AI isn’t just helping US workers. For one in five of them, it’s now doing jobs that used to land on a coworker’s desk. That’s the headline from a new representative survey by Epoch AI and polling firm Ipsos, as detailed in The Decoder. The researchers interviewed 1,106 employed US adults between July 10 and 19, 2026, and found that 20 percent of them now delegate at least one task to AI that a human previously handled.
What stands out here is the framing. This isn’t about AI making people slightly faster at email. It’s about tasks quietly moving off human to-do lists and onto a machine.
What the researchers did
Epoch AI built the survey around ten common work tasks pulled from US Department of Labor data, chosen to reflect typical knowledge work. Respondents were asked whether they use AI for each task, how much of the work AI handles, whether it saves time, and how heavily they edit the output. Workers reported using AI across all ten tasks, though the adoption gap between them is wide.
Where AI shows up most
Adoption tracks closely with the type of work. Technical tasks lead by a big margin, according to The Decoder’s coverage of the study.
- Software and systems design: 57% of workers who do this task use AI
- Data analysis: 46%
- Reading work documents: 39%
- Record-keeping: 25% (the lowest of the ten)
But using AI doesn’t mean the machine runs the whole show. Most workers describe it as partial support. Full or near-full task completion by AI only reaches 10 percent in software development and stays below 7 percent everywhere else.
Tasks are shifting, not disappearing
The clearest sign of substitution is in data analysis, where 7.1 percent of respondents say AI took over work people used to do. Reading documents follows at 5.7 percent, then record-keeping at 5.3 percent. Epoch AI is careful to note this doesn’t mean those workers are being displaced. It calls AI a “versatile but usually not self-sufficient workplace tool,” and the data points to a redistribution of tasks between humans and machines rather than whole jobs being automated away.
The time-savings picture is messier than you’d think
More AI involvement correlates with more reported time saved. When AI handles only part of a task, workers report saving time on 37 percent of them. When AI does most or all of the work, that jumps to 53 percent.
Here’s the catch. About one in six AI-assisted tasks now takes longer than before. The researchers suggest that’s either because working with AI eats up time, or because people use the freed-up capacity to do the job more thoroughly.
Two-thirds of AI output ships with barely any edits
This is the finding practitioners should sit with: 66 percent of AI output gets used unchanged or with only minor tweaks. Just 6 percent goes out completely untouched, and only 5 percent gets heavily rewritten. Epoch AI warns that low editing effort isn’t proof of high quality. It found no consistent link between how much workers edited and how much time they saved.
What you can do with this
- Audit your own delegation. If AI is quietly absorbing tasks that used to go to a teammate or contractor, name it. That’s a budgeting and headcount signal, not just a productivity tweak.
- Watch the light-editing trap. Two-thirds of output shipping with minimal review is a quality risk. Build a review step for anything client-facing or high-stakes.
- Measure real time, not vibes. One in six tasks getting slower means “AI saves time” isn’t automatic. Track it before you assume it.
The fine print
The results are self-reported. Neither actual time savings nor output quality were measured objectively, and the tasks were picked from employment data, not by how exposed they are to AI. Still, the direction matches other work. A Gallup survey from August 2025 found 45 percent of US workers use AI on the job, and a recent Anthropic survey of Claude users saw roughly half say AI could already handle 50 percent or more of their work.
The trend line is consistent across all three. AI is moving from assistant to task-owner, one line item at a time. Full details are available in the original report at The Decoder.