Anthropic's 2030 Economy Scenarios: Half the Economy Grows, Half the Payroll Is Exposed

Last Updated: 2026-09-12

Anthropic's economics team has published a model of what AI does to the US economy by 2030, and its three scenarios span almost the entire argument: GDP between 1.6% and 32% higher than otherwise, and knowledge-worker unemployment anywhere from unremarkable to nearly one in five. The consistent finding across all three is a split: the disruption concentrates in screen-based cognitive work, while wages for physical and hands-on occupations rise in every scenario. It is a planning tool, not a forecast, and the paper says so itself.

What Anthropic Actually Published

Why does one company's model deserve attention alongside the forecast noise it arrives into? Because most AI-economy claims are a number without a mechanism. This one is the opposite: a fully specified task-based model, published on 9 September 2026 with a companion interactive explorer, that maps a small set of assumptions about AI capability, adoption and autonomy into paths for GDP, wages, employment and labour's share of income between 2026 and 2030.

According to the paper, every occupation is treated as a bundle of tasks, and for each task AI can do one of four things: leave it untouched, augment the person doing it, automate it outright, or create new work around it. According to Euronews's coverage of the release, the team built the economy as bundles of tasks in this way precisely so different assumptions could be compared on the same machinery, rather than argued as slogans.

The authors are explicit about the limits, and it matters that they are. The three scenarios carry no probabilities and are not predictions; their purpose is to make the consequences of different assumptions comparable. If someone quotes a single headline figure from this work as "what will happen", they are quoting the paper's own output against its stated intent.

The two islands the model runs on

The model sorts all work into two groups, and the paper's own examples make the split vivid. The cognitive group covers management, professional, sales and office work, the tasks AI directly affects. The all-other group covers manual and interpersonal work, construction workers and electricians, which AI does not touch directly. Almost every interesting result in the paper comes from what happens at the boundary between these two islands: displaced cognitive workers must search for jobs in the other group, switching occupations is slow and hard, and that friction is where unemployment appears in the model.

The Three Scenarios in Numbers

What do the scenarios actually say? The modest case has AI adding less than half a percentage point to GDP growth by 2030, leaving GDP just 1.6% higher than it would otherwise be, with unemployment up by barely a tenth of a point. The substantial case has AI capable of roughly half of all knowledge work by 2030, GDP 8.3% higher, and cognitive employment down about 4%, with unemployment among those workers around 4.5%. The extreme case is without historical precedent: annual growth of about 15%, GDP 32% higher, and the labour share of national income falling from 60% to 45%.

The wage divergence in the extreme scenario is the starkest single result in the paper. Cognitive wages land 11.5% below where they would have been without AI, while wages in all other occupations finish 34% above theirs. Average pay rises across the economy in every scenario; what changes is who receives it.

Who Gains and Who Is Exposed

Why do the physical occupations come out ahead in a report about artificial intelligence? Because in the model, AI raises productivity in cognitive work, which makes the projects and services that depend on it cheaper and more plentiful, and those projects still need people to build, install, maintain and staff them. Demand for the untouched group therefore rises along with output, and so do its wages, in all three scenarios.

The exposed side is exposed by task shape rather than by seniority or salary. What the two groups share is a distinction between work done on screens and work done in the world: a business whose output is documents, analysis, tickets, code or correspondence sits in the cognitive group, whatever it sells, and a business whose output is a finished roof, a wired building or a cared-for patient does not. According to the paper, a software engineer may find it difficult to retrain as an electrician, and the difficulty of that move, not the displacement itself, is what generates the unemployment figures above.

The divergence arrives after 2027

The timing detail is easy to miss and useful for planning. Across all three scenarios, almost all of the divergence arrives after 2027. The model describes a transition that is gradual at first and then steep, which is a very different thing to plan against from a step change on a fixed date.

