Google Research analyzed 15 million anonymized Gemini interactions across the Gemini App, Google's AI Mode, and the Gemini API and concluded that white-collar workers are not automating themselves out of a job. Only 21% of tracked work-related tasks cleared the study's minimum threshold to be classified as a 'Gemini task.' For 29% of occupations, not a single relevant task hit that bar. The paper, released last week and titled the AI & Economy ATLAS, is the largest first-party disclosure to date of how a frontier model is actually being used at work.
The study's framing matters because it comes from Google, a company with every commercial incentive to overstate AI's workplace penetration. Instead, the authors write that Gemini use 'remains shallow and overwhelmingly collaborative in nature, with end-to-end task automation limited in scope.' The researchers mapped prompts against the Bureau of Labor Statistics' Standard Occupational Classifications and O*NET's task database, then had human reviewers validate the automated classifier.
The occupation breakdown lines up with intuition but not with the loudest predictions. Financial and market analysts, software developers, and systems administrators dominate Gemini usage relative to their share of the US workforce. Salespeople, transportation workers, and food preparation staff are heavily underrepresented. Cognitive tasks accounted for 86% of interactions by volume, with interpersonal and manual work far below their real-economy weight.
Key facts
- 01Google Research analyzed 15 million anonymized Gemini interactions across the Gemini App, AI Mode, and Gemini API.
- 02Only 21% of tracked work tasks met the 'Gemini task' threshold of 25+ related interactions in the sample.
- 0329% of occupations saw zero tasks reach non-negligible Gemini usage; another 30% saw fewer than one-quarter.
- 04Just 3% of occupations use Gemini regularly for at least three-quarters of their relevant tasks.
- 05Cognitive tasks made up 86% of measured Gemini interactions, concentrated on low-expertise, non-routine work.
Depth is the more striking finding. For another 30% of occupations, fewer than one-quarter of tracked O*NET tasks saw significant Gemini use, meaning humans still do the overwhelming majority of the component work. Only 3% of occupations use Gemini regularly for at least three-quarters of their relevant tasks — a group that includes software quality assurance analysts and testers, human resources specialists, and document management specialists.
Even inside the jobs where Gemini shows up, the model is not doing the hard parts. Google's team measured task complexity using the entropy of O*NET descriptions and found that low-expertise, non-routine work is heavily overrepresented in the sample. Rewriting material in different languages, drafting product specifications, and information retrieval dominate. High-expertise cognitive work — the parts of the job that actually justify the salary — is largely absent from the prompts.
Blue-collar use exists but looks different. The researchers found thousands of industrial machinery mechanics prompting Gemini for 'analyzing test results and machine error messages,' and tens of thousands of auto mechanics using it for 'testing vehicle components and systems, rewiring systems, and inspecting parts for wear.' These workers were far more likely to attach a photo than to type a paragraph, a usage pattern that suggests multimodal input is doing real work in the trades even when text-first metrics undercount it.
“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.”— Google Research, AI & Economy ATLAS study authors
The researchers stop short of predicting the trend line. They allow that patterns could shift 'as new AI breakthroughs emerge,' particularly if models improve on high-expertise tasks or if AI-powered robots close the gap on manual work. But the base case in their data is that new workflows 'will maintain a degree of complementarity between workers and AI systems' — humans keeping the non-routine, high-expertise core of the job while offloading the routine cognitive edges.
The counterweight is real. This is a single-vendor snapshot; it does not capture Claude, ChatGPT, or Copilot usage, and enterprise deployments running through custom pipelines may look very different from consumer Gemini App traffic. Employees who fear being flagged as replaceable have obvious reasons not to run their most sensitive automation through a logged consumer product. And the study measures current usage, not future capability — a gap that closes fast in this market.
The market implication cuts against the more aggressive automation narratives that have driven a year of hiring freezes and layoffs justified by AI leverage. If Google's own telemetry shows that even its most active users are handing off the shallow end of their jobs, the productivity gains being priced into enterprise software valuations are real but narrower than the pitch decks suggest. Vendors selling agentic end-to-end automation now have to explain why their internal numbers should look different from the ones the model provider just published.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
