AI, Work,
and India's
Next Bargain
Press P to switch into slide-show mode. Press P again, or Esc, to return to the single continuous reading version.
India is vulnerable, but has optionality.
AI is best understood as a general purpose technology: it will not simply "take jobs" or "create jobs." It will force a search for new patterns of production, new contracts, and new ladders into skilled work.
The first shock lands where India has proudly built capability.
Code, back-office work, customer support, analytics, content, testing, and entry-level professional services are unusually exposed to software that can read, write, summarize, classify, and act.
The next gains come from faster discovery.
English-language talent, an IT services base, digital public infrastructure, frugal firms, and a large domestic market give India many places to try new combinations.
Sources: Bresnahan and Trajtenberg on general purpose technologies; Kling on PSST; Anthropic Economic Index; ILO, India Employment Report 2024.
A 90-minute map in five moves.
Start with history
What makes a technology "general purpose," and why social change usually arrives through complements.
Use the task lens
Separate tasks, jobs, firms, contracts, and measured productivity.
Read today's evidence
Usage is broad, labor-market effects are concentrated, and capabilities are moving fast.
Bring it to India
Map the shock to cities, services, youth, migration, manufacturing, and state capacity.
End with choices
Policy, safety nets, taxation, and the capabilities young Indians should build.
Clear-eyed optimism
We should expect stress and still build for more experiments, not fewer.
A GPT is not one invention. It is a long reorganization.
It can be used across a large part of the economy, not just in one narrow industry.
The technology itself keeps improving, often through learning curves and engineering feedback.
Its full value depends on new processes, skills, institutions, infrastructure, and business models.
That third feature is the important one for society. Steam, electricity, computers, and the internet mattered because factories, cities, offices, schools, firms, and laws changed around them.
Sources: Bresnahan and Trajtenberg, General Purpose Technologies; David, The Dynamo and the Computer; Helpman, General Purpose Technologies and Economic Growth.
Earlier GPTs changed society by changing the production map.
Factories and transport
Power became less tied to muscle, animal energy, and local water sites. Industry and railways changed where work could happen.
The factory got redesigned
Motors did not instantly raise productivity. Factories had to be rebuilt around flow, layout, and flexible power.
Information became cheaper
Office work, logistics, finance, and design changed once calculation, storage, and communication became programmable.
Markets became searchable
Search, e-commerce, cloud software, and platforms reduced matching costs and reorganized distribution.
The lesson is not "all change is good." The lesson is that the first-order technology is only the beginning; the complements decide the social result.
Hayek explains why the relevant knowledge cannot be centralized in advance; Amara explains why timing fools us while local actors discover complements.
Sources: David on electricity and productivity delay; Bresnahan and Trajtenberg; Helpman, GPT volume; Hayek, The Use of Knowledge in Society; Greenstein on Amara's Law.
Output can rise while the social bargain becomes unstable.
When technology changes the relative scarcity of skills, capital, and coordination, society has to renegotiate who gets paid, who bears risk, and who gets a path into the new system.
Sources: Acemoglu and Restrepo, Automation and New Tasks; Acemoglu, The Simple Macroeconomics of AI; Trammell on capital substitution.
People rarely fight the machine. They fight the bargain around the machine.
Work standards
The Luddite protests of 1811-1816 were not a generic hatred of technology. They were a fight over wages, skill, and control in textile work.
Codified skill
The Jacquard loom used punched cards to automate patterned weaving, making a beautiful craft more programmable.
A useful myth
The shoe-thrown-into-the-machine origin story is probably folk etymology. But the myth survives because it captures a real social fear.
A better global frame: every society has its own version of "what happens when a productive pattern breaks before a new one is ready?"
Sources: Britannica on Luddites; Britannica on the Jacquard loom; Etymonline on sabotage; Acemoglu and Restrepo on displacement and reinstatement.
Automation is not fate. It is a tug of war between three effects.
Capital or software performs tasks that workers previously did.
