Workers may be spending almost 47 working days a year using AI, but around 20 of those days are spent troubleshooting errors and refining prompts rather than completing productive work, according to new BambooHR research.
The survey of 1,608 full-time, US desk workers found employees spend an average of 87 minutes a day using AI. Respondents said 42 percent of that time goes on troubleshooting and prompt iteration, compared with 35 percent spent on work that actively progresses their workload.
The findings come as 63 percent of organisations have increased their AI tool budgets in response to employee usage, according to BambooHR.
“Workers today are turning to AI all the time to learn things, and it's incredible to be able to ask a question and get an answer immediately. However, there's a difference between knowing something and developing the judgment to know what to do with it,” said Nicole Csizar, Senior Director of HR Services at BambooHR.
“Organisations need to be intentional about making sure efficiency doesn't come at the expense of mentorship and development, because those human conversations are where growth happens.”
AI Usage Does Not Automatically Equal Value
BambooHR found that 65 percent of workers feel confident and enthusiastic about using AI at work. Time savings were the main motivation for 58 percent of respondents.
However, the research suggests that enthusiasm and usage are not necessarily translating into measurable productivity.
A growing number of active users, prompts and AI licences can demonstrate adoption. They do not necessarily show whether an employee completed a task faster, whether the output was accurate, or how much rework was required before it could be used.
That distinction is increasingly important as AI is embedded across meetings, messaging, calling, documents, enterprise search and other everyday collaboration workflows.
An AI assistant that requires repeated prompt revisions, produces unreliable summaries or lacks the necessary context can introduce friction at scale. Employees may still use the tool, but the time saved on an initial task can be lost in verification, correction and adaptation.
A separate UK study published by Glean in July reached a similar conclusion. Its Work AI Index found that workers could lose 6.3 hours a week to “bot sitting”, including supplying context, supervising outputs and correcting errors.
“LLMs are trained on the vast corpus of the internet – they’re often not trained deeply on our own work, our team’s work, our organisation’s work. That is the number one driver of bot sitting,” Dr Rebecca Hinds, Head of the Work AI Institute at Glean, told UC Today.
Measuring the Work Around AI
BambooHR's findings reinforce a challenge already facing enterprise AI programmes: access to technology is not the same as effective implementation.
Organisations can provide licences and run pilots, but employees still need clear, task-specific guidance, trustworthy data sources and workflows designed around the technology’s limitations as well as its strengths.
For buyers, that means evaluating AI deployments against outcome-based measures, including the time to complete a defined workflow, output acceptance and accuracy rates, the volume of work requiring correction or rework, the level of human review needed before an output can be used, and whether the tool reduces escalations or creates new ones.
Generic prompt training may help workers use AI more confidently, but it is unlikely to solve every deployment issue. The more practical opportunity is to identify repeatable, high-volume workflows where AI can reliably support work, then equip employees with approved tools and clear expectations.
The report's findings also point to security and governance risks surrounding workplace AI use – revealing that 59 percent of workers use personal AI accounts for work tasks. Among those respondents, 71 percent said they had entered client data, proprietary strategy or other sensitive company information into tools their employer cannot see.
More than half of organisations, 54 percent, do not have a clear and consistently communicated AI usage policy.
A Better AI Performance Review
The study suggests organisations need to look beyond licence counts and prompt volumes when assessing AI investments.
The central measure should be whether AI reduces the time, effort and rework required to complete a specific task to an acceptable standard.
Leaders should also investigate where employees are encountering friction, what types of work trigger repeated prompt iteration, and whether workers are turning to unapproved tools because sanctioned AI options do not meet their needs.
AI can still create meaningful gains across collaboration and productivity workflows. But those gains will be difficult to demonstrate if organisations measure usage while overlooking the time employees spend getting the technology to work.
BambooHR conducted the research from June 26 to July 15, 2026. The study included 1,608 US full-time salaried desk workers, including a subgroup of 520 HR professionals holding a manager title or above.
