Does teaching people to use AI well actually shrink the time they spend fixing what it produces?

This explores whether training people to use AI shrinks the cleanup work that comes after the AI produces something. The surprising answer from the corpus is that training doesn't seem to shrink cleanup. Trained workers often do more of it. In a Zapier survey of enterprise AI users, the average worker spent about 4.5 hours a week revising AI output, and the trained, heavy users reported both the biggest productivity gains and the most cleanup time How much time do workers really spend fixing AI mistakes?. So training doesn't remove the work. It seems to make people use AI more, and more confidently, and that means more output to check.

One way to make sense of this is to stop treating cleanup as waste. Research on how AI changes time use finds that it rarely cuts total task time. Instead it moves time away from doing the work yourself and toward writing prompts and reading, judging and fixing what comes back Does AI really save time, or just change how we spend it?. Narayanan and Kapoor describe a similar pattern across whole jobs: AI speeds up the middle 'execute' step, while deciding what to do and making sure the result is fit to deliver stay the same or grow Does AI really compress all layers of knowledge work equally?. Seen this way, cleanup is the 'deliver' step growing to fill the space AI freed up. A skilled user may spend more time there because they know what good output looks like.

The Workday data adds a twist. About 40% of reported time savings disappeared into fixing errors and checking outputs, and only 14% of employees consistently came out ahead. The ones who did tended to work at organizations that retrained staff and also redesigned roles, rather than just handing out tools Where does AI's time savings actually go in practice?. That points to a distinction. Training people to use the tool doesn't seem to be enough. Changing the job so that checking AI output is a planned part of the work, not hidden overtime, is what appears to make the gains real. That hidden part matters: most workers in Anthropic's interview study played down their AI use Why do workers hide productivity gains from AI use?, so cleanup time probably goes underreported in many workplaces.

There's also a skills angle you might not expect. Productivity gains from AI show up when people apply skills they already have, not when they're learning something new When does AI actually boost worker productivity?. That suggests good cleanup depends on knowing the subject, not just knowing the tool. If AI takes over the error-fixing that used to build that knowledge, the ability to catch AI's mistakes could weaken over time. Learners who handed debugging to AI kept fewer skills than those who ran into errors and fixed them themselves Does AI assistance remove a core learning channel through error work?. Nielsen pushes back on how such studies are designed, though. He argues that the useful question is how skills develop when AI is always available, not what happens when it's taken away Does removing AI tools actually measure real skill loss?.

A caveat: none of these sources directly tests training against cleanup time over a long period. The evidence is surveys and related studies. The pattern they share is still clear: cleanup doesn't shrink with training, and it may be where skilled AI users now spend much of their time.