Cheaper production expands demand and can raise demand for remaining tasks.
New human work appears: design, maintenance, judgment, sales, coordination, and care.
Baumol's cost disease is the slow-motion version: when some sectors get more productive, wages rise economy-wide, but hard-to-automate services become relatively expensive. AI matters because it may raise service productivity directly - or push cost pressure into the remaining human tasks.
The key question for AI is not whether it automates. It clearly does. The question is whether productivity and reinstatement are strong enough, and fast enough, to create better ladders for people.
Sources: Acemoglu and Restrepo, Automation and New Tasks; Acemoglu, The Simple Macroeconomics of AI; Baumol, Macroeconomics of Unbalanced Growth.
The task is not the job.
Labor markets hire jobs, not isolated tasks. A job is a bundle of predictable work, messy exceptions, accountability, relationships, and tacit knowledge.
What AI is good at first
Codified, text-heavy, pattern-rich tasks: drafting, summarizing, translating, coding, classifying, testing, answering, and planning.
What remains bundled
Authority, trust, physical presence, taste, moral responsibility, tacit know-how, local context, and the ability to handle weird exceptions.
A vivid case: U.S. travel-agent employment is now more than 60% below its dot-com peak - yet the agents who remain saw average weekly pay climb from 87% of the private-sector average in 2000 to 99% by 2025. The machine took the weak, separable part of the bundle and left people the strong part.
Cowen's Average Is Over generalizes the point: when software handles average, measurable work, returns rise to people who can supervise, combine, and complement machines.
If AI can do 40% of your tasks, which 60% becomes more valuable, and which 20% should never have been in your job?
Sources: Luis Garicano, The task is not the job; Alex Imas on scarcity and the human sector; Anthropic Economic Index task data; New Yorker on Cowen's Average Is Over.
Optimism should come from search capacity, not inevitability.
Arnold Kling's PSST frame sees the economy as patterns of sustainable specialization and trade. A technology shock breaks some patterns and opens a larger search space for new ones.
Workers learn new tools.
Firms redesign workflows.
Customers discover new wants.
Schools change credentials.
Cities absorb migration.
Finance funds experiments.
Regulators allow sandbox learning.
Public goods reduce friction.
Bad bets die quickly.
Good patterns scale.
This is where the optimism sits. AI gives us more possible patterns. The national question is whether India can test, kill, scale, and teach those patterns quickly enough.
Public-choice caveat: schools, regulators, public-good builders, and exit rules can become sludge if incumbents use them to protect old patterns. The test is whether institutions lower the cost of entry, feedback, and failure.
Sources: Arnold Kling, Macroeconomics: Some Defects; Greenstein on Amara's Law.
AI touches the white-collar ladder itself.
Natural language becomes a way to command software and produce drafts.
Code generation turns expert workflows into semi-automated loops.
Models increasingly act across tools, not just answer inside a chat window.
Digital work can diffuse much faster than physical machinery.
The risky part is not that every job disappears. The risky part is that the old apprenticeship model for professional skill may be partially automated before the new one is built.
Sources: Anthropic Economic Index; Anthropic geography and enterprise adoption report; METR long-task benchmark; Stanford Digital Economy Lab labor-market review.
AI is already broad, but not yet evenly economic.
Approximate weekly users by July 2025 in OpenAI's usage study.
Share of Anthropic Economic Index conversations classified as augmentation rather than automation.
Share of Claude.ai conversations that are "directive" - full task delegation rather than collaboration - in the Anthropic Economic Index.
Usage data says "diffusion is real." It does not yet say "productivity transformation is complete." That distinction matters for policy and careers.
Sources: OpenAI, How People Use ChatGPT; Anthropic Economic Index; Anthropic geography and enterprise adoption report.
And the leading edge is already moving from chat to delegation.
Growth in weekly active users of OpenAI's Codex agent in the first half of 2026 - rising fastest outside the original software-developer base.
Rise since January in users delegating at least one task estimated to need 8+ hours of expert human time.
