from becoming a label for cheap temporary labour. Standards should specify supervision, compensation, learning content and progression, and public funding should follow verified outcomes – placement, wages, retention – not enrollment or certification.

Education reform and skilling should proceed on two timelines. Assessment can change quickly: greater weight should be put on oral defence, source verification, problem definition and the ability to explain and challenge AI-generated work. Foundational literacy, numeracy, reasoning, and domain knowledge should be the longer-term priority – people cannot verify output in subjects they do not understand – and they are not achievable through short AI courses. Entry pathways should also be designed to include women, smaller cities, and people outside elite institutions.

Protect Existing Workers

The most immediate regulatory gap concerns workplace decisions insulated from scrutiny, whether through automation or unilateral management. Workers should know when systems are being used to make consequential decisions about recruitment, allocation of work, pay, performance, discipline, suspension, or termination. They should receive understandable reasons and have access to meaningful human review.

No worker should be dismissed, penalised, or deactivated solely through an unreviewable automated process. These protections are particularly urgent for platform workers, for whom algorithmic management is already the governing structure of employment.

Firms should also be required to consult workers or their representatives before high-impact workplace deployments, particularly where systems affect monitoring, pay, allocation, recruitment, or termination.

Protection also means income security, and here India’s task is less to invent than to deliver. The Code on Social Security already promises gig and platform workers accident, health, maternity, and old-age cover, financed by contributions from the aggregators that engage them. The promise needs delivery on the ground: a contribution set high enough to fund real benefits, the cover the law already provides is actually activated, and protection made portable across platforms and states so that a worker does not lose it by changing how she earns. Portable social protection should be developed in parallel, treated as structural reform rather than a near-term deliverable.

Where AI produces sustained and measurable gains, workers should have credible routes to share in them. The relevant test is not whether employment rises in exact proportion to output, but whether productivity gains accrue exclusively to owners and a narrow group of specialists while transition costs are transferred to households and the state.

Worker-generated training data requires a more specific framework than ordinary consumer consent. Employers should disclose what is recorded, what systems it will train, how long it is retained, and with whom it is shared; refusal should not attract retaliation where collection is unnecessary to the present role; and where data production is a distinct part of the job, additional compensation or collective negotiation is appropriate. Whether training data whose sole purpose is to replace the worker altogether should be permitted, at least for the time being, is a separate question.

Foundation 3: Govern the Transition Democratically

The third foundation is democratic governance for the transition. This requires two things: a federal architecture and a permanent parliamentary home. 

The architecture should be federal but coordinated. The Centre should establish a national floor of workplace rights, reporting standards, portability and apprenticeship quality; states should experiment above it according to their economic structure.

A permanent Parliamentary Committee on AI, Work, and Democratic Transition should review the observatory’s findings, examine major public deployments and hold regular public hearings in different regions with workers, platform workers, women’s groups, employers, states, young people, and MSMEs instead of a single national consultation. Its purpose is not to remove technological choices from political contests but to give the contest a constitutional forum, a shared factual base, and a durable minimum consensus that survives changes in government.

Supporting measures: Spread productive adoption and create demand for labour

The foundations govern the transition; they do not by themselves create the work. That is the growth-side task.

India should support uses of AI that expand access and raise the capability of workers and smaller organisations. Public-interest missions in health, education, agriculture, and local administration should begin as evaluated pilots. Success should be measured through service reach, quality, error rates, worker capability, and distributional effects, not staff reduction alone.

MSMEs need practical infrastructure rather than appeals to adopt AI: shared tools, Indian-language interfaces, custom workflows, sector-specific resources, independent evaluation, mentors, buyers, and public procurement. India also needs public and university research, language resources, shared compute, technical expertise within government, and institutions capable of auditing systems used in workplaces and public services. Critical systems should not depend entirely on one supplier or jurisdiction.

Public procurement should reward systems that expand access, permit meaningful human review and meet interoperability and data-protection standards. It should not require contractors to preserve fixed headcounts, which would reward artificial labour intensity. Where a deployment substantially reorganises work, the contract should instead require consultation, training, and transparent reporting.

But AI governance cannot substitute an employment strategy. Apprenticeships do not create demand for labour; they improve entry into work that exists. A serious absorption strategy must expand the sectors in which India has unmet needs and where technology can complement labour: care, health, education, housing, construction, urban services, tourism, logistics, climate adaptation, labour-intensive manufacturing, and the formalisation of small enterprises. The direction of public investment matters as much as the design of AI regulation.

Supporting Measure: A Global Compact for AI, Work, and Shared Prosperity

AI systems are developed in one jurisdiction, trained on data and labour drawn from many others, and capable of altering employment across borders with no corresponding system of accountability.

India should begin by building a focused Global South coalition around labour-impact reporting, algorithmic management, independent evaluation, and consultation when the technological or sourcing decisions of global firms affect employment elsewhere. Countries that contribute labour, languages, data, and markets should not remain rule-takers in a system built partly from their resources.

India’s argument should not be defensive: the Global South is a potential producer, adaptor and governor of AI, not merely a site of displacement or a source of training data.

Conclusion

India's central AI risk is that firms become more productive while the economy becomes less capable of absorbing people, developing their judgment, and connecting greater capability to greater security and voice. That outcome would look efficient at the level of the firm and corrosive at the level of society.

India stands at a crossroads. AI can widen access to expertise, lower the cost of organisation, strengthen small firms and improve public services. But upside and downside run through the same mechanisms. A tool that helps one junior perform at a higher level may also reduce the number of juniors hired; an automated system that may improve efficiency may also become arbitrary and make management unanswerable; data that raises the capability of a robot may also be extracted from workers who receive none of its future value.

The democratic bargain requires a credible route into productive work, a reasonable floor of security and a voice over consequential decisions. Rebuilding that bargain means steering technology for mass benefit, treating entry-level work as an apprenticeship institution, sharing productivity gains, protecting workers against opaque systems and directing public investment towards labour-absorbing demand.

India can pursue cheaper production while allowing its entry ladders to decay. Or it can use AI to widen access to expertise, distribute growth across regions and classes, and proactively build the pipelines and institutions that turn inexperienced citizens into capable workers. The question is not whether India will adopt AI. It is whether the pursuit of productivity will be allowed to sever its link with capability, labour income and economic citizenship.