A new survey from Epoch AI and Ipsos finds that 20% of employed US adults now delegate at least one work task to AI that was previously handled by coworkers or outside contractors. The finding points to a quiet shift in how work gets done, one that is more about redistribution than replacement.
The representative survey interviewed 1,106 employed US adults between July 10 and 19, 2026. Respondents were asked about ten common work tasks drawn from US Department of Labor data, chosen to reflect typical knowledge work. Workers reported using AI across all ten tasks, but adoption varied sharply by task.
AI Adoption Varies Widely by Task
Software development leads the pack. Among workers who do that task, 57% use AI. Data analysis follows at 46%, and reading work documents sits at 39%. Record-keeping trails at 25%, the lowest of the ten tasks examined.
The pattern suggests AI has found its footing in some areas faster than others. Coding and data work lend themselves to automation in ways that administrative record-keeping may not. Still, the survey shows AI is now a regular presence across the full range of knowledge work.
Workers mostly describe AI as partial support rather than a full replacement. Full or near-full task completion by AI hits 10% only in software development. For all other tasks, that figure stays below 7%. In data analysis, 7.1% of respondents say AI has taken over tasks that people used to handle. Reading work documents comes in at 5.7%, and record-keeping at 5.3%.
Time Savings Depend on How Much AI Does
The survey also probed what happens to time when AI steps in. When AI handles only part of a task, respondents report saving time on 37% of those tasks. When AI does most or all of the work, time savings are reported in 53% of cases.
But the picture is not uniformly rosy. About one in six AI-assisted tasks takes longer than before. Researchers suggest longer task times may be because interacting with AI eats up time or because workers use freed-up capacity to do tasks more thoroughly or at higher quality.
Whether heavier AI use actually makes people faster or whether workers simply lean on AI more when they want to save time can't be determined from the data alone. The results are based on self-reported data, and actual time savings and quality of AI output were not measured objectively.
Editing Effort Tells a Complicated Story
The survey also looked at how much workers edit AI output. A striking 66% of AI output is used unchanged or with only minor tweaks. Just 6% of AI output is used without any changes at all. At the other end, 5% of AI output is heavily reworked or mostly rewritten.
Stay ahead of the AI curve
The most important updates, news, and content — delivered weekly.
No spam. Unsubscribe anytime.
Researchers note that low editing effort isn't a direct measure of AI output quality. They also found no consistent link between reported time savings and the amount of editing. That suggests workers may accept AI output as-is for reasons other than quality, such as convenience or time pressure.
Epoch AI calls AI a "versatile but usually not self-sufficient workplace tool." The data points to redistribution of tasks between humans and AI, not full automation of entire jobs. Epoch AI stresses that task substitution doesn't automatically mean workers are being displaced entirely.
Earlier Surveys Show a Broader Trend
The new findings build on earlier work. A Gallup survey from August 2025 found that 45% of US workers use AI on the job, but only 10% use it daily. That gap between occasional and daily use suggests AI is still a supplement for most people.
Anthropic, the maker of Claude, conducted its own survey of roughly 9,700 Claude users. About half of those respondents believed AI could handle 50% or more of their work. That sample consisted of users of a specific AI product, so it was not representative of the general population.
The contrast is telling. Power users of a single AI tool express high confidence in its capabilities. The broader workforce, by comparison, is more cautious. The Epoch AI and Ipsos survey, with its nationally representative sample, offers a more grounded view of everyday reality.
Limits of the Data
The survey has clear boundaries. Tasks were selected based on national employment data, not on the likelihood of AI impact. That means the ten tasks may not capture the most AI-exposed corners of the economy.
Self-reporting also carries risks. Workers may overstate or understate their time savings, and the quality of AI output is not objectively measured. The researchers acknowledge these limits in their analysis.
Still, the survey offers a useful snapshot. AI is now embedded in the daily routines of a meaningful slice of the US workforce. It is not yet running the show, but it is increasingly sharing the load.
The findings suggest a workplace in transition. Some tasks are being handed off to AI, while others remain firmly human. The line between the two is still being drawn, task by task, job by job.