More than 10% of users now run three or more agents at once in a given week - a new shape of knowledge work.
Share of OpenAI workers' work-related AI output tokens that now flow through the agent, not chat (June 11, 2026).
Read it carefully
OpenAI's own usage is the frontier, not the average firm - "a preview of agentic work once adoption frictions fall." The shift is real but uneven: smallest for individuals, larger for organisations, largest at the frontier. And it is delegated production, not consultation - the leading edge of the diffusion the last slide measured.
Sources: Johnston et al., The Shift to Agentic AI: Evidence from Codex (OpenAI, June 2026); OpenAI, How agents are transforming work.
The capability frontier is moving from answers toward work products.
The chart is conceptual. METR estimates a fast-rising time horizon for software tasks; GDPVal evaluates economically valuable work products.
Why this matters
Benchmarks are moving away from puzzles and toward deliverables: memos, spreadsheets, code changes, professional analysis, and multi-step tasks. That is closer to how workers are evaluated.
The policy caveat: benchmarks are not the economy. They tell us what may become automatable, not whether firms will reorganize well.
Sources: METR, Measuring AI Ability to Complete Long Tasks; METR arXiv paper; OpenAI GDPVal companion site; GDPVal paper.
The aggregate job apocalypse has not shown up. Concentrated stress has.
No broad collapse
Recent evidence through 2025 finds little sign that AI has caused a meaningful overall decline in hiring.
Entry-level risk
Some early-career workers in highly exposed occupations show clear relative employment declines.
Automation matters
Exposure is not enough. The key distinction is whether AI use substitutes for labor or complements it.
This is exactly the pattern a PSST lens would expect early in a shock: not one national average, but many local adjustments moving at different speeds.
Sources: Stanford Digital Economy Lab, AI and Labor Markets; Brynjolfsson, Chandar, and Chen, Canaries in the Coal Mine; Anthropic Economic Index.
The young professional ladder is the part to watch.
The group where Brynjolfsson, Chandar, and Chen find the sharpest relative declines in exposed occupations.
Relative employment decline for young workers (ages 22-25) in the most AI-exposed occupations, after controlling for firm-level shocks.
The initial adjustment appears more in employment than in compensation.
The hypothesis is intuitive: early-career workers often sell codified knowledge before they have accumulated tacit judgment. AI is strongest where the first kind of knowledge is explicit.
Sources: Brynjolfsson, Chandar, and Chen, Canaries in the Coal Mine; Stanford Digital Economy Lab labor-market review.
Beside the breaking rung, a new one: the age of the solopreneur.
Americans whose primary income is solo work paying $100k+ a year (2023) - roughly double the early-2010s level.
More solopreneurs crossed $5m and $10m in revenue in 2025 than in 2023 - scaling without employees (2× crossed $1m).
More likely a 2025 Stripe cohort hits $1m revenue within a year than the 2023 cohort - and 3× more likely than 2019.
AI-influenced journeys as a share of new Stripe sign-ups versus a year earlier; nonemployer growth tracks AI adoption by industry.
The flip side of the canary
The same technology that erodes the entry-level rung is filling the capability gaps that once required hiring - "the revenge of the idea guys." For India this is optionality, not yet evidence - but a young population, DPI rails, and a vast home market are the raw material for a solo-scaling path, if the skills arrive.
Sources: Tedeschi, Rama and Cruickshank, The Age of the Solopreneur (Stripe Economics, June 2026).
Productivity may appear first as a different firm shape.
AI-native firms are a useful early clue because they are not trying to retrofit yesterday's hierarchy.
Smaller average employment in one 2026 study of AI-native startups.
Higher engineer share in the same study.
Lower entry-level and manager shares, pointing to flatter organizations.
This is not macro proof. It is a mechanism: the frontier may first show up as smaller teams doing more, with fewer classic apprenticeship slots.
Sources: Kim and Koning, AI-Native Firms; Marginal Revolution, AI-Native Firms.
Coase and Cheung explain why productivity statistics move late.
