A survey by Epoch AI and Ipsos, conducted with 1,106 employed adults in the United States between July 10 and 19, 2026, revealed that 20% of American workers delegate at least one professional task to artificial intelligence (AI) that was previously done by colleagues or outside contractors. The study asked participants about ten common activities, selected based on data from the U.S. Department of Labor.
AI adoption varies by activity. Among workers who perform each task, 57% use AI in computing systems and software development, 46% in data analysis, and 39% in reading work documents. The lowest rate is in record keeping, at 25%.
In most cases, AI acts as partial support. Full or near-full task completion by the machine reaches 10% only in software development and is below 7% in the other activities.
Replacement and time saved
The transfer of tasks from humans to machines is most evident in data analysis: 7.1% of respondents say AI has taken over activities previously done by people. Next are document reading (5.7%) and record keeping (5.3%). The researchers stress that this replacement does not necessarily mean a total loss of jobs.
The survey also identified a relationship between the volume of work done by AI and the time saved. When the tool performs only part of the task, 37% of respondents report time savings; when it does most or all of it, the percentage rises to 53%.
On the other hand, about one in six AI-assisted tasks began to require more time than before. The researchers suggest that interacting with the AI itself can take time, or that workers use the freed-up capacity to perform tasks more thoroughly or with higher quality.
The survey also shows that 66% of AI-produced content is used without changes or with minor adjustments; 6% receives no changes and 5% is heavily rewritten. The researchers note that low editing effort does not directly measure the quality of the result, and they found no consistent relationship between time savings and editing.
The researchers classified AI as a “versatile tool, but generally not self-sufficient” in the workplace. The data point more to a redistribution of tasks between humans and machines than to full automation of jobs. The conclusions are based on workers' own reports; neither the time saved nor the quality of the result was objectively measured. The sample was selected based on national employment data, not on the likelihood of tasks being affected by AI.
