The Hidden Cost of Enterprise AI: 6.4 Hours a Week Babysitting Bots

At a glance:

• Enterprise AI adoption costs employees 6.4 hours weekly on bot-sitting tasks
• 87% of digital workers use AI, but only 13% report significant performance gains
• 69% of users admit to shipping unverified AI-generated work (botshitting)

The Productivity Paradox of AI in the Workplace

The integration of AI into enterprise workflows has created a counterintuitive productivity paradox. While AI tools promise to streamline tasks and save time, they instead demand significant human oversight. A survey by Glean’s Work AI Institute of 6,000 full-time digital workers revealed that employees spend 6.4 hours per week on bot-sitting—the unrecognized labor required to feed AI context, debug outputs, and manage multiple tools. This time investment, though framed as efficiency, actually reduces net productivity. The paradox lies in AI’s ability to automate tasks while simultaneously increasing the cognitive load on workers to ensure its reliability.

The report, co-authored by experts from Emory University, Stanford, and UC Berkeley, highlights that AI’s value is often undermined by its lack of enterprise-specific knowledge. Large language models (LLMs) are trained on public data, not proprietary company information, forcing employees to repeatedly provide context about products, customers, or processes. Rebecca Hinds, head of Glean’s Work AI Institute, notes that this friction creates frustration. "Workers feel exhausted when tools don’t understand day-to-day operations," she said. The result is a cycle where AI requires more human intervention than it delivers, eroding the very efficiency it aims to achieve.

Workers Struggle with Bot-Sitting and Bot-Shitting

The survey identified two key behaviors: bot-sitting and bot-shitting. Bot-sitting refers to the manual work of preparing AI for use, such as refining prompts or cleaning up errors. Bot-shitting, conversely, involves shipping unverified AI-generated work due to time constraints or overwhelm. Sixty-nine percent of respondents admitted to botshitting, with 41% unable to explain their AI outputs if asked. This behavior risks propagating errors, as unverified outputs may lack critical context or contain inaccuracies.

Hinds emphasizes that botshitting is not just a time-saving tactic but a form of offloading critical human judgment. "You’re delegating thinking that must stay with humans," she explained. The issue is exacerbated by the use of multiple AI agents, which can spiral out of control without proper governance. Users overwhelmed by fragmented tools often abandon verification efforts, leading to cascading errors that require extensive cleanup. This dynamic creates a feedback loop where AI’s scalability amplifies human inefficiencies.

The Role of Multiple AI Tools

A significant driver of bot-sitting is the proliferation of unconnected AI tools. The survey found that 54% of high AI achievers use unapproved or noncompliant tools, while 36% hide their AI usage. This fragmentation forces workers to switch between platforms, repeating prompts and context across systems. Hinds notes that "multiple tool usage is a major contributor to exhaustion." Without integration, AI becomes a productivity drain rather than a collaborator. Employees must manage permissions, ensure data consistency, and troubleshoot compatibility issues—tasks that detract from core work.

The problem is compounded by AI’s variability. Outputs may appear polished but lack accuracy or relevance. Debugging requires workers to cross-check results, often without prior involvement in the initial generation. This not only consumes time but also creates accountability gaps. As Hinds points out, "You don’t see the negative impacts until several steps down the line," making it harder to trace errors back to their source.

Successful Organizations Addressing the Challenges

Companies leading in AI adoption are tackling these challenges proactively. Instead of focusing solely on AI tool usage, they invest in the work around it: setting clear context, defining quality standards, and building human judgment into workflows. The report highlights that transformative organizations prioritize training, governance, and rewards for AI skills. For example, they provide structured onboarding to help employees understand AI’s limitations and best practices. Hinds stresses that leadership must model AI use, sharing both successes and failures to normalize its role.

Governance is another critical factor. Successful companies treat AI policies as "living and breathing," regularly revisiting them to adapt to new tools and risks. They also align AI initiatives with existing KPIs, measuring not just efficiency but quality and employee engagement. By embedding AI into broader workflow strategies, these organizations reduce the need for bot-sitting. Hinds notes that "it’s less about surveillance and more about feedback," emphasizing collaboration over top-down control.

The Need for Context and Governance

A key takeaway from the survey is the importance of context in AI effectiveness. LLMs, while powerful, lack knowledge of specific enterprise environments. This gap necessitates human input to tailor AI outputs to organizational needs. Hinds argues that this requires a shift in how companies approach AI. "It’s not just about clicks or tokens," she said, "but real skills and learning." Training programs must focus on teaching workers when and how to use AI, rather than just tool adoption.

Governance frameworks must also evolve. The report calls for clearer definitions of what tasks should never be delegated to AI. For instance, sensitive data handling or high-stakes decisions may require human oversight. By establishing these boundaries, companies can mitigate risks while preserving AI’s value. Hinds also highlights the role of psychological safety: employees are more likely to use AI effectively if they feel supported and not penalized for experimentation.

The Future of AI Adoption

The survey reveals a Goldilocks problem: workers are using AI extensively but facing diminishing returns. Over 50% of high achievers use unapproved tools or conceal their AI reliance, fearing negative perceptions. This suggests a tension between demonstrating AI fluency and maintaining professional credibility. Hinds warns that without balance, AI adoption could lead to burnout or reduced trust in the technology.

Looking ahead, the report suggests that AI’s role will depend on how well organizations integrate it into human workflows. Low-code, no-code tools with low learning curves are gaining traction, as workers prefer AI embedded directly into their tasks. Hinds predicts that future success will hinge on "real skills, real learning," not just technical proficiency. Companies that prioritize this holistic approach may avoid the productivity paradox and unlock AI’s full potential.

What’s Next for Enterprise AI

The findings underscore a critical lesson: AI is not a silver bullet. Its value depends on how humans interact with it. As AI tools become more sophisticated, the need for context, governance, and human oversight will only grow. The report calls for a paradigm shift from treating AI as an autonomous solution to viewing it as a collaborative tool. This requires rethinking workflows, investing in employee training, and fostering a culture where AI enhances—not replaces—human judgment.

The hidden cost of enterprise AI is not just financial but human. As Hinds concludes, "The massive, massive human labor at the core of this needs to be addressed." Without systemic changes, the productivity gains of AI may remain elusive, trapped in a cycle of bot-sitting and bot-shitting.