The tool arrives first. Then firms have to discover which contracts, boundaries, and accountability systems make the new division of labor legible.
Why firms?
Firms exist when internal coordination is cheaper than repeated market contracting.
Which contract?
The real choice is among contracts when output, effort, quality, and risk are hard to measure.
What changes?
Monitoring, drafting, matching, and delegation get cheaper. But responsibility does not disappear.
That is why "firms organize fast enough" is not a small detail. It is the channel through which task-level capability becomes social productivity.
Sources: Coase, The Nature of the Firm; Cheung, The Contractual Nature of the Firm; Kim and Koning, AI-Native Firms.
The same GPT lands on a very different economy.
The U.S. story is not India's story. India has a services-heavy GDP, agriculture-heavy employment, a thin formal job ladder, strong IT capability, weak manufacturing absorption, and enormous internal variation.
The question is not "Will AI help India?" The question is: which Indians, which cities, which firms, and which ladders?
Sources: ILO, India Employment Report 2024; DataForIndia, PLFS explainer; CSEP, India at Work.
India's output structure and work structure do not line up neatly.
Approximate visual: services dominate output, while agriculture still absorbs roughly half of employment.
The policy trap
AI first hits many formal, urban, service-sector tasks. But India's larger employment challenge remains moving people from low-productivity work into better, more stable work.
That means India's AI question is also a development question.
Sources: ILO, India Employment Report 2024; DataForIndia, PLFS explainer; CSEP, India at Work; DataForIndia on manufacturing jobs.
Services are not one thing.
India's services strength is real. But the services that scale globally, pay well, and train young workers are a much thinner bridge than "services share of GDP" suggests.
The arbitrage is headcount. India's IT-services majors run on the order of $50,000 of revenue per employee; the US software benchmark for a "good" company is now near $300,000. That five-to-six-fold gap is the business model - and it is exactly what AI compresses, by letting smaller teams do the codified work India sold in volume.
Sources: ILO, India Employment Report 2024; CSEP, India at Work; Anthropic Economic Index; OpenAI usage study; Lightspeed on India SaaS revenue per employee.
China's lesson is not "copy China." It is "learning needs a machine."
The China contrast
China built dense manufacturing ecosystems, infrastructure, process-learning, and state capacity at unusual speed. That created ladders for rural migrants and suppliers.
India's missing middle
India's manufacturing problem is not just too little factory employment. It is too few large, labor-absorbing, productivity-raising firms in the middle of the distribution.
One deep root is human capital. China expanded education bottom-up - mass primary literacy first, then a heavy tilt toward engineering and vocational training - while India went top-down, growing secondary and higher education before universal basic schooling. Education inequality accounts for roughly a quarter of wage inequality in India, against under an eighth in China. Composition, not just headcount, shapes who can climb.
AI makes this sharper. If tradable services absorb fewer entry-level graduates, manufacturing and city-building cannot remain permanently "next decade's reform."
Sources: ADB, firm size distribution in Indian manufacturing; Bharti and Yang, Human capital in China and India, 1900-2020; DataForIndia on manufacturing jobs; Joe Studwell, How Asia Works; Dan Wang, Breakneck; Peter Hessler, Other Rivers.
The first Indian AI shock will be local, urban, and occupational.
Bengaluru
Software products, IT services, startups, cloud tooling, analytics.
Hyderabad
IT services, global capability centers, pharma services, back-office operations.
Gurugram and Noida
Consulting, BPO, analytics, finance operations, customer support.
Pune and Chennai
Engineering services, automotive software, IT services, support operations.
This is not a prediction that these cities collapse. It is a claim about exposure: their growth stories are tied to tasks that AI can increasingly perform or radically reshape.
Sources: Anthropic Economic Index occupational task exposure; ILO, India Employment Report 2024; CSEP, India at Work.
What China did to some U.S. towns, AI may do to some Indian task clusters.
The China Shock papers connect a national shock to local labor markets. For India, map AI exposure to city-industry-occupation clusters.
