Will AI make India rich faster? Will it take away the office jobs a generation studied for? Both? We took the Anthropic Institute's model of how AI changes an economy and ran it on India.
Same model, same assumptions about AI. Only the economy is different. How we built this is at the end.
Every job is a bundle of tasks. AI leaves some tasks alone, helps with some, does some on its own, and creates a few new ones. Take a loan officer at a bank branch in Indore.
Now count every loan appraisal, every KYC check and every field visit across every branch in the country. Add every other job. That is India's economy: about ₹346 lakh crore of work a year. What AI does to the tasks decides what it does to the economy.
The model only lets AI touch office and professional work: managers, professionals, technicians, clerks and sales staff. Everyone else is left alone. That one fact drives most of what follows.
India has about as many office and professional workers as the US, around 110 million. They are just a much smaller slice of a much bigger workforce.
The scenarios differ in how good AI gets, how widely it is used, and how much of the work it does on its own. The settings are the Anthropic team's. The results are India's.
AI touches a third of office work and is used on a fifth of it. Think Copilot in IT firms and banks. The effect is real but hard to see in the national numbers.
AI can do half of office work and is used on 40% of it, mostly on its own. India grows at 8% a year. Office hiring shrinks and office pay stalls while other wages rise.
AI does 80% of office work by itself and creates no new tasks for people. India grows at nearly 11%. One in four office workers is out of work and their pay falls.
Answer five questions about AI. The model works out the India they imply in 2030. Every combination of answers has been run through the full model.
Only 23 of every 100 workers are exposed, so the upside is smaller. In rupees, the extreme case adds an economy the size of Maharashtra by 2030.
percent above the no-AI path
percent per year; 6.5% without AI
AI shrinks office work and grows demand for everything else. Workers have to cross over, and crossing over is slow. In India the crossing is from a salaried desk to farm, site or shop work, at a fraction of the pay.
unemployment rate, percent
unemployment rate, percent
unemployment rate, percent
Pay in construction, transport, farming and care work rises because a richer economy buys more of what they do. Office pay goes the other way. In the extreme case it ends 14% below today's level after inflation.
pay, percent above the no-AI path
pay, percent above the no-AI path
pay, percent above the no-AI path
Today 53 paise of every rupee India produces goes to workers and 47 to capital. When AI does a task, the pay for that task moves from a worker to whoever owns the AI.
Same AI, same model, extreme scenario. India gains a third as much and its office workers lose more.
The model is deliberately simple. For India, some of what it leaves out is the story.
IT and back-office firms earn $246 billion a year from clients abroad. Whether that work gets automated is decided in a US bank's budget meeting, not in India. Indian adoption can lag and the jobs still go.
When pay moves from workers to capital, the model books it as Indian capital income. Much of it is a bill to Microsoft, Google, OpenAI or Anthropic. GDP rises by these numbers. National income rises by less.
India's workforce grows 1.5% a year. A shrinking office sector gets there by not hiring. It shows up as graduate unemployment, already 11 to 13%, before it shows up as layoffs.
The model moves a displaced worker into "other work" at that group's pay. In the US that group includes electricians and nurses. In India it is mostly farm, site and daily-wage work at a third to half of office pay.
MGNREGA guarantees rural manual work. Nothing covers an unemployed software tester in Pune. Holding office incomes at their no-AI level in the extreme case would cost about 6% of GDP, half of what the Centre collects in tax.
Food is 40 to 47% of household spending. An AI dividend gets spent on physical things and in-person services. That is why the model raises pay for construction, transport and farm workers. An AI boom shows up on building sites first.
Every setting in the model is something we can measure. The next two years of Indian data will say which scenario we are in. The paths only separate after 2027.
This page is an adaptation of Scenarios for our Economic Future by the Anthropic Institute (Korinek, Jones, Sacher, Cotter and McCrory, working paper 2026-02, September 2026). Their model treats the economy as a set of tasks that either people or AI can do, with two groups of workers (office and professional work, which AI can touch, and everything else, which it cannot), a capital market, an ideas channel, and a monthly job-search block with slow-moving office wages.
