Anthropic’s October 16, 2026, workplace study reports that about 80% of elapsed time in the company’s interactive, multi-turn Claude sessions was spent waiting for human input. The estimate describes those sessions at Anthropic, not a representative sample of US employers.
What the 80% figure measures
Anthropic calculated the waiting time by subtracting an agent’s active working time from the session’s total elapsed, or wall-clock, time. The result is a share of interactive-session time, not a measure of how much time an agent spends actively working.
The study also gives a separate measure: by August 2026, employee-directed interactive agents logged runtime equal to six times employees’ work hours. For that calculation, Anthropic counted inference and active tool use, excluded waits longer than 30 minutes, and assumed employees worked 40 hours per week. This runtime-to-human-hours ratio uses a different basis from the 80% session estimate.
How Anthropic measured meta-work
Anthropic analyzed company records on employment, AI use, output, surveys, communications, calendars and hiring decisions, alongside 17 employee interviews. The study describes Anthropic’s own AI-intensive workplace.
Anthropic reported that meta-work rose from 10% of the average employee’s interactive Claude sessions in fall 2025 to 22% in spring 2026. The company used automated classification on a sample of 2,000 one-line chat summaries. The percentages refer to the share of sessions classified as meta-work, which includes preparing agent instructions and plans, reviewing work, monitoring pipelines, building dashboards and documenting processes.
These activities surround the task an agent is assigned. Employees still need to set up work, monitor progress and organize results, so greater agent output does not automatically remove the human work involved in managing it.
What the coding comparison can show
Anthropic followed 192 engineers who adopted coding agents and 74 who did not. Its within-person analysis found that adopters added more code in the months after adoption. But adoption was not random, and adopters were already increasing output faster. The comparison therefore does not establish that AI alone caused the increase.
Anthropic also found no difference in rates of code lines later removed or reverted in its comparison. Those are limited quality measures: the company notes that code quality has multiple dimensions and that AI errors may be harder to identify. More code, on its own, does not establish better code or greater business impact.