Which tasks?
Code, QA, support, documentation, moderation, analytics, finance operations.
Which cities?
Measure city employment shares in exposed service clusters.
Which buffers?
Firm upgrading, new entry, migration options, education quality, local services.
The model to build is not "AI exposure for India." It is AI exposure multiplied by local specialization and adjustment capacity.
Borrow the method, not the analogy: China's was a goods-trade shock to factory towns; India's is digital-task displacement in services - a faster but more concentrated adjustment.
Sources: Autor, Dorn, and Hanson, The China Syndrome; Autor et al., The China Shock; Anthropic Economic Index; DataForIndia on PLFS.
A city shock is also a family, education, and migration shock.
Push-pull changes
If services ladders weaken, the pull of major cities changes for graduates and families.
Urban pressure
Job-market stress interacts with rents, commutes, infrastructure, and city governance.
Credential bets
Families invest in degrees and coaching based on yesterday's ladder into formal work.
Wider effects
Urban professional earnings support consumption, siblings, parents, and local aspirations elsewhere.
The India story cannot end with a productivity statistic. It has to include the social geography of adjustment.
Sources: ILO, India Employment Report 2024; CSEP, India at Work; Peter Hessler, Other Rivers; Dan Wang, Breakneck.
AI exposure may become a new layer on old inequalities.
The distributional question is not only whether a task is exposed. It is who gets access to exposed, well-paid tasks, and who gets displaced from them.
A 2026 working paper links AI-exposed graduate jobs in India to existing caste inequality.
Mobility, safety, care work, and social norms shape who can move into new AI-complementary roles.
AI opportunity is likely to cluster around cities, English fluency, firms, and institutions.
A fair optimism has to ask not just how much AI raises output, but how widely Indians can enter the new patterns.
Sources: Mishra, The Privilege of Exposure; ILO, India Employment Report 2024; DataForIndia on PLFS.
India has more shots on goal than the pessimistic story admits.
A large English-using professional class can adopt frontier tools early - though English is double-edged: it is also what makes India's entry-level tasks the most automatable.
India already has service firms that know global clients, software delivery, and process discipline.
Digital public infrastructure can lower transaction and verification costs.
A large domestic market allows adaptation for local language, cost, and institutional realities.
The strategic question is whether these advantages become learning machines, not whether they look impressive in a slide about "India's demographic dividend."
Sources: India Stack; ILO, India Employment Report 2024; CSEP, India at Work; Joe Studwell, How Asia Works.
The fragile part is the ladder into good work.
Codified tasks
Many first jobs train people through exactly the tasks AI can now draft, test, summarize, and classify.
Uneven quality
Degrees do not reliably certify the judgment and tacit skill needed when basic output is automated.
Weak insurance
Many workers lack the buffers that make retraining and migration feasible.
Underbuilt
The places that attract talent often make adjustment expensive through housing and infrastructure limits.
Whether this is a structural break or a passing adjustment turns on how Indian firms adopt: use AI mainly to cut costs (automation) and the entry-level rung snaps; use it to extend what people can do (augmentation) and the rung bends but holds. The 57/43 augmentation-to-automation split from the usage evidence is the number to watch, sector by sector.
Sources: ILO, India Employment Report 2024; DataForIndia on PLFS; CSEP, India at Work; Brynjolfsson et al., Canaries.
India should build an AI shock dashboard before the shock is obvious.
Task exposure
Map Indian occupations and tasks to AI exposure, separating automation from augmentation.
Local concentration
Track city-industry clusters, hiring, entry-level wages, layoffs, and migration signals.
Adoption quality
Measure whether firms are using AI for cost cutting alone or for new products, new markets, and worker leverage.
The PLFS redesign is helpful, but AI diffusion needs higher-frequency, task-level, city-level data. Otherwise policy will arrive after the adjustment has already happened.
Sources: DataForIndia, PLFS explainer; ILO, India Employment Report 2024; Anthropic Economic Index methodology; Mishra, India graduate AI exposure.
