Prospere AI - AI Automation Solutions for Business Growth

Anthropic's economic scenarios for 2030: which side of the ledger is your business on?

September 11, 2026

Anthropic's economic scenarios for 2030: which side of the ledger is your business on?

On Wednesday, 9 September, Anthropic's economics team published a model of the US economy in 2030 and an interactive scenario explorer. You type in what you think AI will be able to do in four years, and the model shows the GDP, wages and unemployment that follow. Most headlines boiled it down to one line about unemployment worse than the pandemic. That is a description of one column in a three-column table. The most important sentence in the report never made it into a headline.

In all three scenarios the economy grows. In all three, a smaller share of that growth goes to wages and a larger share goes to the owners of machines and systems. The scenarios differ only in size and speed. The authors say so themselves: "What the scenarios disagree about is size and speed."

If you run a company of 10, 50 or 200 people, this is not an article about whether AI will take your people's jobs. It is an article about which side of that shift your company will be on.

Three scenarios in numbers

The model has five dials: what share of tasks AI can do, what share of those it is actually used for, whether it replaces the person or assists them, how much productivity it adds per task, and how long a displaced worker takes to find a new job. The three scenarios are three settings of those dials. The starting point is mid-2026, the finish line is January 2030.

In 2030 Modest Substantial Extreme
US GDP vs. a world without AI +1.6% +8.3% +32.4%
Annual growth rate (2% today) 2.4% 5.4% 15.4%
Share of all tasks affected by AI 4% 12% 30%
Wages in cognitive occupations +0.4% −0.3% −11.5%
Wages in all other occupations +1.1% +5.9% +33.6%
Unemployment among cognitive workers (2.9% today) 2.9% 4.5% 17.9%
Labour share of income (60% today) 59.4% 56.1% 45.2%

Source: Korinek, Jones, Sacher, Cotter, McCrory, "Economic Scenarios for Transformative AI", The Anthropic Institute, September 2026, Table 3.

The modest scenario is AI as another internet: visible inside companies, invisible in the statistics. Substantial is a technology bigger than the internet or the railroad. Growth of 5.4% a year beats the dot-com record, when the US economy grew 4.7% in 1999. Extreme requires AI that improves itself, adopted quickly across all office work. The economy then doubles in under five years, and one in five cognitive workers is out of a job.

The authors attach no probabilities. They write that none of the three can be ruled out today and that the paths only separate after 2027. Until then all three look the same in the data. For a company that means one thing: you have roughly a year to get on the right side of this, and during that year no statistic will tell you which scenario is playing out.

The economy is bundles of tasks. So is your company

The most useful thing in this report is not the numbers. It is the method. The model does not look at jobs. It looks at tasks.

Every job is a bundle of tasks, taken from the US Department of Labor's O*NET database. Anthropic illustrates it with a nurse: rounds, drawing blood, triage, charting vitals, ordering supplies for the ward. For every task, AI can do one of four things:

  • nothing, because the task cannot be handed to a machine (bathing a patient),
  • augment the person: the nurse drafts discharge notes faster, monitors patients remotely, plans the shift,
  • automate: AI charts the vitals and orders the supplies on its own,
  • create a new task: someone has to check how the AI triaged patients and sign off on the care plan it proposed.

The bundle changes, the job stays. Thirty years ago nobody monitored patients remotely, and today almost nobody keeps paper charts.

You count a single company the same way. When we audit a client's processes we never ask "which position do we replace". We write down what people actually do in a week and put one of four letters next to each task. In a clinic, a voice agent took over one task from the reception bundle: answering the phone and booking appointments after hours. The dozen others stayed with people. At a metal-industry manufacturer we automated the preparation of quote emails, but a person approves every send. In a public-tender monitoring system, rules written in code decide what matters and the model only summarises documents. That is augmentation, not automation.

The fourth category, new tasks, is the one most often skipped and the most expensive to skip. In September we audited an assistant on a client's website. In one language version, zero out of sixteen enquiries had reached the database for several weeks, because an access token had expired and nobody got an alert. The machine was working. Nobody was watching it. Supervising AI is a new task in the bundle, and it has to be in someone's job description. If you want to see that worksheet for one of your own processes, book 30 minutes.

Who gets the money

Today, of every dollar the US economy produces, about 60 cents goes to workers and 40 to capital. In the substantial scenario the labour share falls to 56%. Four points in four years, roughly the entire decline the US saw over the four decades after 1980. In the extreme scenario it falls to 45%, and capital income is 81% higher than in a world without AI.

