Anthropic published two workplace studies on October 16, 2026. One measures expectations; the other examines work inside the company. Their useful distinction for business readers is between what people believe AI will do and what an organization has actually observed. Read the expectations report and the employment study.
What the survey measures
In August, surveyed Anthropic staff expected AI could perform an average of 75% of their current work hours by 2030. Employed US respondents estimated 28%. These are beliefs about future capability, not measured automation or projected job losses. Source: survey findings.
The public survey reached 10,980 adults through an online opt-in panel and used demographic weighting. The internal poll drew 1,334 staff and was unweighted. Public respondents selected ranges for work-hour exposure; staff used a slider. Those differences matter when interpreting the comparison. Source: methodology and Figure 5.
What the workplace study observes
The companion study reports rising output after employees adopted agents, alongside pressure on human attention and a declining share of junior technical hires. Adoption was not randomized, and the authors explicitly discuss selection bias. A frontier lab is not a representative employer. Source: employment study.
How to use this evidence without overreading it
Our interpretation: neither report should become a staffing target. A percentage of work hours is not a percentage of removable roles. Before acting, write down what would have to be true for a forecast to apply to your organization.
For example, consider an invoice-triage pilot. Define the allowed task narrowly: categorize incoming requests and propose a routing decision. Keep payment authorization outside its scope. Compare completed, correctly routed cases against the existing process, rather than counting generated responses.
Include the review burden. Record the time needed to inspect suggestions, correct errors and resolve exceptions. Decide in advance what error level would stop the trial and who owns the final decision. This is a proposed evaluation approach, not a result established by Anthropic's research.
The decision to make now
Choose a reversible workflow experiment instead of committing to a 2030 headcount scenario. Preserve a baseline, document failures and ask whether the new process improves accepted output after oversight costs. If it does, expand the task boundary deliberately. If it merely transfers work to reviewers, redesign the process before scaling it.
The practical question is not which group has the better forecast. It is what evidence your own team needs before changing responsibilities.