By Caroline Falkman Olsson, Yafah Edelman, and Amreeta Das, Epoch AI
Epoch AI's previous survey, fielded March 3-5 of this year, found that AI has become a common workplace tool in the US, with 51% of employed AI users reporting they use it at least as much for work as for personal tasks. To zoom into workplace AI usage, Epoch AI then surveyed 1,106 employed US adults about how they use AI for ten work tasks. The surveyed tasks were selected to be broadly representative of US knowledge work.
This reallocation of work to AI appears across all ten tasks measured. It is most common for “analyzing data” (7.1% of respondents), followed by “reading work documents” (5.7%) and “maintaining records” (5.3%). Although the survey shows task-level substitution, this does not necessarily equal full worker displacement.
This breadth of use does not mean AI performs the whole task. In most cases, workers describe it as assisting with only part of the work. Full or nearly full task reallocation to AI is most common for workers designing computer systems or software, at 10%, but remains below 7% for other tasks.
There are several possible reasons for this association. One possibility is that AI taking on a larger share of a task saves workers time. Another is that workers hand more of a task to AI when they are trying to save time. It may also be the case that these findings simply reflect tasks that AI handles particularly well.
However, AI use is not always accompanied by reported time savings. Roughly one in six AI-assisted tasks now takes more time than before. The share is similar whether AI assists with part of the task or performs most or all of it. Possible reasons for this may be that engaging with AI makes these tasks take longer or that workers are spending more time on a task for other reasons, such as AI freeing them up to do more of it or a more high-quality version of the task.
To learn more about how AI is used in the workplace, visit the Polling on Usage explorer to find the full survey results, methodology, and interactive visualizations!
The survey was fielded July 10–19, 2026, and includes 1,106 employed US adults.
Probability-based sampling reduces self-selection bias and produces estimates that are more representative of the employed US population. All estimates are weighted to be representative of employed US adults. A small set (n=3) of invalid responses was removed from the analysis.
The tasks are common work activities drawn from O*NET, the US Department of Labor’s occupational database, and were selected based on employment share to reflect common knowledge work.
Read more about the methodology here.
All point estimates and counts are weighted unless otherwise specified. Confidence intervals are 90% and are estimated using Taylor series linearization, treating the respondent as the primary sampling unit. For findings measured at the task level, this keeps the multiple tasks a worker reported together, so the intervals account for that clustering. The full survey questionnaire, data files, and graphs are available at the polling hub.
About the authors
Caroline Falkman Olsson is a researcher at Epoch AI, focusing on polling, with a background in economics and statistics. She has previously worked as a predoctoral researcher at LSE’s International Inequalities Institute (III) and as a data analyst at the Institute for International Economic Studies (IIES) at Stockholm University.
Yafah Edelman is the head of data and trends at Epoch AI. She is interested in understanding and measuring the inputs that allow AI to scale.
Amreeta Das is a Ph.D. student in Political Science at the University of California, Merced. She is interested in measuring the societal impacts of AI.
Reviewed by Irfan Ahmad.
Read next:
• AI is making disinformation harder to spot – but we’ve found a new way to catch it
• 66% of Americans Have Missed Important Calls, Lose 27 Hours to Spam, and 88% Suspect Unknown Calls
Epoch AI's previous survey, fielded March 3-5 of this year, found that AI has become a common workplace tool in the US, with 51% of employed AI users reporting they use it at least as much for work as for personal tasks. To zoom into workplace AI usage, Epoch AI then surveyed 1,106 employed US adults about how they use AI for ten work tasks. The surveyed tasks were selected to be broadly representative of US knowledge work.
AI is now delegated work previously given to other humans
Reallocation from humans to AI is one concrete way that AI is entering workflows. Epoch AI asked workers whether AI now handles mostly all, or all, of any task that they or their team previously delegated to a contractor or coworker. One in five said that this has happened for at least one of the ten tasks measured.This reallocation of work to AI appears across all ten tasks measured. It is most common for “analyzing data” (7.1% of respondents), followed by “reading work documents” (5.7%) and “maintaining records” (5.3%). Although the survey shows task-level substitution, this does not necessarily equal full worker displacement.
