Workers Spend 20 Days Fixing AI—Productivity Is Not the Whole Story

BambooHR’s September 1 research release describes an online survey of 1,608 full-time salaried U.S. desk workers, including 520 HR professionals, and estimates 87 minutes of daily AI use, or 22,526 minutes annually; respondents allocated 42% of that time to troubleshooting and prompt iteration, 35% to productive work and roughly 20 eight-hour days to the first category.

HR Dive’s September 2 coverage independently summarized the central estimates as approximately 47 workdays of annual AI use and 20 days spent fixing errors or iterating on prompts, while also noting that 65% of employees felt confident and enthusiastic about AI and 58% identified time savings as their motivation.

The headline number therefore captures a substantial amount of friction, but it does not establish 20 days of net productivity loss. It annualizes workers’ estimates of how they divided their AI time; it does not compare their output with an equivalent period in which they completed the same tasks without AI.

How the survey’s AI time divides into workdays

The published inputs make the annual allocation reproducible. Multiplying the annual total of 22,526 minutes by each share and dividing the results by a 480-minute workday produces this breakdown:

• All AI use: 22,526 minutes, or 46.9 eight-hour days, rounded to 47 days.
• Troubleshooting and prompt iteration: approximately 9,461 minutes, or 19.7 days, rounded to 20 days.
• Productive work: approximately 7,884 minutes, or 16.4 days.
• Other AI activity: the remaining 23%, or approximately 5,181 minutes and 10.8 days.

This arithmetic supports the annualized 20-day estimate. It does not show that employees produced nothing during those hours: revising a prompt may be avoidable rework, but it may also be part of producing an acceptable result. The categories reflect respondents’ own classification of their time rather than an independent assessment of whether each minute added value.

The study measures perception, not productivity

The sample is specific: adults in the United States who held full-time, salaried desk jobs, with a separate subgroup of HR managers and more senior HR professionals. It does not represent hourly employees, frontline occupations, independent contractors or workers outside the country, so its estimates should not be generalized to the entire labor force.

Self-reported time also introduces uncertainties that an activity log would avoid. Participants had to remember how long they used AI, decide when refinement became troubleshooting and distinguish directly productive work from other potentially useful activity. The methodology does not include application telemetry, employer time records or an examination of completed work products to validate those classifications.

The defensible finding is consequently narrower than a verdict that AI reduces productivity. These respondents estimated that correction and iteration occupied a larger share of their AI time than activity that directly advanced their workload. The survey cannot determine whether AI still helped them finish faster, improve quality or produce more than they would have without it because it provides no non-AI baseline or task-level performance comparison.

Policies, sensitive data and hiring add operational risk

BambooHR’s full survey findings put the policy gap at 54% of organizations, personal-account use for work at 59%, and sensitive-data entry at 71% within that personal-account group; they also show that 67% of HR professionals would trust AI to conduct a candidate interview alone, versus 34% of employees, while 80% of HR respondents had encountered bias or other problems in supporting language models and 33% perceived a pattern of demographic bias.

These are respondent claims, not evidence of a particular data breach or proof that a named model discriminated. The sensitive-data percentage applies only to people who used personal AI accounts for work, while the bias figures reflect HR professionals’ perceptions rather than the results of an independent audit.

Even within those limits, the findings connect time lost to correction with weak governance. An organization can increase AI use while leaving employees without clear rules about approved accounts, permissible data, human review or which tasks have been validated for automation. Those controls address the behaviors identified by the survey more directly than a blanket instruction to use—or avoid—AI.

The hiring results create a related accountability problem. Trusting a system to interview candidates independently while users are also perceiving model problems leaves unresolved who reviews its questions, detects disparate treatment and can override its output. The study does not show that automated interviews caused discriminatory decisions, but it identifies a gap between HR professionals’ willingness to delegate and their confidence in the underlying tools.

The productivity verdict remains unresolved

The survey establishes that a defined group of U.S. desk workers believes troubleshooting consumes a substantial portion of its AI-use time. It also identifies uneven policy coverage, workplace data entering personal accounts and strong HR interest in automated hiring despite perceived model shortcomings.

What remains unknown is the net effect on completed work. Resolving that question would require observed comparisons of equivalent tasks performed with and without AI, including completion time, output quality and later rework. Until those measurements exist, the 20-day figure is best read as an estimate of friction and a warning about governance—not a complete productivity balance sheet.

Also read:

• Delta’s Premium Cabins Just Out-Earned Economy for the First Time — And the Real Story Is More Interesting Than “The Rich Are Getting Richer”
• Consumers Are Ready to Spend 29% More Time with AI Content by 2029
• Autism Simulator Is Still Online—but Its Workplace Story Needs Context