Source: Anthropic, “How employment is changing at Anthropic” (ai-research/anthropic-employment-at-anthropic-2026-10-27.md) and “The AI expectations gap” (ai-research/anthropic-ai-expectations-gap-2026-10-27.md), both published 2026-10-16; AP on the Public Benefit Fund (ai-research/ap-anthropic-public-benefit-fund-2026-10-27.md); commentary from The Artificial Intelligence Show ep. 247 (raw/Ep._247_-_OpenAI_Safety_Concerns_Anthropic_s_AI_Expectations_Gap_MAICON_2026_Recap_Gemini_4_Argon.md); The Neuron Daily headlines (raw/newsletter-theneurondaily-com-df67e8a9fe.md, raw/newsletter-theneurondaily-com-4f8dd7da56.md).

Author: Anthropic Economics team (employment report: Nathan Wilmers, Regina Catipon, Maxim Massenkoff, Peter McCrory; expectations-gap report: Szymon Sacher, Tess Cotter, Charles I. Jones, Anton Korinek, Peter McCrory) URL: https://www.anthropic.com/research/employment-at-anthropic and https://www.anthropic.com/research/ai-expectations-gap Published: 2026-10-16 Podcast commentary: Paul Roetzer and Mike Kaput, The Artificial Intelligence Show ep. 247 (YouTube, https://www.youtube.com/watch?v=nvyNQmOTKfY, recorded 2026-10-19)

In the week of 2026-10-16 Anthropic published two reports about itself. One measures how hiring, output and daily work changed inside the company as agents took over more of the work. The other compares what 1,334 Anthropic staff and 10,980 US adults expect AI to do by 2030. Both are first-party data, not a forecast about other companies. They are still the most detailed public picture of an organisation run mostly on agents, and they give operators concrete numbers for workforce-planning conversations. Two days later Anthropic proposed putting 5% of its shares into a new Public Benefit Fund, partly to address labour-market effects.

Both fetched reports are tool-generated summaries of the Anthropic pages, not verbatim copies. Check exact wording on the live pages before quoting.

Key Takeaways

  • Agents now out-work their handlers. Combined agent work hours passed human hours in March 2026. By August, agents tied to individual employees logged about 6× their handlers’ hours. Kaput reports the comparison assumes a 40-hour human week.
  • Junior technical hiring collapsed in share, not in total. The junior share of technical hires fell from 27% to 9%, and the junior share of technical staff fell from 21% to 14% since early 2025. Junior postings fell sharply after December 2025, while technical hiring kept growing in absolute terms. The Neuron notes the report does not show whether the number of junior hires fell.
  • Managers report hiring less. About a third of ML managers and a fifth of software-engineering managers said they hired less because of AI (April 2026 survey). Technical departments fell from 61% to 56% of headcount.
  • Output per person rose. Commits per engineer reached about 5× their level six months earlier. Contracts reviewed, invoices processed and deals closed also rose. Coding-agent sessions by non-technical staff went from 2 a month (August 2025) to 90 (May 2026).
  • Human attention is the bottleneck. Reading all interactive agent output would take non-technical staff more than half their work week. About 80% of interactive session time is agents waiting for human input. The share of sessions spent managing agents, building tools and monitoring pipelines rose from about 10% to 22%.
  • Work has to be visible to agents. Private work in unlogged meetings and direct messages is hard for agents to see and is discouraged. Decisions are expected to be documented in the open.
  • Hiring criteria moved to judgment. Mentions of candidates’ AI tool use roughly tripled for technical roles and quadrupled for non-technical roles versus 2024. Interviewers weight autonomy and judgment over domain expertise.
  • The expectations gap is about the future, not the present. Staff and the public broadly agree on what AI can do today. On 2030 they split: 91% of staff vs 39% of US adults expect a Nobel-level discovery, and staff expect AI to do 75% of their current work hours against 28% for US workers.
  • Anthropic says its own fixes are not enough. It is building an internal early-warning system and redeployment process, but says these “will not prevent” AI job disruption or the junior-hiring slowdown. It calls for government retraining and income support.

The employment report

Method. Internal headcount, AI-use, output, survey, Slack, calendar and hiring data, plus 17 employee interviews. The productivity comparison uses 192 engineers who adopted coding agents against 74 who did not, with person and month fixed effects. Each session’s agent output tokens were converted to words at average reading speed to estimate review load (Roetzer reads this method out on air).

Findings not listed above:

  • Administrative headcount grew more slowly than the company. Roetzer reads the report’s phrase that agent assistance, automated review and ticket-service outsourcing have left Anthropic “clerical light”.
  • Median meeting time rose about 20% year over year (16% for technical individual contributors).
  • Agents appeared in nearly half of active public Slack channels, mostly dedicated human-agent channels.
  • Written standing instructions for agents grew from a few hundred (July 2025) to about 8,500 a year later.
  • Long-tenured employees were about 50% more likely to move across departments each month in 2026 than in 2025.
  • The share of employees who said their work processes made them as productive as possible rose from 44% (late 2024) to 58% (November 2025).
  • Read on air from the report: “humans are a thin layer on top [of] a growing contingent of AI agents”. Engineers “describe becoming leaders and managers of agents”. For non-technical staff, the strategic “20%” of the job “becomes 90%”. Interpersonal work such as sales calls and interviews stays human. One engineer described “an army of Claudes with everyone asking for my attention”.
  • Per Roetzer’s reading, at five years old Anthropic takes in 2 to 4 times the revenue per employee that Meta and Google did at the same growth stage. Epoch puts the current figure at ~$9M per employee (see Compute Economics).