The report follows this to the end. In the extreme scenario, total labour income in 2030 is almost exactly what it would have been without AI. The economy is a third larger, and all of that increase goes to the owners of capital. Compensating cognitive workers for their loss would take a transfer of about 9% of GDP, comparable to Social Security and Medicare combined. The authors add that transfers of that scale have no precedent.

"But whether and how those resources reach the people who bear the cost is not something growth delivers by itself."

Korinek et al., conclusions of the report.

For a business owner this paragraph has one consequence. In this model, capital is not a factory and not the stock market. Capital is systems that perform tasks instead of people. A company that has put AI into quoting, phone handling or reporting is on the capital side and collects the shift. A company that sells its people's hours and competes on price with those who have deployed is on the labour side. In the substantial scenario the difference is a few points of margin. In the extreme scenario it is survival.

There is a second side to the same effect. Wages outside cognitive work rise in every scenario: 5.9% in the substantial one, 33.6% in the extreme one. If you have people on the shop floor, on site or in the field, that is your future wage pressure. The report explains the mechanism: faster designs and permits mean more construction, and more construction means demand for contractors. AI savings in the office will be paying for raises on the floor.

Adoption decides the scenario, not the model

This part should change how you think about buying AI.

Alongside the report, Anthropic published a survey: nearly 11,000 US adults asked in August what AI will be able to do in 2030. The median answer: AI will handle 6 of the 8 tasks described at a professional level. That is more than the substantial scenario assumes. Yet the median respondent's outcome lands in the substantial scenario, not the extreme one. The same people expect AI to be used on only 40% of the tasks it can do.

The model shows it in black and white. In the substantial scenario AI affects 12% of tasks in the economy even though it could do more. Deployment is the limit, not capability. The extreme scenario is half better models and half companies actually using them.

That is exactly what we see with clients. For two years the models have been able to do more than companies get out of them. The bottleneck is whether someone has mapped the process, connected the data, set up permissions and supervision. We wrote the same thing a week ago about Jacob Coxon leaving Anthropic: the pace of model development will not slow down, and the advantage comes from deployment, not from waiting for a better version.

The authors put it in the language of economics: every parameter of the model can be measured, including "the share of firms using it". Your company is one data point in that statistic. Which scenario materialises depends, in part, on what you do.

What to do about it in a company of 10 to 200 people

Four things that follow directly from the model.

  1. Count tasks, not headcount. Take the three most expensive positions and break a working week into tasks. Next to each: nothing, augment, automate, new task. After one such worksheet it is usually clear that two or three tasks are candidates for automation, not "a position". A task has a clear output and clear inputs. A position has relationships, exceptions and a customer calling at 4:58 pm. As a bonus, that worksheet records knowledge that currently lives in people's heads and leaves the company with them.
  2. Automate what has a clear output. Augment the rest. Data entry, orders, quotes from a price list, answering the phone after hours. Decisions, relationships and anything that requires taking responsibility stay with people, with AI as a tool.
  3. Put supervision in someone's job description. The report's "new task" is, in practice, an agent's action log, an alert when something fails, and one person who reviews it once a week. Without that, zero out of sixteen enquiries will happen to you too.
  4. Compute your own labour share. What percentage of revenue goes to office salaries, and what percentage to salaries on the floor or in the field? The model says the first number will fall and the second will rise. A company that sees this early plans raises and hiring differently from one that finds out from the market.

If it has to be a single decision, it is this: pick one process with a clear output, put AI into it this quarter, and write supervision into someone's duties. The rest will come from the data.

What this model does not say

Three caveats, so we do not overshoot in the other direction.

The model covers the United States. Europe has different wages, a different employment structure and different labour markets. The task mechanics are the same, though, and companies in Warsaw or Madrid buy the same models as companies in California.

The model has no robots. The authors note that they did not model a scenario with physical machines on the shop floor or in the home. Non-office occupations are safe in this model because it assumes so, not because it will be so.

The model does not say which scenario will happen. It shows that all three start from the same point and separate after 2027. The data will show where we are in a year or two. A company that waits for that data will be deploying under conditions where everyone else already has.

Book a 30-min call. We will take one process from your company and break it into tasks using the report's method: what to automate, what to augment, where supervision is needed. If you would rather read first, start with what AI for business means in practice, or with the numbers from twelve deployments.

Sources