What Americans Already Expect

Where does public expectation sit against these scenarios? Anthropic surveyed US adults alongside the modelling, and according to the paper the median respondent's answers line up with the substantial scenario: GDP roughly 8% higher by 2030, cognitive employment down about 4%. The public, in other words, is not priced for the extreme case; the typical expectation is a materially changed but recognisable economy, not a discontinuity.

That convergence between a model and the people it models is unusual, and the sensible reading is not that the public has validated the model. It is that the middle scenario, the one where AI can do half of knowledge work but adoption lags behind capability, is the assumption most people already live their hiring and investment decisions by, whether or not they have ever seen an economic model of it.

Capability Is Not the Same as Adoption

The quiet core of the paper is the gap it draws between what AI can do and what actually gets used. The substantial scenario is defined by exactly this lag: AI is capable of half of knowledge work autonomously, yet the majority of tasks are still performed without it. The extreme scenario closes that gap, and every dramatic number in the report lives on the far side of the closing.

For a business owner, this reframes the question the report is usually summarised as answering. The interesting variable between now and 2030 is not which scenarios are right about capability, something no owner controls, but how quickly organisations like theirs actually absorb the capability they already pay for. The model's frictions are occupational retraining and search; the frictions on the ground are messier, and mostly internal: which tasks are documented, who owns them, and whether anyone can tell whether a change to them worked.

One of those frictions has nothing to do with willingness, and it is one we run into in conversation with owners. A business hands its IT to a managed service provider, the provider covers email, licences, devices and the helpdesk, and AI enablement is simply not in the service scope. The owner has budget, appetite and a clear idea of which work is wasting hours, and no route to any of it through the vendor relationship they already have. Nothing in the model captures that, because it is not a retraining problem or a search problem: it is a coverage gap, and the useful question for an owner reading this report is a narrow one, which is what their own provider's contract actually covers before they conclude AI is not available to them.

What an Owner Can Usefully Take From It

The honest use of this report is smaller and more practical than its headlines. It is evidence, from a source with a stake in the technology and the candour to label its own output as scenario planning rather than prediction, that the work most likely to be reshaped by 2030 is task-shaped: bundles of screen-based, repeatable cognitive work, exactly the bundles a small business already has a name for internally.

That task shape is the part an owner can act on, and in our experience it is not something anyone can hand over as a list. In discovery workshops the list gets built in the room, by walking through a normal week task by task, and the useful separation appears about halfway through: the repetitive screen-based bundles on one side, the work that genuinely needs a person on the other. The same conversation tends to surface two things the owner did not come in to discuss. The first is where the money is actually going, because putting tasks into categories makes visible which of them move value through the business and which are dead wood quietly eating the margin. The second is how much of the process depends on one person's memory, which is usually the founder's.

What we consistently see is that the categorising is worth doing whether or not any AI follows it. Hearing the people who run a process describe where it snags, first-hand and in the same room, is often the first time an owner sees the whole chain end to end rather than the parts that reach them. It also changes who adopts what comes next: the people involved in that conversation early are the ones who tend to champion the agents afterwards, and the ones handed a finished system tend not to.

An owner who has been through that, and has documented which workflows are the repetitive screen-based ones, who owns each, and what each currently costs in hours, is holding the only asset this report cannot supply: a baseline for their own business, against whichever scenario arrives.

Sources

• Anthropic Economics Team (Korinek, A. et al.), "Economic Scenarios for Transformative AI", published 9 September 2026. https://www-cdn.anthropic.com/files/4zrzovbb/website/cf58f84d46a4a76bf5a5b039ac695fba6b80041c.pdf
• Euronews, "Anthropic says AI could bring both 15% growth and mass unemployment by 2030", 11 September 2026. https://www.euronews.com/business/2026/09/11/anthropic-says-ai-could-bring-both-15-growth-and-mass-unemployment-by-2030

Background reading, qualitative framing, no figures: i-scoop, "How AI could shape economic growth in 2030 according to Anthropic", 10 September 2026.