How much time do workers really spend fixing AI mistakes?

Zapier, which bills itself as "the most connected AI orchestration platform," commissioned Centiment to survey more than 1,100 U.S. enterprise AI users (screened as AI users at companies with 250+ employees, fielded November 13–14, 2025, unweighted, margin of error about ±4% at 95% confidence). The survey finds that while 92% of workers say AI boosts their productivity, the average employee spends 4.5 hours a week — "more than half a workday" — revising, correcting, and sometimes completely redoing AI-generated output; the release calls this gap "AI workslop." Only 2% say they generally don't need to revise what AI produces, and 74% report at least one negative consequence from low-quality AI output, including work rejected by stakeholders (28%), security incidents (27%), and customer complaints (25%). Data analysis and visualization top the cleanup list at 55%, ahead of writing tasks at 46%.

The survey's own explanation is a training and context divide, not a model-capability one. Workers without AI training are "6x more likely to say AI makes them less productive" (6% vs. 1%) and report benefit far less often (69% vs. 94% of trained workers) — yet trained workers also spend more time on cleanup, because, as the source puts it, they "use AI more aggressively, more frequently, and in higher-stakes contexts where both the benefits and the cleanup requirements are greater." Cleanup time is framed as rising with usage intensity rather than falling with skill. Zapier's stated remedy is explicitly its own product category: "the solution isn't fewer tools, it's better infrastructure" — orchestration platforms, company context fed into workflows, prompt libraries, and mandatory training, each tied in the data to a specific productivity figure (97%, 96%, 95%).

How much work that employees receive is actually unhelpful AI content? puts a different number on the same term, estimating a share of received work (15.4%) and a per-instance time cost (1h51m); this survey instead asks the person doing the cleanup for a weekly total (4.5 hours) rather than a share of incoming work, and ties the burden to training and usage intensity rather than to who sent the work. It also complicates Does AI assistance erode the skills needed to oversee it?: both surveys find self-reported productivity gains sitting alongside a hidden correction cost, but where Anthropic's engineers describe an oversight gap that widens with delegation, Zapier's trained users report the opposite direction — more use produces more perceived benefit even as cleanup hours rise. That pattern cuts against Does AI assistance help less experienced workers most?, a measured field result where gains concentrated among the least experienced; Zapier's self-reported pattern runs the other way, with trained, heavy users reporting the largest gains — though the two are not measuring the same thing, behavior against perception.

This is a vendor survey: Zapier sells the orchestration tools its own data recommends, and the release's "solutions" section reads as product marketing as much as finding. The 97%/96%/95% productivity figures for orchestration, context, and prompt libraries are self-reported perceptions from an unweighted convenience sample, not a controlled comparison, so they cannot show that adopting those tools causes the gain rather than correlating with the kind of worker who already adopts them. The 4.5-hour cleanup figure is likewise self-reported, not validated against logged time or output audits. What the source supports is narrower than its headline: in this one sample, workers who say AI helps them also say they spend substantial time fixing what it produces, and that time rises with how much and how aggressively they use it — not that better orchestration or training actually reduces the cleanup burden, which the survey recommends but does not test.