Anthropic Modeled the Economy. One Dial Is Yours.
Anthropic's economic model runs on two inputs: how capable AI gets, and how widely it's used. Nobody outside a handful of buildings touches the first. Every company touches the second. And in Anthropic's own middle scenario, AI can do half of knowledge work by 2030 and most of that work still isn't done that way.
- anthropic
- economics
- ai-adoption
- market-structure
- workflow-discovery
In Anthropic's middle scenario, AI can do half of all knowledge work by 2030, most of it without a person involved, and most of that work still gets done the old way.
That sentence is carrying more than it looks like. The capability arrives. The usage doesn't follow. And the distance between those two facts isn't a rounding error in the model, it's the thing the model is about.
Anthropic's Economics team published its Econ Scenario Explorer in September, alongside a technical report by Anton Korinek, Chad Jones, Szymon Sacher, Tess Cotter and Peter McCrory. The approach is to treat every job as a bundle of tasks, borrowing the US Department of Labor's O*NET taxonomy, and to ask of each task what AI does to it: nothing at all, helps a person do it faster, takes it over, or creates a new task that didn't exist before. Add up every instance of every task across the country and you have the $30 trillion the US economy produced last year.
Then the model hands you two inputs. How capable AI gets. How widely it's used.
Only one of those is yours
Nobody reading this sets the first one. Frontier capability gets decided in a handful of buildings by a few thousand people, and the most any company can do about it is read the release notes and adjust.
The second is different. Adoption is the sum of thousands of small decisions made inside ordinary companies by ordinary managers, one workflow at a time. It's the only input to the entire model that anyone outside those few buildings actually touches.
It's also the one Anthropic doesn't try to explain. The explorer takes adoption as a number you supply, not an outcome it predicts. That's a defensible choice for a model working at national scale, and it leaves the interesting question sitting there: if AI can do half of knowledge work and half of knowledge work still isn't done that way, what's in the way?
By 2030, in that scenario, it isn't capability. It isn't price either. What's in the way is the step nobody counts as work. Somebody has to be able to say what the job consists of. Which tasks, in what order, triggered by what, handed to whom, failing how often and in what way. O*NET does this for occupations in the abstract, which is why Anthropic could build the model at all. Almost no individual company has done it for itself.
The scenarios differ on use, not on models
Worth being precise about what separates the three futures, because it isn't model quality in any simple sense.
In the modest scenario, AI's effect resembles the internet's: real gains, arriving gradually, inside the historical range for new technology. In the substantial scenario, the economy grows at twice its normal rate, wages for knowledge workers stay flat, and workers outside knowledge work gain. In the extreme scenario, GDP growth reaches 15% a year, the economy doubles every four and a half years, knowledge-worker wages fall by more than 10% by 2030, and unemployment goes past what recessions normally produce.
Anthropic attaches no probabilities to any of them and says plainly they aren't predictions. It also published its reviewers' objections, which is more than most vendor research does. Several of the economists who read the draft, a list including Daron Acemoglu, David Autor and David Romer, thought the extreme case reads better as a thought experiment, while others thought the modest one understates what's already visible in the data. None of them were asked to endorse the conclusions.
The obvious thing is worth saying too. Anthropic sells the first dial. A company whose revenue depends entirely on capability has published a model in which adoption does most of the work of deciding the outcome. Read that as a point in the model's favor.
The median American already expects the gap
In August, Anthropic surveyed more than 10,000 Americans with Morning Consult about present and future AI capability and adoption. The typical respondent's answers imply something close to the substantial scenario: GDP about 10% higher by 2030 than it would otherwise have been, unemployment around 5%. Roughly one in ten hold views consistent with the extreme case.
So the median expectation is already the one where what AI can do runs well ahead of what it gets used for. That isn't pessimism about the technology. Anyone who has watched a company buy an enterprise license for a tool that four people open has run the small version of this study.
Your company has a scenario too
The model is national, and Anthropic is clear about what it leaves out: policy responses, business cycles, demand effects from the data center buildout, capable robots, and any view of what happens to individual workers. Fine. The structure still transfers.
Your company has the same two dials, and its position on the first is the same as everyone else's. You'll have access to roughly the same models as your competitors, at roughly the same time, at roughly the same price. The place a difference can open up is the second dial: what share of your work you can actually point AI at.
That share has a ceiling you've already set, whether or not you know the number. It's the proportion of your operation that someone could describe precisely enough to redesign. Above that line the work is unreachable, not because the model can't do it, but because nobody can say what it is.
Which is why so many AI programs finish smaller than the budget implied. A company that can't describe its own work commissions a strategy, receives a deck built from its most legible use cases, ships three of them, and lands in the modest scenario while believing it's in the substantial one. The models kept improving the entire time. That was never the constraint.
Anthropic built a tool that lets you forecast the economy, which is the one system you have no control over. There's no equivalent for the system you do control, and the reason is that it would need an input most companies don't have: a written account of what the place actually does all day.
By 2030 the first dial will have moved whether you did anything about it or not. The second one won't.
If you want that question answered for your specific situation, the Forge Playbook does it. Answer a few questions about your business and we'll put together a tailored outline of which workflows are worth automating and what a realistic budget looks like for each. Free, no obligation, takes about three minutes.