We rebuilt the model from the paper's equations and checked it by reproducing the paper's US results (their Table 3) to within 0.1 percentage points on every line. Then we changed only the inputs that describe the economy, listed below, and left every assumption about AI as the paper set it. The 2030 amount of work AI affects is scaled so that the same fraction of office work is affected as in the US: 32%, 48% and 80% of it. The explorer above runs the same model on all 4,375 combinations of answers, with the paper's survey settings for the inputs the questions do not cover.
| Input | India | US (paper) | Basis for the India value |
|---|---|---|---|
| Share of workers in office and professional work | 23.0% | 62.4% | PLFS 2025 via ILOSTAT, ISCO-08 groups 1 to 4 plus sales workers (52): 2.8% managers, 5.9% professionals, 2.7% technicians, 2.3% clerks, 9.4% sales |
| Share of income going to workers | 53% | 60% | Penn World Table 11, which imputes the labour income of the self-employed |
| Growth without AI: output per worker; labour force | 5.0%; 1.5% | 1.67%; 0.33% | GDP 6.5% (RBI 6.7% and IMF 6.4% for FY27); implied productivity growth 2.65%, close to India KLEMS |
| Capital to output; net return to capital | 3.0; 10.7% | 3.5; 6.5% | Investment 31% of GDP over growth plus 5% depreciation |
| Normal unemployment; office; other | 4.3%; 6.6%; 3.6% | 3.8%; 2.9%; 5.4% | PLFS weekly status Jan to Mar 2026. India's unemployment is concentrated among the educated, the reverse of the US |
| Monthly chance of finding a job | 0.15 | 0.23 | Assumption. PLFS panel data show slow moves into salaried work and long spells for educated job seekers |
| Work AI already affects, mid-2026 | 5.1% of tasks | 14% | Same exposure within office work as the paper measures (22%), times India's 23% share |
| Work actually done with AI, mid-2026 | 5% of exposed tasks | 10% | Anthropic Economic Index: India's use per office worker is a quarter to a third of the US. Large IT firms are at parity; small and informal firms near zero |
| Everything about AI in 2030: diffusion, gain per task, share done on its own, new tasks, ease of switching jobs, wage stickiness, capital supply | Kept as in the paper | Modest, substantial and extreme settings unchanged | |
Substantial and extreme results, changing one Indian input at a time. The paper counts office workers by heads. They earn about twice as much as others, so by pay their share is nearer 35% than 23%. Weighting by pay raises the growth numbers by half. The truth is between the two rows.
| Variant (substantial · extreme) | GDP, % above no-AI | Office pay, % | Office unemployment, % | All unemployment, % | Workers' share, % |
|---|---|---|---|---|---|
| Base India | 3.0 · 9.3 | −5.0 · −27.8 | 9.6 · 26.2 | 4.5 · 7.6 | 51.7 · 48.3 |
| Office share by pay, not heads (35%) | 4.5 · 14.6 | −3.4 · −23.2 | 8.1 · 24.4 | 4.7 · 9.6 | 51.0 · 45.8 |
| Adoption lag: 2030 use half the US level | 1.6 · 5.4 | −2.6 · −12.0 | 7.8 · 14.2 | 4.3 · 5.2 | 52.3 · 50.6 |
| Scarce capital (imported compute, ε = 1) | 2.3 · 6.6 | −6.4 · −32.1 | 9.5 · 25.8 | 4.4 · 7.5 | 51.3 · 47.1 |
| Unlimited capital (ε = ∞) | 3.7 · 12.0 | −3.5 · −23.2 | 9.6 · 26.6 | 4.5 · 7.6 | 52.0 · 49.5 |
| Sticky formal-sector wages (ξ = 0.9) | 2.6 · 8.0 | −0.7 · −4.9 | 11.2 · 34.5 | 4.9 · 9.2 | 51.7 · 48.4 |
| Fully flexible wages (ξ = 0) | 3.8 · 10.7 | −15.3 · −62.0 | 6.0 · 5.7 | 3.5 · 3.2 | 51.6 · 48.2 |
| Slower job finding (0.10 a month) | 3.1 · 9.3 | −5.1 · −28.0 | 9.7 · 26.4 | 4.3 · 7.5 | 51.7 · 48.3 |
| United States (paper) | 8.3 · 32.4 | −0.3 · −11.5 | 4.5 · 17.9 | 4.6 · 11.9 | 56.1 · 45.2 |
Limits to keep in mind. The model pays everyone in a group the same wage, so it cannot show the pay cut a displaced office worker takes. It is a closed economy with Indian-owned capital, so it misses exported office work and foreign-owned AI. It has no informal sector and no robots, so the 77% it treats as safe will not stay safe past 2030. And, as the paper says, it has no business cycles, no policy response and no catastrophic risks. These are not forecasts. They show what would have to be true of AI, and of India, for each future to happen.
Siddhant Mundra, September 2026.