Workers Report Limited Gains from Prompt-and-Pray AI

Glean's Work AI Institute, working with AI researchers from Stanford University, UC Berkeley, and Harvard University, surveyed 6,000 knowledge workers for its inaugural Work AI Index 2026 report. The survey found that AI automation saves workers an average of 11 hours a week, and that most workers say AI makes them individually more productive, but only 13% say their organization's overall performance has significantly improved. The report's authors write that "high AI achievers don't just prompt and pray": a top-performing 13% of the sample reports both productivity and quality gains, while the rest see time savings absorbed elsewhere.

The gap comes down to unpaid "maintenance" work. Glean found workers spend about 6.4 hours a week "botsitting" - checking, correcting, and cleaning up AI output - and that more than a third of AI sessions fail outright and require a redo, according to Rebecca Hinds, head of the Work AI Institute, in comments reported by CIO Dive. Only 27% of the time workers spend interacting with AI tools goes toward learning them or building agents; the rest goes to management overhead. Separately, 53% of workers say information critical to their jobs is not accessible through their AI systems, and in what the report calls "context-rich" organizations, workers are 64% less likely to feel worn out by AI, 52% less likely to ship work they can't explain, and spend meaningfully less time on both botsitting and "botshitting" (producing confident-sounding but ungrounded output).

The report's clearest signal for teams standing up enterprise AI: treat it as a work-design and context problem before a procurement problem. Top-performing organizations in the survey were far more likely to redesign workflows around AI (90% versus 54% of laggards), provide adequate training (90% versus 52%), and formally reward AI skills (84% versus 48%). Notably, high performers were also 18% more likely to deliberately withhold AI from certain tasks, a reminder that knowing where not to use AI is itself a maturity signal, not a failure to adopt.

Watch whether more companies start measuring AI success by quality and explainability rather than seat counts or token usage, since Hinds explicitly warns against treating adoption as "a vanity metric." Also watch for follow-on data on "shadow AI" use: the study found 54% of workers use unapproved tools or approved tools in noncompliant ways, and 36% hide how much AI is helping them, often because official tools are too slow or disconnected from real workflows.