Anthropic Staff Expect AI to Do 75% of Their Work by 2030; US Workers Say 28%

TL;DR

On October 16 Anthropic's economics team published The AI Expectations Gap, which sets a Morning Consult survey of 10,980 US adults (August 11-23) beside an internal poll of 1,334 Anthropic employees (August 9-12). Both groups agree on what AI can do today. They split hard on 2030: staff expect AI to be able to do 75% of their current work hours, US workers say 28%, and 91% of staff expect a Nobel-level discovery against 39% of the public. A companion report, How Employment is Changing at Anthropic, shows what that belief looks like on a payroll: employee-directed agents logged six hours for every human hour by August, junior roles fell from 27% to 9% of technical hires, and about 80% of an interactive Claude session is the model waiting for a person.

Two polls, one question set

The authors (Szymon Sacher, Tess Cotter, Charles I. Jones, Anton Korinek and Peter McCrory) ran the same questions past two very different rooms. Morning Consult fielded the public survey as an online opt-in panel, weighted to be nationally representative, with half the respondents asked about 2027 and half about 2030. The staff poll was internal, unweighted, and asked everyone about 2030. A September follow-up with 7,613 US adults nudged the public's Nobel number from 39% to 44%, but the paper reports the August figures throughout.

Capability questions covered eight tasks, each phrased as "as well as a skilled professional," from drafting routine emails up to a Nobel-level discovery. On the present, the two groups rank the tasks in almost the same order, and about a quarter of each group thinks AI can already design, run and publish a scientific study. The disagreement is entirely about the slope. By 2027, about nine in ten staff expect that study-publishing capability; 38% of US adults do. Depending on the task, 19% to 40% of US adults say AI will never manage it. Among staff that figure tops out at 2%.

That last chart is the most useful one in the paper for anyone selling AI tooling. The public is not one block. Roughly a fifth of respondents use AI regularly at work, and 56% of them expect AI to clear seven of the eight tasks by 2030, double the share among people who have never used it. Exposure moves expectations, which is either evidence that users see something real or evidence that familiarity breeds optimism. The paper, to its credit, does not pick.

Your job, as they see it

The headline number comes from asking people what share of their own current work hours AI could do by 2030. Staff averaged 75%, and more than half said 80% or more. US workers averaged 28%, and about one in 20 said 80% or more. Split by occupation, manual trades, production and transport workers said 25%, computer, engineering and science workers said 33%, and the subset of those technical workers who use AI regularly said 40%. Even the most exposed, most AI-fluent slice of the public sits 35 points below the lab.

The unemployment question is where the two groups suddenly agree. Respondents were told the current US rate is about 4.2% and asked for 2030. Workers who expect AI to do less than half their job forecast 7% to 8%. Above that line the forecast climbs with exposure: people who expect 80% or more of their hours automated put 2030 unemployment at 12.2% on average, above the peak of the financial crisis. Every extra 10 points of expected exposure comes with about 0.4 points more forecast unemployment. Anthropic staff show the same gradient, at much higher exposure, and land in the same place as the public once you line them up on that axis.

They are less worried about their own employer. 95% of staff expect Anthropic's headcount to grow over the next 12 months, with a median guess of 66% growth and 27% expecting it to at least double. The paper is careful to call these personal views rather than company plans, which is fair, and it is also the answer you get when you ask people whether their employer should hire more of them. Asked how long a worker displaced in 2030 would need to find a job in a new occupation, staff put the median at 12 months and the public landed in the 7 to 12 month band, against a normal-times baseline of about three months.

What it looks like from inside

The companion report, by Nathan Wilmers, Regina Catipon, Maxim Massenkoff and Peter McCrory, pulls Anthropic's internal employment, usage, Slack, calendar and hiring-scorecard data, plus 17 employee interviews, and treats the company as a leading indicator. Whether a five-year-old AI lab with 2 to 4 times the revenue per employee that Meta and Google had at the same age tells you much about a regional accounting firm is a fair question. The numbers are still a rare look at what heavy agent use does to an org chart.

The agent-hours figure is the one that will get quoted. Counting only interactive, employee-directed sessions (subagents pooled, waits over 30 minutes excluded, humans at 40 hours a week), agent hours first passed human hours in March 2026 and reached six to one by August. About 80% of that interactive session time is the model waiting for a person to answer it, so the bottleneck has moved from generation to attention. One engineer described "an army of Claudes, with every one asking for my attention." The report estimates that reading the interactive Claude output of a non-technical employee would take over half a work week.

