A new survey gives one of the clearest pictures yet of how AI is actually being used inside American workplaces, and the number that stands out is one in five. That is the share of employed Americans who say AI now handles a task that used to go to a coworker or an outside contractor. The survey comes from the research group Epoch AI and pollster Ipsos, who talked to 1,106 employed adults in July 2026 about ten common office tasks pulled from government labor data.
The pattern is uneven. Software development leads by a wide margin, with 57 percent of workers who do that kind of work using AI for it. Data analysis follows at 46 percent. Reading work documents sits at 39 percent, and plain record keeping trails at just 25 percent. The order makes sense: the more a task involves clean digital text or numbers, the easier it is for AI to help with it.
Full replacement of a task is still rare. Even in software development, only about 10 percent of workers say AI now does nearly all of the task by itself. In every other category, that number stays under 7 percent. Most AI use today is a helper sitting next to a person, not a replacement standing in for them.
Time savings are not automatic either. When AI does only part of a task, workers report saving time about 37 percent of the time. When AI takes over most or all of a task, that rises to 53 percent. But roughly one in six AI-assisted tasks actually takes longer than it used to, possibly because checking the AI's work, or using the spare time to polish the output, eats up the gains.
There is a quieter finding buried in the numbers that matters for anyone managing a team: 66 percent of AI-generated work gets used with little to no editing, and only 5 percent gets heavily reworked. Low editing is not the same as high quality. If nobody is set up to catch errors, they can travel straight into a report, a contract, or a client email.
Here is the part the survey does not answer on its own, but that a separate tracker fills in. The Budget Lab at Yale has been watching national employment and wage data for jobs exposed to AI, and month after month it finds no measurable change. Put the two findings together and a clearer story appears: AI is moving pieces of work out of one person's hands and into a chatbot, without necessarily removing anyone's job title. A worker can stop sending data analysis to a contractor and hand it to AI instead while staying in the exact same role.
That is the real shift business owners should watch. It is not mass layoffs showing up in the jobs report. It is a slow redrawing of who does what inside a job that still exists on paper. For a manager, the useful move is to ask, task by task, which pieces of work in your business are already quietly going to AI, and whether anyone is still checking that the output is right, fast, and actually cheaper than the old way.