The current debate around AI at work often jumps straight to replacement. A new study from Google Research points to a more measured reality: workers are using AI, but most are not handing over entire jobs to it.
The study, called the AI & Economy ATLAS, looked at 15 million anonymized AI interactions across the Gemini App, Google’s AI Mode, and the Gemini API. Its central finding is that workplace AI use is real and widespread in some areas, but it remains limited in depth and mostly collaborative in nature.
What Google Measured
Google researchers examined work-related Gemini interactions and classified them against the Bureau of Labor Statistics’ Standard Occupational Classifications and O*NET’s more detailed database of specific work interactions. Because real prompts can be ambiguous, the classification involved some probability-based judgment, but human reviewers found the method reliable for understanding how Gemini was being used for work.
The study did not simply ask whether a worker used AI. It looked at the occupational tasks connected to those interactions and asked a more useful question: how much of a job’s actual task structure is showing meaningful Gemini use?
That distinction matters. A worker might use AI every day for one narrow part of a job, such as drafting, research, or checking a document. That is different from AI performing the full range of tasks that make up the occupation.
AI Use Is Uneven Across Jobs
The data shows stronger Gemini use in some white-collar fields than in others. Occupations in computers, finance, and arts and entertainment were overrepresented compared with their share of the US economy. Financial/market analysts, software developers, and systems administrators were among the heavier users for job-related tasks.
Other roles appeared much less often in the Gemini work data. Salespeople, transportation workers, and food preparation/service workers were heavily underrepresented. That does not mean these jobs have no possible AI use. It means the measured Gemini interactions did not show the same level of workplace adoption.
The researchers also found that many jobs have barely crossed the threshold for measurable AI involvement. Across the O*NET task database, only 21 percent of all work-related tasks were classified as “Gemini tasks,” meaning they had at least 25 related interactions in the sample.
- For 29% of occupations, no relevant work task met that threshold.
- For another 30 percent of occupations, less than one-quarter of tracked tasks showed significant Gemini usage.
- Only 3 percent of occupations had Gemini regularly consulted for at least three-quarters of relevant tasks.
The small group with the broadest task-level use included software quality assurance analysts and testers, human resources specialists, and document management specialists. Even there, the study describes use of Gemini across tasks, not proof that the occupations have been fully automated.
Most Workplace AI Use Looks Collaborative
The researchers’ conclusion is direct: they “[did] not find evidence… to support the claims that AI is about to cause massive automation and displacement of white-collar work…” Instead, the study describes AI as a complement to human work.
In the researchers’ words, “AI is currently serving primarily as a complement to existing work” and “AI appears useful for a subset of tasks performed within occupations, but they do not currently appear to be comprehensively used for performing the work currently done by humans.”
That framing is important because it separates task assistance from job replacement. If Gemini is used for a portion of a worker’s activities, the worker may still be responsible for judgment, coordination, context, final decisions, and the parts of the job that do not map cleanly onto model output.
The study also leaves room for change. The researchers note that new AI breakthroughs could alter the current pattern. They also say future workflows may continue to keep workers and AI systems in complementary roles.
The Tasks Workers Give AI
When the study looked beyond occupations and into task types, cognitive work dominated the Gemini sample. Cognitive tasks accounted for 86 percent of the measured Gemini interactions by volume. Interpersonal and manual tasks were underrepresented compared with their presence in the workplace.
Still, the data did include examples from more manual jobs. Industrial machinery mechanics used Gemini for “analyzing test results and machine error messages.” Auto mechanics used it in tens of thousands of conversations involving “testing vehicle components and systems, rewiring systems, and inspecting parts for wear.” These workers were more likely than others to provide Gemini with a photo instead of text.
For cognitive work, the largest categories were not full automation. The study found many significant Gemini tasks connected to “drafting and generation” of ideas or “information retrieval and learning.” A smaller share involved automation of work tasks, even when the tasks were considered “routine.”
The researchers also found that workers tended to use AI more for lower-expertise cognitive tasks. Examples in the source include rewriting material in different languages and writing and reviewing product specifications. The pattern suggests that people are less likely to rely on Gemini for the most complex parts of their jobs.
What This Means For The Future Of Work
The clearest implication from the study is that workplace AI adoption should be judged at the task level, not only through broad claims about occupations. A job can contain dozens of different activities. Some may be easy to support with AI, while others still depend heavily on human skill.
The current Gemini data points to augmentation more than displacement. Workers appear to use AI to speed up or support parts of their work, especially routine cognitive tasks, while also collaborating with AI on non-routine cognitive work.
The researchers describe this as evidence that employees are not “automating themselves out of existence.” Instead, the observed pattern points to “greater returns to human skill in [the] non-routine dimensions” that still make up much of many jobs.
That does not settle the long-term question. The study itself notes that the picture could change if future models become more useful for high-expertise tasks or if AI-powered robots become better at manual tasks. For now, though, this large sample of real Gemini use shows a workplace where AI is active, useful, and uneven, but not yet taking over most jobs end to end.