Google Read 15 Million AI Conversations — 'AI Touches 68% of Jobs, But Only 21% of Their Tasks'

Seventeen researchers across Google and Google DeepMind published "Google's AI & Economy ATLAS v1.0: Mapping Gemini Usage in the Economy" on July 23. ATLAS stands for Activity, Task, Landscape, and Adoption Study — an attempt to measure how AI is actually used in the economy directly from anonymized conversation data. It belongs to the same research lineage as Anthropic's Economic Index, which analyzed Claude usage; here Google raises the stakes with a far larger sample and broader coverage.

What they looked at — 14.65M conversations, 800 occupations, 150 countries

The dataset is 14,653,926 anonymized interactions sampled from the Gemini App, Google AI Mode, and the Gemini API between April 6 and April 19, 2026. Google mapped these conversations onto more than 800 detailed occupations and 4,000 work tasks built on the U.S. Department of Labor's O*NET framework, plus 300 household activities. Geographically it spans 150 countries and 140 languages. The methodological core is that it aggregates not by individual chat but by "which task in which occupation is AI being used for."

Wide, but shallow — what '68% vs 21%' means

The most-quoted figure is 68%. Across the world's detailed occupations, 68% show AI usage above a minimum threshold — equivalent to 88.4% of U.S. civilian employed workers. On the surface, that reads as "AI has reached nearly every job."

But the number Google emphasizes more is 21%. Even within occupations where AI use appears, the share of tasks AI actually touches is a median of 21%. If a job is made up of 20–30 tasks, AI has entered only five or six of them. Reach and depth are different stories, and ATLAS's value is that it separates the two.

Not replacement, but collaboration — automation intent under 10%

The third key point is automation intent. By classifying what users actually want from the model, Google found that fewer than 10% of non-routine cognitive conversations aimed to hand off a task end-to-end. The large majority were augmentation — collaboration, ideation, strategy, information retrieval, and learning. The dominant pattern is people working with AI, not handing work to it.

That conclusion throws empirical data at the usual "replace vs. augment" framing of the AI-and-labor debate. The caveats are real, though: the sampling window is a short two weeks in April; a user's "intent" need not be full automation to still reduce labor over time; and because Google analyzed usage of its own products, the "collaboration over automation" conclusion happens to align with a narrative that flatters the company.