The Hidden Cost of AI: Workers Spend as Much Time 'Botsitting' as Producing

What happened?

The Work AI Institute, sponsored by Glean and with contributors from Stanford and UC Berkeley, surveyed 6,000 digital workers in the United States, United Kingdom, and Australia between December and January. The report, led by Paul Leonardi (UC Santa Barbara), reveals that workers spend an average of more than six hours per week 'botsitting'—an activity that includes supervising, correcting, and providing the tacit knowledge needed for AI to function properly. Of that time, 37% goes to supervision and correction tasks, compared to 36% for actual production. Additionally, more than a third of AI sessions fail, requiring restarts or rework. The study, covering 6,000 digital workers across three countries, shows that for every hour of useful AI production, approximately another hour is needed to make it usable. This 'invisible layer of human work' is the main reason why, although 75% of individuals report productivity increases, only 13% of organizations see significant business gains.

Why is it important?

The study reveals a key paradox in AI adoption: individual productivity does not automatically translate into organizational results. Leonardi compares it to having to act as a 'manager' of bots, an invisible burden not accounted for in traditional metrics. This phenomenon is not new: during the Internet revolution, companies invested in technology without redesigning processes, leading to Solow's productivity paradox. Now, with AI, the situation repeats: time saved on routine tasks is consumed by managing the AI itself. Moreover, 41% of workers admit to delivering AI-generated content they could not explain if asked, posing risks to quality, legal liability, and ethics. This data is especially concerning in regulated sectors like finance or healthcare, where explainability is mandatory. The study also notes that workers are transferring their personal judgment to bots, which can erode critical skills and increase technological dependence.

Consequences and outlook

If left unaddressed, 'botsitting' can erode trust in AI and slow its adoption. Companies need to redesign workflows to better integrate AI supervision, or invest in more reliable tools. The study suggests that individual productivity does not automatically translate into organizational results, forcing a rethink of performance metrics. In the long term, a new role—'AI manager'—could emerge, similar to human team managers today. However, this also implies additional training and hiring costs. Compared to robotic process automation (RPA) a decade ago, where 'bots' required constant maintenance, generative AI adds a layer of complexity by needing contextual tacit knowledge. Companies like Glean, which sponsor the study, could benefit from tools that reduce 'botsitting,' but the report warns that the solution is not just technological: cultural and organizational change is needed. The AI management software market, valued in the billions, could grow if companies prioritize seamless integration. For workers, the recommendation is to document their tacit knowledge and participate in training systems, while for startups, the opportunity lies in creating tools that automate AI supervision.

“What's happening with these generative AI tools is that we are essentially expecting individual contributors to act as managers. They are managing AI tools, AI agents, and we expect them to produce much more, but we are not accounting for all the work involved in managing them.” — Paul Leonardi, co-author of the study.

What readers should know

AI adoption is not a 'plug and forget' affair. It requires investment in training, processes, and tools that minimize failures. Workers must document their tacit knowledge for AI to be effective, and companies should measure not just output but system maintenance time. The promise of total efficiency clashes with the reality of an invisible layer of human work. The Work AI Institute study is a call to action for managers and employees to recognize 'botsitting' as a real cost and design strategies to reduce it. Otherwise, AI could become a burden rather than a boost. History shows that transformative technologies require process reengineering; ignoring it led to the failure of many ERP initiatives in the 1990s. Now, with AI, the risk is similar: without proper management, 'botsitting' could consume more time than it saves, hampering long-term productivity.