On October 16, 2026, The Anthropic Institute released a dual research report titled 'The AI Expectations Gap'. Authored by Szymon Sacher, Charles I. Jones, Anton Korinek, Peter McCrory, and Nathan Wilmers, the study compared survey results from 10,980 representative US adults with an internal census of 1,334 Anthropic employees. The survey findings were matched against internal telemetry and productivity metrics tracking the lab's day-to-day operations. The resulting data highlights a vast divide between how frontline builders and the wider workforce perceive the pace of technological change.
The divergence is immediately apparent in current productivity experiences. Approximately 59 percent of Anthropic staff reported that AI allows them to complete specific workplace tasks in one-quarter of the baseline time or even faster. In contrast, only 26 percent of regular AI users across the broader US workforce experienced comparable speedups. Furthermore, 29 percent of regular workers reported that performing tasks with AI currently takes the same amount of time or longer. Among Anthropic's own technical staff, only 5 percent reported such friction.
Looking ahead to 2030, expectations diverge even more dramatically regarding job automation. Anthropic staff projected that automated systems will handle an average of 75 percent of their current work hours by the end of the decade. More than half of internal respondents anticipated that at least 80 percent of their responsibilities will be fully automated. In contrast, typical US workers estimated that only 28 percent of their own workload would be taken over by AI. Only 5 percent of the broader public believed that 80 percent or more of their work could be automated within that timeframe.
The accompanying workplace report documented the rapid shift toward agentic organizational structures inside Anthropic. Since March 2026, autonomous software agents have collectively logged more working hours than the lab's entire human workforce. By August 2026, the deployment expanded to an average ratio of six active software agents for every human employee. This operational shift demonstrates how autonomous tools are moving from experimental aids into the core engine of technical production.
As routine creation automated, the operational bottleneck within the company moved decisively from generation to oversight. Drafting documents, generating software code, and formulating analysis are no longer the primary constraints. Instead, during interactive sessions, Anthropic employees spent roughly 80 percent of their time reviewing, directing, and validating agent outputs. This dynamic, dubbed the human attention bottleneck, has established human judgment and error correction as the primary ceiling on output velocity.
Finally, these contrasting operational realities translate directly into macroeconomic forecasts. Both surveyed cohorts linked intensive corporate automation with downward pressure on employment, but their specific projections differed significantly. Anticipating widespread automation, Anthropic employees forecast a US unemployment rate above 12 percent by 2030. In contrast, the average US citizen anticipated an unemployment rate of around 8 percent. The findings underline that direct immersion in advanced agentic workflows fosters an outlook profoundly different from that of the general population.