How much of your AI time is wasted fixing its output?

42% of Your AI Time Is Troubleshooting: What to Do About It

Roughly 42% of it, according to BambooHR's September 1, 2026 survey of 1,608 US desk workers: 87 minutes of daily AI use, with 42% spent troubleshooting errors and iterating prompts and only 35% on productive work, which is about 20 workdays a year. Ed Miller's The Logic of Sports Betting calls that spread the hold, and the fix is the same: stop delegating tasks where AI's hold exceeds the time it saves.

The number that stopped me this week was not the headline. It was buried further down: vice presidents and C-suite executives now spend 101 minutes a day with AI, nearly twice the 54 minutes of individual contributors. The people with the most expensive hours are spending the most of them talking to a model, and by the survey's own arithmetic almost half of that conversation is repair work.

The study is BambooHR's Redesigning Work: AI's Performance Review, released September 1, 2026, based on 1,608 full-time salaried US desk workers including 520 HR professionals. The averages: 87 minutes of AI use per day, which annualizes to 22,526 minutes or about 47 eight-hour workdays. Respondents said 42% of that time went to troubleshooting errors and iterating on prompts, 35% went to work that actually advanced their workload, and the remaining 23% was other activity. HR Dive's September 2 write-up ran the same figures to roughly 20 days a year spent getting AI to work correctly. And the mood is untroubled: 65% feel confident and enthusiastic about AI, 58% cite time savings as their main motivation, and 63% of organizations have already raised AI tool budgets without clear evidence the spend pays back.

I read this through an unusual book for a coaching practice: Ed Miller and Matthew Davidow's The Logic of Sports Betting. Their core teaching is that amateurs think about whether a bet is a good opinion, while professionals think about the hold, the built-in spread between what the house charges and what a fair price would be. A typical NFL market has a 4.5% hold, and Miller's warning is that it destroys your margin for error: make four good bets and one mistake, and the hold quietly eats all four wins. Every AI delegation is a bet with a hold. The hold is the correction time, and BambooHR just measured the market-wide average at 42 cents on the dollar. At that price, an executive whose AI-assisted work is only somewhat better than their unassisted work is losing hours while feeling productive, which is exactly the trap Miller describes: most bettors lose, and most of them are enjoying it.

Miller's other lesson is that you do not shop for a bet and then look at the price; you find the zero-hold markets first and only then decide what you like. Translated into how I run my own AI use, that became a protocol I now apply before delegating any recurring task. I call it the Hold Test:

Price the manual version. Time yourself doing the task without AI once. That is your fair line. Done when you have a real number in minutes, not a guess.
Log the full AI cycle, not just the prompt. Start the clock at the first prompt and stop it when the output is actually shipped, including every re-prompt and every edit. Done when the number includes the fixing.
Compute the hold. Correction minutes divided by total AI minutes. Above 40% means you are betting into BambooHR's average market; above 60% means you are the sucker at the table.
Keep only positive-expectation tasks. Delegate where AI time is under the manual line with a hold below 25%. Everything else goes back to doing it yourself or to a narrower, better-specified version of the task.

When I logged my own sessions for a week in August the results embarrassed me. Drafting client-facing summaries had a hold around 15%: the first output was usually usable and I saved real time. Anything requiring current numbers, from a competitor price check to a market sizing, ran a hold above 55% because I verified every figure and rewrote most of them. I did not stop using AI for research; I stopped delegating the synthesis and kept it for the search, which is Miller's move of splitting a bad market into a related one that is priced well. The executives I coach who report the least frustration have made the same split without naming it: they delegate transformation of material they already understand and refuse to delegate judgment about material they do not.

The limits deserve the same honesty. This is a self-reported perception survey with no activity logs and no non-AI baseline, so it cannot show that AI made anyone slower; a worker who spends 42% of AI time iterating may still finish faster than without it. The line between troubleshooting and legitimate refinement was drawn by respondents themselves. The sample is US salaried desk workers, not founders or frontline staff. And the 101-minute executive figure may partly reflect that senior people are given the tools first. None of that changes the practical point: nobody in this survey knew their own hold, and once you measure yours the decision about what to delegate stops being a matter of enthusiasm.

Sources: BambooHR press release, "Workers Lose 20 Days of Productivity to Troubleshooting AI Each Year," September 1, 2026 (bamboohr.com); BambooHR, "Redesigning Work: AI's Performance Review" full findings (bamboohr.com/resources/data-at-work); HR Dive, "Almost half the time spent on AI is on fixing its output, BambooHR says," September 2, 2026 (hrdive.com); Quasa's methodological breakdown of the annualized figures, September 5, 2026 (quasa.io); Ed Miller and Matthew Davidow, The Logic of Sports Betting (2019).