Product1 publisher3 min readPublished
Anthropic's 2030 model makes the planner fill in the adoption number
The Anthropic Institute's three scenarios for 2030 run from half a point of extra GDP growth to 15% a year. The interactive model behind them makes you supply the missing variable, which is how much of what AI can do actually gets used.
The Product Desk · Product desk
What happened
- Researchers at The Anthropic Institute published a framework and an interactive model for how AI affects jobs, unemployment and GDP growth to 2030, built around three scenarios they call modest, substantial and extreme.
- In the modest case AI adds less than half a point to US GDP growth by 2030 and lifts unemployment by a tenth of a point, a change the researchers say would be hard to find in macroeconomic data.
- The substantial case has AI capable of half of all knowledge work by 2030, most of it autonomous, with the economy growing at twice its normal rate while knowledge-worker wages stay flat.
- The extreme case puts GDP growth at 15% a year with nearly one in five cognitive workers unemployed and their relative wages falling immensely.
Compiled by The Product DeskSomething wrong?How this is made
Why it matters
- constraint A band this wide cannot size a 2027 headcount plan, and because its ends are measured against different workforces it can only be used to test which commitments survive all three worlds.
- decision The model moves the planning question off what AI can do and onto what share of eligible work an organisation will actually let it do, and that share is a number the buyer sets rather than the vendor.
- exposure If the extreme case lands, the researchers say the money to cushion displaced cognitive workers exists but its distribution becomes a policy question, which puts the worker's downside outside any employer's plan.
- precedent The researchers' own claim that the picture clears within a year or two sets a short expiry date on any strategy justified by a single scenario.
The tool's central field asks how many of every 100 instances of a task AI can do in 2030 it will actually be doing [15]. Anyone who has shipped a rollout knows the answer is not 100, and the middle scenario rests on that gap: it holds that most knowledge work tasks would still be completed without AI in a 2030 where the capability to do them exists [7]. Capability is set by the lab. The share of eligible work that actually moves is set by whoever owns the process, the review step, and the person who signs off.
The two ends of the band are also not measured against the same population. The tenth of a point in the modest case is unemployment across all workers; the near one in five in the extreme case is cognitive workers only [17]. There is no dial to slide between them and read off a middle number, which is presumably why the model asks the user to construct the middle themselves [2].
Scale helps here. The researchers put the value of every task performed in the US last year, by people, machines and software alike, at roughly $30 trillion [5]. Run their two outer growth figures against that base and the extreme case adds on the order of $4.5tn of output a year while the modest case adds around $150bn, a thirty-fold spread in the same currency [20]. Rough arithmetic, since the base itself grows, but it is the honest width of what a frontier lab's own institute is willing to put in writing, alongside the statement that the outcome is "extraordinarily uncertain" [3].
The survey underneath the work is a measure of expectations, not of outcomes. Roughly 11,000 Americans were asked to predict AI use, productivity gains, automation versus augmentation, and displaced work [12]. Most landed near the substantial scenario, with GDP 10% higher by 2030 than it would otherwise be and unemployment around 5% [13]. About one in ten held views consistent with the extreme case [14], which on that sample is roughly 1,100 people [16]. That minority matters less as a forecast than as a fact about the workforce a rollout has to persuade.
Here is what the model is useful for on Monday, which is not sizing a 2027 headcount. Take each process you own and mark two things separately: whether a 2030 model can perform the task, and whether your organisation would let it run without a human signature. Work that scores yes on both is where the substantial scenario's productivity actually shows up. Work that can be automated but will not be is where the money goes into review steps rather than reduced headcount, and the middle scenario says that quadrant is the larger one [7]. Commitments that only pay off in one quadrant are bets on a scenario; commitments that hold in both are decisions you can defend either way.
The researchers say which world we are heading toward "may become clearer within a year or two" and that preparing for disruption is the prudent course [4]. That is a short review clock on any plan built on one column, and it is the part of the publication a planning committee can actually act on.
What to watch
- Whether the Anthropic Institute releases the full distribution of the 11,000 responses rather than the clustering around the substantial scenario.
- The researchers' own one-to-two-year visibility window: which of the three scenarios 2026 and 2027 data rules out.
- Whether the tool's per-100 adoption question starts appearing as a stated assumption in corporate planning documents.