A growing body of research is challenging the central sales pitch of generative AI: that it saves workers time. Multiple studies published in recent months find that AI tools often create new categories of labor, from verifying outputs to fixing errors, that can offset or even exceed the hours they were supposed to reclaim. The pattern holds across white-collar professions, software development, and manufacturing, raising hard questions for organizations that have bet heavily on AI-driven efficiency.
Faster Drafts, Heavier Mental Loads
The most direct evidence comes from a peer-reviewed qualitative study of professionals using generative AI on the job. Researchers found that while AI tools can speed up an initial draft, they shift work from doing to managing, a process that includes learning the tool, crafting effective prompts, and, above all, verifying the accuracy of what the model produces. The study, published in Business and Information Systems Engineering, introduces a framework it calls “effort management” to describe this phenomenon. Time-to-first-draft may fall, but cognitive load rises because workers must now evaluate and correct machine-generated material rather than simply producing their own.
That finding resonates with a separate empirical field study of software developers. Researchers observed programmers using tools like GitHub Copilot and chat-based assistants during real work tasks. Developers reported perceiving higher productivity, yet the study documented higher cognitive load from monitoring and context-switching between their own reasoning and the AI’s suggestions. The gap between feeling faster and actually being less burdened is significant. It suggests that speed gains can mask a quieter form of work intensification that does not show up in simple output metrics.
Both studies point to the same underlying dynamic: generative AI tools often move workers into a supervisory role. Instead of directly performing a task, they orchestrate and audit a machine that is capable of making confident, fluent, and sometimes dangerously wrong assertions. That supervisory work is cognitively demanding and, in many cases, continuous. A single overlooked hallucination in a contract, report, or codebase can negate any time “saved” earlier in the process.
4.5 Hours a Week Spent Fixing AI Mistakes
If the academic research describes the mechanism, survey data puts a number on the cost. A recent poll of workers by Zapier found they spend an average of 4.5 hours per week revising, correcting, or redoing AI outputs. That is more than half a standard workday lost each week to cleanup, a figure that complicates the narrative of net time savings.
The consequences extend beyond wasted hours. The same survey reported that flawed outputs have led to rejected work, security incidents, and customer complaints. These are not hypothetical risks. They represent real organizational costs that rarely appear in vendor pitch decks or internal adoption dashboards. When a marketing team submits AI-drafted copy that contains fabricated statistics, or when a customer-facing chatbot gives incorrect information, the resulting damage requires human intervention that was never part of the original time budget.
In many workplaces, this cleanup work is also unevenly distributed. More experienced employees may find themselves acting as de facto “AI editors,” reviewing colleagues’ drafts for subtle errors the tools introduced. That can shift hidden labor onto senior staff whose time is already scarce, eroding any headline gains from faster drafting at the individual level.
New Tasks That AI Creates
One of the less discussed effects of AI adoption is the generation of entirely new work categories. A National Bureau of Economic Research working paper studying Danish workers provides some of the clearest evidence on this front. The researchers linked representative worker surveys on chatbot adoption with administrative labor market records covering wages and employment across 11 occupations exposed to conversational AI. Their finding: AI use can generate new tasks such as prompting, editing, and monitoring that did not exist before the tools arrived.
That last point deserves attention. If AI were truly eliminating large swaths of work, wages and employment in exposed occupations should show measurable shifts. The fact that they do not, even as adoption spreads, suggests something more subtle is happening. The tools are not so much replacing human effort as rearranging it. Workers spend less time on some tasks and more time on others, with the net effect washing out at the aggregate level. The much-promised productivity revolution, at least for now, looks more like a productivity reshuffle.
These new tasks also require different skill sets. Knowing how to frame a problem so an AI system can handle it, how to spot when the model has gone off track, and how to integrate its output into existing workflows are all forms of expertise that must be learned. That learning curve is itself a cost, especially in organizations that lack dedicated training or clear guidelines on when and how to use generative tools.
Manufacturing’s Productivity Paradox
The pattern is not limited to knowledge work. Research from MIT Sloan examining AI adoption in industrial settings identifies a productivity paradox in manufacturing: companies are rolling out AI-based systems on factory floors at increasing rates, yet output gains frequently lag behind expectations. The study indicates that certain types of firms outperform others in translating AI investment into results, but the broader trend is one of adoption running ahead of realized benefits.
This echoes a historical pattern. When personal computers entered offices in the 1980s, economist Robert Solow famously observed that “you can see the computer age everywhere but in the productivity statistics.” The current wave of generative and industrial AI may be following a similar trajectory, where the gap between installation and impact takes years to close. Until organizations redesign workflows, retrain staff, and adjust management practices around the new capabilities, much of the potential remains stranded.
In manufacturing, that redesign often involves rethinking maintenance schedules, quality-control procedures, and decision rights. An AI system that predicts equipment failures may only deliver value if managers trust its recommendations enough to change production plans. Without those complementary changes, AI becomes another dashboard to monitor, one more stream of information for humans to reconcile with existing processes.
Why the Dominant Narrative Falls Short
Most coverage of AI productivity treats the question as binary: either AI saves time or it does not. The research tells a more complicated story. Tools can genuinely accelerate specific subtasks, particularly first-draft generation and pattern detection. But they simultaneously introduce verification burdens, context-switching costs, and entirely new task categories that the original productivity estimates never accounted for.
The deeper problem may be structural. Many organizations deploy AI tools within fragmented technology ecosystems where different platforms do not share context. A worker using one assistant for email drafting, another for data analysis, and a third for code generation must repeatedly re-enter information across each system. Every handoff between tools is a potential point of failure, and every failure requires human correction. The 4.5 hours of weekly cleanup time reported in the Zapier survey likely reflects not just model error rates in isolation, but the compounding effect of disconnected environments that force workers to serve as the integration layer between systems that cannot talk to each other.
This has direct implications for how companies should evaluate AI investments. Measuring only time-to-first-draft or the number of automated interactions misses the real picture. A more honest accounting would track end-to-end task completion time, error rates, rework, and the distribution of cognitive load across teams. It would also consider the opportunity cost of workers spending hours each week on oversight and cleanup instead of higher-value activities.
For leaders, the emerging research offers a clear message: generative AI is unlikely to be a simple shortcut to efficiency. It can be powerful when carefully integrated into workflows, paired with training, and supported by realistic expectations about oversight. But without that groundwork, organizations risk trading visible speed for invisible strain, moving faster on the surface while workers quietly spend more time than ever keeping the machines in line.
More from Morning Overview
*This article was researched with the help of AI, with human editors creating the final content.
