AI and the Economy: Three Futures for Jobs, Wages and Who Gets the Money
A new economic model sketches three versions of the next five years, from mild to wild. In every one the economy grows. What changes is who ends up holding the gains.
Ask ten people what artificial intelligence will do to jobs and you get ten answers, most of them shouted. A team of economists has tried something more useful: instead of predicting one future, they have built a model that lets you pick your assumptions and see what kind of economy those assumptions would produce. Three of those futures are worth walking through, because the differences between them are not really about growth. They are about who ends up better off.
The work discussed here is the economic scenarios project published by Anthropic's economics team in September 2026, along with a companion research paper and a survey of more than ten thousand people. It models the United States, not India. Two things worth keeping in mind as you read: it is a simplified model rather than a forecast, and it comes from a company that sells the technology it is modelling. Neither fact makes it wrong. Both are reasons to read it as a way of thinking rather than as a prediction.
In all three futures the country produces more. In two of them, a growing share of that extra output goes to whoever owns the machines rather than to whoever does the work.
First idea: your job is a stack of small jobs
The model starts somewhere unusual. It does not ask "will AI replace accountants". It breaks every job into the individual tasks a person actually performs, then asks what happens to each task on its own.
Take a nurse. Over one shift she checks on patients, draws blood, decides who needs attention first, writes up readings, and orders supplies for the ward. That is one job title and a dozen different activities. Machines have nothing to say about some of them and quite a lot to say about others.
Four things can happen to any single task:
The task still needs a person in the room. Nobody is going to have a machine help a patient wash.
The person still does it, but faster or better. Drafting discharge notes, planning the shift, keeping an eye on patients from a distance.
The task stops needing a person at all. Logging routine readings. Reordering ward stock.
Work that did not exist before. Checking whether the machine sorted the queue sensibly. Signing off on a plan it drafted.
That last one matters more than people expect. Every previous wave of technology destroyed some tasks and invented others. Nobody hand-writes ward charts any more. Nobody monitored patients remotely thirty years ago. The bundle is always being reshuffled; the question is the balance and the speed.
Second idea: add up everyone's tasks and you get the economy
Now scale that up. Every task in the nurse's day happens thousands of times on thousands of wards. Add every task performed by every worker in a country over a year, and the total is the economy itself. For the United States that came to more than thirty trillion dollars of output last year.
Which turns a vague argument into three answerable questions.
- How much of the work can machines actually take on? Not in a demo. In the ordinary run of a working day.
- How much faster does the work get done? A tool that saves two minutes an hour is a different economy from one that saves two hours.
- How quickly do people and firms actually start using it? Capability sitting unused changes nothing. This is usually the slow part.
Change the answers and you change the economy that comes out the other end. That is the whole machinery of the model.
Three futures, side by side
Real gains, arriving slowly, hard to pick out in the national figures. Roughly what other useful technologies have delivered before.
By 2030 machines can handle about half of all desk work, and mostly on their own. Most of that work is still done the old way, because take-up lags. The economy grows at about twice its usual pace.
Machines outperform people at nearly all desk work, do it without supervision, and create almost no new desk work in return. Growth reaches roughly fifteen percent a year, which doubles the economy about every four and a half years.
Notice what the three have in common. The country gets richer in every one of them. Nobody in this model is arguing that the economy shrinks. If you only read the growth line, all three look like good news.
Growth is the easy part. The split is the hard part.
Here is the number that does the real work in this study. Of every rupee an economy produces, some goes to people as pay and the rest goes to whoever owns the equipment, the buildings and the technology. Today, in round terms, that split is about sixty to forty in favour of pay.
Where each rupee of output goes today
In the bigger two futures the model expects this line to shift, with a larger slice moving to the ownership side even while total pay rises.
Why would that happen? Because if machines can be pointed at more and more kinds of work, then owning machines becomes more valuable, and demand for them pushes their price up. The pie gets bigger and the owner's slice gets bigger faster.
This is the part that gets lost in headlines about growth. You can have an economy that is genuinely, measurably richer, in which a typical desk worker is no better off than before. Both things fit inside the same growth figure.
Desk work and hands-on work go in opposite directions
The model's second uncomfortable finding is that the pain and the gain land on different people.
Work done at a desk
- Most exposed, because much of it is reading, writing, checking and deciding
- In the big future, pay barely moves
- In the wild future, pay falls by more than a tenth by 2030, and unemployment in these roles climbs past the levels a bad recession usually produces
- Examples the study uses: writing software, handling customer calls
Work done with your hands
- Least exposed, because presence is the point
- Pay rises, in every one of the three futures
- Demand goes up partly because desk work got faster: quicker drawings and approvals mean more building actually happens
- Examples the study uses: electricians, nurses, construction
Average pay across everybody rises in all three futures. That average hides a widening gap, which is exactly why an average is a poor thing to reassure yourself with.
