Workers say AI makes them faster — but does that feeling hold up once you count the cleanup time and whether the company actually benefits?

This asks how far we can trust workers' and executives' own estimates of how much AI speeds them up. The corpus doesn't say these estimates are useless. Its answer is that they measure something real but narrow: the moment a task feels faster. They tend to leave out the cleanup afterward, the cost to coworkers, and whether the organization actually gains anything. If you follow these surveys from the individual worker up to company results, the gap between reported and measured gains gets wider at almost every step.

Start with the most careful survey work. Bick, Blandin and Deming turn self-reported time savings into an estimate of how much AI has raised productivity across the whole economy: about 1.1% What is generative AI's actual aggregate productivity gain today?, and up to 1.3% in pooled data Did generative AI actually boost productivity growth?. That isn't far from expert-based estimates, and it roughly matches the speed-up in official US productivity figures since ChatGPT launched. So self-reports aren't wildly wrong when you aggregate them carefully. Read the industry surveys next to them, though, and a pattern shows up. Zapier found that 92% of enterprise AI users report productivity boosts, yet the average worker spends 4.5 hours a week fixing AI output. The heaviest users report the biggest gains and also do the most cleanup How much time do workers really spend fixing AI mistakes?. A Workday survey estimates that nearly 40% of reported time savings go into rework and checking AI outputs, and only 14% of employees consistently come out ahead Where does AI's time savings actually go in practice?. People appear to remember the fast first draft and forget the slow corrections.

A less obvious blind spot is that your own time savings can turn into someone else's workload. In a Stanford and BetterUp survey, workers estimated that 15.4% of the work they receive is unhelpful AI-generated content. Each instance takes them almost two hours to deal with, longer than if the sender had done the work properly How much work that employees receive is actually unhelpful AI content?. A survey that only asks you about your own productivity will never catch this. BCG found a related effect: self-reported productivity rises with up to three AI tools and then falls with four or more, mainly because of the effort of supervising them Does using more AI tools always boost worker productivity?. Self-reports can pick up that kind of overload, but only when the survey asks the right question.

The gap is widest between individuals and the organization. McKinsey found that 80% of respondents report personal productivity gains, but only 37% see company profits improve. The few firms that do see results have redesigned their workflows rather than just handing out tools Why do individual AI gains not translate to business results?. In a survey of 750 executives, Baslandze and coauthors found that perceived gains exceed measured ones, partly because revenue lags behind operational improvements Do AI productivity gains feel larger than they actually measure?. That lag cuts both ways. Brynjolfsson, Rock and Syverson's 'productivity J-curve' argues that official statistics can undercount early AI gains. Companies spend heavily on training, process changes and software, and none of that shows up as output at first, though it can lead to overcounting later Why does AI productivity appear weak when investment is high?. So when perception and measurement disagree, the survey isn't necessarily the one that's wrong. Sometimes the hard data is late.

The idea worth taking away comes from aviation research on automation. Simkute and colleagues argue that AI moves people's work from doing a task to checking it, interrupts their focus, and makes easy tasks easier while hard ones stay hard Does generative AI productivity actually improve or decline in practice?. Narayanan and Kapoor make a similar point: AI shrinks the 'execute' part of knowledge work, while deciding what to do and delivering the result stay the same size or grow Does AI really compress all layers of knowledge work equally?. That explains why self-reports mislead. People are mostly reporting on the part of their work that AI speeds up, and the slower parts around it are easy to overlook. The corpus has many surveys that compare self-reports with each other, but few studies that check them against directly measured output for the same workers. That gap is still open.