The narrative around AI and employment tends toward extremes. Either the technology is a minor productivity tool, or it is an existential threat to the workforce. A large-scale study from Google Research offers something more grounded: actual data on how workers are using AI, drawn from 15 million real interactions. The picture that emerges is considerably less dramatic than most headlines suggest.
What the Data Actually Shows About Workplace AI Use
The study, which Google Research calls the “AI & Economy ATLAS” (an Activity, Task, Landscape, and Adoption Study), analyzed 15 million anonymized interactions across the Gemini App, Google’s AI Mode, and the Gemini API. Researchers used the Bureau of Labor Statistics’ Standard Occupational Classifications alongside O*NET’s detailed database of work tasks to categorize how those interactions mapped onto real jobs.
The headline finding: across the full O*NET database of work tasks, only 21 percent qualified as what the researchers call “Gemini tasks,” meaning they reached a minimum threshold of 25 related interactions in the sample. For 29 percent of all occupations, not a single relevant work task hit even that modest threshold. For another 30 percent of occupations, fewer than one quarter of tracked tasks showed significant Gemini usage. Only 3 percent of occupations showed Gemini being regularly consulted for at least three quarters of that job’s relevant tasks.
Those numbers describe a technology that is present in many workplaces but has not penetrated deeply into most of them. The researchers’ own conclusion is direct: AI is “currently serving primarily as a complement to existing work,” not a replacement for it.
Where AI Is Actually Being Used, and How
The distribution of AI use across occupations is uneven in predictable ways. White-collar roles in computers, finance, and arts and entertainment are overrepresented in the Gemini interaction data relative to their share of the broader economy. Financial and market analysts, software developers, and systems administrators are among the heavier users. Salespeople, transportation workers, and food preparation and service workers appear far less frequently in the data.
The type of work being offloaded to AI is also telling. Cognitive tasks account for 86 percent of the interactions measured by volume. Within that category, the dominant uses are drafting and generating ideas, and information retrieval and learning. Automation of specific work tasks, even routine ones, represents a smaller share.
Crucially, the tasks workers are handing to AI tend to be low-expertise ones. The researchers measured task complexity using the language workers used to describe what they needed, and found that simpler tasks such as rewriting material in different languages or reviewing product specifications were heavily overrepresented. Workers appear much less likely to use Gemini for the most complex, high-expertise parts of their jobs.
This pattern holds even in sectors not typically associated with AI adoption. The study found industrial machinery mechanics using Gemini to analyze machine error messages, and auto mechanics using it to help with testing vehicle components and inspecting parts for wear. These workers were notably more likely to include photos in their prompts rather than relying on text alone, reflecting the practical, hands-on nature of their questions.
Why This Matters Beyond the Hype Cycle
This is what most coverage of AI and work gets wrong: the gap between what a technology can theoretically do and what workers actually ask it to do is enormous, and that gap tells you something important about how change really happens in labor markets.
The Google Research findings suggest that workers are not passively waiting to be automated. They are actively choosing which parts of their jobs to delegate to AI, and they are consistently choosing the lower-stakes, lower-complexity tasks. The high-expertise, non-routine dimensions of most jobs remain firmly in human hands.
The researchers themselves frame this carefully. They note that current patterns point to “greater returns to human skill in the non-routine dimensions” that still dominate most job descriptions. They also acknowledge that this could shift if future models become more capable at high-expertise tasks, or if AI-powered robotics improve enough to affect manual work at scale. Neither of those developments has arrived yet.
What the data describes, for now, is a workforce that is using AI to handle cognitive overhead: the drafting, the lookups, the routine reformatting. That frees up attention for the work that actually requires judgment, experience, and expertise. The technology is augmenting existing roles, not hollowing them out.
In Short
A Google Research study of 15 million real AI interactions found that most tasks, in most jobs, are not being automated by AI. Only 3 percent of occupations showed Gemini being used for three quarters or more of their relevant tasks. Workers are predominantly using AI for low-complexity cognitive tasks such as drafting and information retrieval, while keeping the most demanding parts of their jobs to themselves. The data does not support predictions of imminent mass displacement. It describes, instead, a technology that is useful at the margins of most jobs and central to very few.
Based on reporting from Ars Technica.