The PSST policy goal: make good experiments cheaper.
Workflow labs
Help small firms test AI for sales, compliance, inventory, customer service, and accounting.
Procurement as learning
Buy measurable AI improvements in courts, health, education, inspections, and citizen services.
Interoperability
Support open data standards, audit trails, procurement templates, and liability clarity.
Do not subsidize "AI adoption" as a slogan. Subsidize discovery of repeatable, measurable patterns that raise output and widen access.
Sources: Kling on PSST; Anthropic Economic Index project; India Stack; ILO, India Employment Report 2024.
Frontier access decides who discovers what AI is for.
If frontier models cannot diffuse as general-purpose inputs, labs are pushed to capture downstream applications themselves. That may solve inference economics, but it narrows the social search process.
If only approved institutions can touch frontier capability, who gets to decide what AI is for in Indian conditions?
Sources: Dean Ball, What Should Be Done; Séb Krier on endogenous social equilibria; Hayek, The Use of Knowledge in Society; Kling on PSST.
AI policy is also infrastructure, energy, cities, and education policy.
Sources: IEA, Energy and AI; ILO, India Employment Report 2024; CSEP, India at Work; Dan Wang, Breakneck.
The UBI question for India is really a state-capacity question.
A universal income floor is attractive when adjustment is broad and hard to target. But India's practical question is what kind of floor the state can finance, deliver, and update without weakening the search for new work.
Basic security
Cash support can reduce desperation and make retraining or migration less fragile.
Active help
Placement, apprenticeships, wage insurance, and training matter when the shock is occupational.
Hard tradeoffs
India must compare UBI to health, schooling, infrastructure, nutrition, and city investment.
The best Indian frame may be a layered system: a reliable cash floor, portable benefits, and aggressive support for moving into new patterns of work.
Sources: Economic Survey 2016-17, UBI chapter; ILO, India Employment Report 2024.
A global token tax sounds elegant until you ask what it taxes.
The temptation
If AI automates labor, tax compute or tokens and use the revenue to fund the social transition.
The problem
Compute is an intermediate input - globally mobile, hard to define, and often exactly what society wants cheaper. The Diamond-Mirrlees rule says tax final output and redistribute, but keep intermediate goods untaxed; a compute tax is like "taxing steel during the industrial revolution."
Sources: Brian Albrecht, A compute tax is a really dumb idea; Korinek and Lockwood, Public Finance in the Age of AI; Anthropic Economic Index project; Economics of Transformative AI, Coasean singularity chapter.
For India's youth, the right question is: what stays scarce?
Use the machine well
Prompting is too small a word. Learn delegation, verification, tool choice, context design, and error detection.
Know something real
Healthcare, law, finance, logistics, education, energy, agriculture, design, operations, public policy.
Own responsibility
Taste, ethics, accountability, client trust, ambiguity, and the courage to decide under uncertainty.
Design education around projects where students must use AI, explain what they trusted, show what they rejected, and defend the final judgment.
Sources: Arnold Kling on project learning and AI; Alex Imas on scarcity and human value; Brynjolfsson et al., Canaries; ILO, India Employment Report 2024.
The future is not a forecast. It is a coordination problem.
AI increases the space of possible production patterns. Some old ladders will weaken. Some new ladders will be built. India should not wait to find out which is which.
Build task, city, firm, and youth dashboards before aggregate statistics move.
Let firms, schools, states, and cities test new productive patterns quickly.
Protect people through the transition without taxing away the tools that help them adapt.
India is vulnerable, but optionality is real if we build the institutions that turn experiments into ladders.
Sources: Kling on PSST; Acemoglu and Restrepo; ILO, India Employment Report 2024.
A reading list for turning the talk into an essay.
The deck is designed to become a longer essay. These are the sources I would keep closest while writing it.
Source note: This reading list was selected from the user's Obsidian clippings, Economics of AI folder, and Readwise notes, then complemented with linked public sources for the deck.