Caveats the report states. Tool adoption was not random, so estimates may be biased. The data covers one company during rapid growth. Code volume is not business impact, and the quality proxies are imperfect.

Recommendations. Shared ownership through employee stock, an early-warning system that reads role-level demand and productivity against AI capability and use, and internal redeployment. The authors note that redeployment commitments could make employers reluctant to hire, and argue that prioritising redeployment pushes firms toward creating new tasks instead of “so-so” automation.

The expectations-gap report

Samples. 10,980 US adults (Morning Consult, 2026-08-11 to 08-23, weighted, online opt-in) and 1,334 Anthropic staff (internal, unweighted, 2026-08-09 to 08-12). A September re-survey of 7,613 adults moved the Nobel figure from 39% to 44%.

Question (2030 unless noted)Anthropic staffUS adults / workers
Nobel-level discovery by 203091%39%
AI can design, run and publish studies by next year9 in 1038%
Share of own current work hours AI could do75% average; over half expect 80% or more28% average; about 1 in 20 expect 80% or more
Tasks AI could do that will be done with AI65%42%
AI-suited task takes half the time or less with AI85%typical answer is half; 29% of AI-using workers say as long or longer
Says some task will never be automatableat most 2%19% to 40%, by task
Model-implied 2030 GDP growth (median)17.1%5.4%

Other findings:

  • US workers who expect AI to do 80% or more of their work forecast 12.2% unemployment in 2030, above the Great Financial Crisis peak. Each extra 10 points of work affected adds about 0.4 points to the forecast, for staff and public alike.
  • Regular AI users at work rank data privacy as the top obstacle (42%). Others rank customers preferring a person first.
  • Median staff estimate of the share of frontier-model-building work done by AI: 51% today, 80% by end of 2027, 95% by end of 2030. The report cites Anthropic’s figure that over 80% of merged code is written by Claude.
  • The report makes no policy recommendations. Its caveat: staff and public use very different amounts of AI at work, which may explain part of the gap.

The Public Benefit Fund (2026-10-18)

  • AP: Anthropic committed 5% of its shares to a new Anthropic Public Benefit Fund, “a new entity within the company”. Examples of uses are “large donations to traditional nonprofits or direct distributions to specific populations”, with labour-market impacts named as one area. Anthropic has already committed $200 million to studying AI’s effects on the labour market and says “bolder approaches may be needed”. With earlier pledges, roughly 20% of shares are set aside for the “public good”.
  • Read on air from Anthropic’s post (Ep. 247), and repeated by The Neuron: the commitment is subject to stockholder approval, with an additional 2% at the board’s discretion “for every 50% increase in the company’s value”. Anthropic’s own post (anthropic.com/news/public-benefit-fund) was not fetched.
  • The fund’s structure, investor reaction and IPO context are in Anthropic’s Position.
  • Roetzer’s objection (opinion): the labs “just assume this all works out” if they “give everybody money back”, and the reports never discuss pacing the impact on workers.

Try It

  1. Make decisions agent-visible. Log decisions where your agents can read them: shared channels, written meeting notes, a decision doc. Anthropic discourages unlogged meetings and DMs for exactly this reason. Roetzer links it to Zapier’s public-Slack practice and to Tobi Lütke’s line: “If every interaction with an agent happens in a private window, the only person who learns anything is the person at the keyboard.”
  2. Size review load before adding agents. Copy Anthropic’s metric: convert a week of agent output tokens to words and divide by reading speed. If the result is more than a few hours per person, the next investment is dashboards, monitors and standing instructions, not more agents.
  3. Use the junior-hiring numbers carefully in client conversations. 27% → 9% is a share of hires at one fast-growing AI lab. It says nothing about total junior headcount, and the report itself calls the setting unusual.
  4. Run Kaput’s prompt. Load the employment report into your AI tool and ask: “if even a fraction of this became true, what does that mean for my career, my company, and our people”.
  5. Plan for the expectations gap with clients. Customers who don’t use AI at work expect far less change than people who do (the table above). Set expectations from your own usage data, not from either group’s forecast.

Open Questions

  • Is Anthropic a leading indicator or an outlier? Roetzer guesses frontier-adopting companies are “probably 18 months behind” what Anthropic describes. That is opinion. The report itself says the data covers one company in rapid growth.
  • What counts as “junior”? Secondary coverage notes the report does not define it precisely, and no raw count of junior hires is given.
  • Fund governance. Who decides grants, what the baseline valuation is, and when stockholders vote are not in the sources fetched.
  • Verbatim wording. Both reports were captured as fetch-tool summaries. The “thin layer” and “clerical light” phrases are confirmed only through the podcast reading.