Google ATLAS Study Finds AI Assists More Than It Automates at Work
A study of 15 million interactions suggests workplace AI use is spreading widely while remaining selective, collaborative and rarely fully autonomous.
The loudest question in the AI labour debate is usually which jobs will disappear. Google’s first AI & Economy ATLAS report suggests a more immediate question: which parts of a job are people already asking AI to help with?
ATLAS—short for Activity, Task, Landscape and Adoption Study—draws on 15 million aggregated and de-identified interactions across the Gemini App, AI Mode and the Gemini API. Google says those products collectively serve more than one billion monthly users. The study spans more than 150 countries, 140 languages, 800 occupations and 4,000 tasks.
Broad adoption, shallow use
The headline finding is a gap between occupational reach and task-level depth. Workplace AI activity appears across 68% of occupations representing 90% of U.S. employment. Yet in a typical occupation, the technology is used for about 21% of tasks.
Most work interactions in the dataset involve assistance: retrieving information, learning, developing ideas or supporting strategy. Fewer than 10% fully automate a task. That does not prove that automation will remain uncommon, but it complicates the idea that present-day adoption translates directly into whole-job replacement.
The data also pushes beyond the white-collar stereotype. Google found that workers in manual and technical occupations use conversational AI for adjacent work such as diagnostics, troubleshooting and interpreting results. When those workers use Google’s AI tools, they are twice as likely to use multimodal features.
The economy outside the office
More than 86% of ATLAS interactions occurred outside work. People used AI for purchases, appliances, household administration and government services such as taxes and licensing—activities that conventional productivity measures may not capture neatly.
The global picture is equally revealing. English accounts for only about one-third of conversations in the dataset, and users do not consistently abandon their native language for complex tasks. Adoption generally rises with national income, although Google identifies middle-income countries in South America and the Middle East that perform more like wealthier markets.
What the study cannot settle
ATLAS is large, but it is not a neutral census of all AI use. It observes Google’s own products, and interaction frequency does not establish whether an answer was correct, whether work improved or how non-users are affected. Google describes the methodology as privacy-protected and de-identified, including removal of personally identifiable and sensitive information before aggregation.
The report is best read as a map of behaviour inside a major AI ecosystem. Its early message is not that automation has failed to arrive. It is that AI’s current economic footprint looks more like selective collaboration distributed across many occupations than wholesale replacement concentrated in a few.
Status
Analysis. The dataset and reported findings are official Google research. The interpretation above distinguishes observed usage from unproven claims about productivity or future employment.
Sources
Update note: Last reviewed 2026-07-24. We will revise this post as Google releases further ATLAS datasets or methodology updates.
Sources
- Google — Understanding the AI economy — official
Drafted with AI assistance from source briefs; reviewed for citation completeness and label accuracy.