Junior hiring is the uncomfortable part. Junior roles fell from 21% to 14% of technical departments since early 2025 and from 27% to 9% of technical hires, with postings dropping sharply after December 2025. By April 2026 about a third of ML managers and a fifth of software-engineering managers said they were hiring less because of AI. Interview scorecards tell the same story from the other side: comments about autonomy and judgment grew, comments about domain expertise shrank as a share, and mentions of a candidate's track record with AI tools, about 5% of technical assessments in 2024, tripled.

Output went up, with the usual asterisks. Commits per engineer hit 5x their level from six months earlier. A within-person comparison of 192 engineers who adopted coding agents against 74 who did not shows code added rising sharply after adoption, and for code by the same person in the same month and directory, agent-written and human-written lines were removed or reverted at the same rate. The authors also note that adopters were already on a steeper trajectory before they adopted, that more code is not more business impact, and that the company was growing explosively at the same time. Sales reps who adopted agents from September 2025 closed deals with more clients; one said automations replaced about 20% of their prior tasks. Non-technical staff went from 2 coding-agent sessions a month in August 2025 to 90 in May 2026, and "meta-work" (sessions about managing other sessions) rose from 10% to 22% of interactive use.

The GDP number, and how to read it

The paper's final move is to feed each group's median answers through the growth model from the Institute's earlier Economic Scenarios for Transformative AI, which takes five inputs (capabilities, diffusion, automation, productivity, and how hard it is for displaced workers to move) and holds everything else at its "Substantial" setting. Think of it as a crop-yield calculator: you type in your guesses for rainfall and soil, and it hands back bushels per acre. The output is only as good as the guesses, but it turns a vague "AI will be big" into a number you can argue with.

The median US adult's answers imply 5.4% growth in 2030, which happens to match the model's "Substantial" scenario and is already more than double the typical US rate. The 23,075 people who completed the public scenario explorer in its first 48 hours, from more than 140 countries, land at 12.5%. Staff medians imply 17.1%, past the model's own "Extreme" scenario, which the authors describe as transformative and historically unprecedented. The distributions overlap widely, and the paper shows 10th to 90th percentile bands rather than pretending otherwise.

Staff were also asked what share of frontier-model building AI already does. The median answer was 51% today, 80% by the end of 2027 and 95% by the end of 2030, with research-team members at 60% today versus 53% for everyone else. A footnote adds that more than 80% of code merged into Anthropic's codebase is now written by Claude. These are perceptions, not measurements, and the paper says so.

Caveats, straight

The public sample is an opt-in online panel, so it excludes people without internet access, and weighting cannot fix every way it differs from the population. The staff poll is unweighted, internal, and drawn from people who are both closer to the technology and financially exposed to its trajectory. The employment report's authors work for the company they are studying, the adoption analysis cannot fully separate agent uptake from the company's growth, and the junior-hiring relationship is described as descriptive rather than causal. Anthropic is an AI-native firm, and the report says outright that many of its practices do not transfer economy-wide. Finally, the GDP figures are model outputs under fixed assumptions, not forecasts.

Why a builder should care

Price your product for the 28% room, not the 75% one. Most of your customers' employees expect AI to take a quarter of their job by 2030, say customer preference for a human is the top obstacle to adoption, and rank data privacy second. The one in five who already use AI at work put privacy first, at 42%. The internal data gives you a preview of what heavy use does to a team: six agent hours per human hour, 80% of it idle waiting for review, a reading load that outruns the week, and managers whose commit counts doubled. If you are building agent tooling, the scarce resource in that picture is human attention, and the product that wins is the one that spends less of it.

Key Takeaways

• Anthropic staff expect AI to be able to do 75% of their current work hours by 2030; US workers put it at 28%. More than half of staff said 80% or more, versus about one in 20 workers.
• Both groups agree on present capability and rank tasks in the same order. The gap is about speed: 91% of staff expect a Nobel-level discovery by 2030, against 39% of US adults (44% in a September follow-up).
• Expectations track exposure. 56% of US adults who use AI at work expect it to clear seven of eight tasks by 2030, twice the share among never-users.
• Inside Anthropic, employee-directed agent hours ran six to one against human hours by August 2026, with about 80% of interactive session time spent waiting on a person.
• Junior roles fell from 27% to 9% of technical hires. Commits per engineer rose 5x, with no measured difference in revert rates between agent-written and human-written code.
• Fed through Anthropic's growth model, the public's median answers imply 5.4% US GDP growth in 2030; staff answers imply 17.1%. Both are model outputs, not forecasts.

Sources: The AI Expectations Gap (Anthropic Institute), How Employment is Changing at Anthropic, Economic Scenarios for Transformative AI explorer, TechBooky, IBTimes