AI adds labor, not magic

The new story about AI at work is not that it replaces people. It is that it recruits them into new kinds of labor. Companies keep selling generative AI as a shortcut. Then workers discover that the shortcut comes with setup, supervision, cleanup, and a steady stream of second-guessing. CNBC’s reporting this week captured that gap in plain terms: employees are being told to use AI everywhere, even as the tools create more checking, more rewriting, and more mental strain than the sales pitch admits (cnbc.com). That pressure is coming from the top. In April 2025, Shopify CEO Tobi Lütke told employees that using AI was now a “fundamental expectation,” and that teams should show why AI could not do a job before asking for more headcount (cnbc.com). By September 2025, a survey of business leaders cited by CNBC found that 24% of companies said they required AI use across all roles (cnbc.com). The mandate arrived faster than the operating model. That is why the real work now often happens after the model answers. Workday put numbers on that hidden cleanup in January 2026. Its survey found that employees often did save time with AI, but nearly 40% of those gains were lost to rework: correcting mistakes, rewriting generic text, and verifying output that looked polished but could not be trusted (newsroom.workday.com). CFO.com, reporting on the same study, translated the result into a harsher ratio: for every 10 hours gained, nearly four were spent fixing the machine’s work (cfo.com). That is not automation in the old sense. It is delegated draft work with a human editor chained to it. The strange part is that this does not contradict the research showing real productivity gains. One of the best-known field studies, from Erik Brynjolfsson, Danielle Li, and Lindsey Raymond, tracked 5,179 customer-support agents using a generative AI assistant. Productivity rose 14% on average, and novice workers improved far more than veterans did (nber.org). Stanford’s 2025 AI Index says a growing body of evidence shows AI can boost output and narrow skill gaps, while business adoption has surged, with 78% of organizations reporting some AI use in 2024, up from 55% a year earlier (hai.stanford.edu). The catch is that these gains show up most clearly in narrow, structured tasks. The broader the job, the more supervision rushes back in. That supervision is not a side issue. It is the job design problem. A 2025 overview in *Policy and Society* lists the now-familiar burdens of generative AI governance: hallucinations, data leakage, opacity, control failures, and the need for adaptive oversight (academic.oup.com). The International Labour Organization has made a similar point in blunter terms. Its work on generative AI argues that the biggest effects may land not on job counts but on job quality, including work intensity and autonomy (ilo.org). Once a tool cannot be trusted on its own, every output needs a watcher. Once every worker becomes a watcher, attention becomes the scarce resource. That is why the mood split inside companies has gotten so sharp. CNBC cited a January 2025 Section survey in which 74% of C-suite leaders said they felt excited about AI, while 68% of individual contributors said they felt anxious or overwhelmed (cnbc.com; businesswire.com). Executives see a capacity machine. Workers see another inbox, another interface, another draft that cannot be trusted until they inspect it line by line. Anthropic’s Economic Index, updated on March 24, 2026, suggests that much of real-world use still looks more like augmentation than full delegation, with people collaborating with the model instead of simply handing work over (anthropic.com). The technology is not magic. It is a coworker that never gets tired of producing plausible mistakes.