There is a further catch. Moving from one kind of work to another sounds simple on a chart and is brutal in life. People have to want to move, learn something new, and then actually be hired. The faster a future arrives, the more people are stuck in the gap between the job they had and the job they might get. That gap is what shows up in the figures as unemployment.
What ordinary people actually expect
The team also asked more than ten thousand people in the United States what they thought machines would be able to do, how quickly workplaces would take them up, and how hard it would be to find new work.
That is a striking result. The middle of public opinion is not the comfortable scenario. It is the one where desk work changes substantially and pay for it stops rising.
What a reader in India should take from an American model
This section is my own reading, not part of the study. The model covers the United States and does not attempt to describe India.
Three reasons it still matters here.
- Our exposure is unusually concentrated. India sells desk work to the world: software services, back-office processing, bookkeeping, support and documentation. That is precisely the category the model treats as most exposed. A country whose exports are heavily weighted towards this work has more riding on the answer than a country whose exports are commodities.
- The ownership question decides who benefits. If more of each rupee flows to whoever owns the technology, then it matters a great deal whether Indian firms and households own any of it, or only rent it. That is a question about savings, capital markets and where profits are booked, not a question about machines.
- The exposed work is not the unskilled work. Every previous wave hit routine physical work hardest. This one points at qualified, salaried, sitting-down work: the exact ladder that has carried a generation of Indian graduates into the middle class. That is a different social problem and it needs a different answer.
For anyone in a profession like mine, the practical exercise is the one the model itself suggests. Stop asking whether your job survives. Write down the tasks that actually fill your week, and sort them into the four buckets above. Most people find the honest answer is that a chunk of their week is already in bucket three, and that the valuable part of their work was always in the buckets a machine cannot reach: judgement, responsibility, and being the person who signs.
What the model deliberately leaves out
The authors are unusually candid about this, which is a point in the work's favour. It leaves out government policy, the ordinary ups and downs of the business cycle, financial shocks, the enormous spending on data centres, and any scenario involving capable robots doing physical work. Reviewers also pointed out that it follows groups rather than individual people, so it cannot say much about what a displaced worker actually goes through.
Read it, then, as a tidy way of connecting assumptions to consequences. Change what you believe about machines and you can see, roughly, what that belief implies for jobs and pay. That is genuinely useful and it is not the same thing as knowing what will happen.
The economy grows in every version. Whether that growth reaches ordinary pay packets is not decided by the technology. It is decided by how quickly people can move between kinds of work, and by who owns what.
Frequently Asked Questions
Is this a prediction that unemployment will rise? No. It is a model showing what would follow from a given set of assumptions. Unemployment rises sharply in only one of the three futures, the most extreme one, and stays within normal historical ranges in the other two. The study's own survey found the typical person's expectations land in the middle scenario, where unemployment settles around five percent.
Which kinds of work are most exposed? Work made up largely of reading, writing, checking, summarising and deciding, done at a desk. The study uses software development and customer service as examples. Work that needs physical presence, such as nursing, electrical work and construction, is least exposed and sees pay rise in all three futures.
If the economy grows so much, why would pay stagnate? Because output is split between pay and the owners of equipment and technology, roughly sixty to forty today. If machines can be applied to more kinds of work, demand for them rises and a larger share of each rupee flows to their owners. Total output and total pay can both rise while pay for the affected group goes nowhere.
Does this apply to India? The model is built on United States data and does not claim to describe India. The mechanism it describes is not country-specific, and India's heavy exposure to exported desk work arguably makes the question sharper here, but no number in the study should be read as an Indian figure.
Who produced this and should that affect how I read it? It was produced by the economics team at Anthropic, a company that builds and sells artificial intelligence, with a companion technical paper reviewed by outside academic economists. The reviewers were not asked to endorse the conclusions and several published criticisms remain unresolved. Treat it as serious work with an obvious interest attached, and weigh it accordingly.
What is the single most useful idea in it? That a job is a bundle of tasks rather than one indivisible thing. It turns an unanswerable question about your job title into an answerable one about your working week.
General commentary for readers in India, written 10 September 2026. It is not investment advice, not a forecast, and not a recommendation to do anything. Figures described are the study's own scenario settings and survey results; the underlying work is published in full by its authors and is the version current at the date above.
This piece, as an infographic
Every figure on it comes from the piece above. Share it freely — a link back is all that is asked.