One in five workers say AI now handles work they once delegated
Among employed US adultsAI is used for tasks that cut across widely held jobs
AI use is reported across all ten tasks, although adoption rates vary considerably. Among workers performing the “designing computer systems or software” task, 57% use AI to some extent for that work. This share is 46% among workers “analyzing data” and 39% for “reading work documents”. Uptake is lowest for “maintaining records”, where 25% of workers who perform it use AI.This breadth of use does not mean AI performs the whole task. In most cases, workers describe it as assisting with only part of the work. Full or nearly full task reallocation to AI is most common for workers designing computer systems or software, at 10%, but remains below 7% for other tasks.
Across every regular job task, a quarter to half of workers now use AI
Among employed US adults whose job includes each taskAI usually doesn’t do the whole task, but saves time most often when it does
More AI involvement is associated with a greater likelihood of reported time savings. Respondents report that when AI assists with part of the work, 37% of tasks take less time. Among tasks where AI does most or all of the work, the share rises to 53%.There are several possible reasons for this association. One possibility is that AI taking on a larger share of a task saves workers time. Another is that workers hand more of a task to AI when they are trying to save time. It may also be the case that these findings simply reflect tasks that AI handles particularly well.
Workers more often report time savings when AI does more of the task
Among job tasks where employed US adults use AIHowever, AI use is not always accompanied by reported time savings. Roughly one in six AI-assisted tasks now takes more time than before. The share is similar whether AI assists with part of the task or performs most or all of it. Possible reasons for this may be that engaging with AI makes these tasks take longer or that workers are spending more time on a task for other reasons, such as AI freeing them up to do more of it or a more high-quality version of the task.
Workers accept most AI outputs with minimal revision
Across tasks both partially and fully performed by AI, 66% of AI outputs are used as produced or with only minor revisions. 6% are used without any changes. More extensive revision is reported for 27% of outputs, and 5% of outputs are majorly reworked or mostly redone.Workers keep a majority of AI output without making substantial changes
Among job tasks where employed US adults use AIConclusion
These survey results offer a snapshot of how tasks are divided between people and AI in current workflows. The four key takeaways portray AI as a versatile but usually not self-sufficient workplace tool. Together, they point to tasks being reallocated between people and AI rather than the wholesale automation of jobs.To learn more about how AI is used in the workplace, visit the Polling on Usage explorer to find the full survey results, methodology, and interactive visualizations!
About this survey
The data in this analysis come from an Epoch AI/Ipsos survey on the Ipsos KnowledgePanel, a probability-based online panel recruited via address-based sampling.The survey was fielded July 10–19, 2026, and includes 1,106 employed US adults.
Probability-based sampling reduces self-selection bias and produces estimates that are more representative of the employed US population. All estimates are weighted to be representative of employed US adults. A small set (n=3) of invalid responses was removed from the analysis.
The tasks are common work activities drawn from O*NET, the US Department of Labor’s occupational database, and were selected based on employment share to reflect common knowledge work.
Read more about the methodology here.
All point estimates and counts are weighted unless otherwise specified. Confidence intervals are 90% and are estimated using Taylor series linearization, treating the respondent as the primary sampling unit. For findings measured at the task level, this keeps the multiple tasks a worker reported together, so the intervals account for that clustering. The full survey questionnaire, data files, and graphs are available at the polling hub.
About the authors
Caroline Falkman Olsson is a researcher at Epoch AI, focusing on polling, with a background in economics and statistics. She has previously worked as a predoctoral researcher at LSE’s International Inequalities Institute (III) and as a data analyst at the Institute for International Economic Studies (IIES) at Stockholm University.
Yafah Edelman is the head of data and trends at Epoch AI. She is interested in understanding and measuring the inputs that allow AI to scale.
Amreeta Das is a Ph.D. student in Political Science at the University of California, Merced. She is interested in measuring the societal impacts of AI.
Reviewed by Irfan Ahmad.
Read next:
• AI is making disinformation harder to spot – but we’ve found a new way to catch it
• 66% of Americans Have Missed Important Calls, Lose 27 Hours to Spam, and 88% Suspect Unknown Calls
