An ongoing economic research initiative from Google and Google DeepMind, with authors including Zanna Iscenko, Scott Strand, Alex Imas, Julian Jacobs, Juan Mateos-Garcia, and James Manyika. ATLAS stands for Activity, Task, Landscape, and Adoption Study.
Data and method
Built on 15 million de-identified interactions across the Gemini App, Google AI Mode, and the Gemini API, mapped "using privacy-preserving algorithms as well as established and bespoke classification methods" to "over 800 occupations, 4000 tasks, 300 household activities, 150 countries, and 140 languages."
It is the Google counterpart to Anthropic's Economic Index, and shares its structural limitation: usage data from a single provider's products.
Findings
Workplace: broad but shallow. "AI adoption spans occupations covering just above 88% of US employment," but "penetration remains shallow and overwhelmingly collaborative in nature, with end-to-end task automation limited in scope."
That combination is the report's most consequential result, and it converges with the Remote Labor Index's 2.5% end-to-end automation rate from an entirely different method — usage telemetry rather than benchmark completion. Breadth of adoption and depth of substitution are separate quantities, and only the first is large.
Non-work usage. AI "spans activities making up about 98% of Americans' non-sleep time, with disproportionately high use in high-friction tasks such as engaging with government and professional service providers, likely delivering economic value that standard national accounts may miss." The measurement point is that unpaid friction reduction does not appear in GDP.
Global distribution. "Adoption scales with national wealth and has broad linguistic distribution, with English queries representing only around a third of volume."
Earnings and education. "Workers in occupations with higher median earnings and education levels are more intensive users of Gemini," measured as conversations per occupation divided by that occupation's employment in BLS OEWS data. This cuts against the assumption that AI substitutes first for lower-paid work — the heaviest users are in the better-paid occupations.
The report cautions that "AI-related impacts will not be cabined to routine tasks," and that "further work is also needed to understand the entanglements of AI usage and skills, as well as their downstream labor market implications."
Placement in the literature
The report situates itself against a divided evidence base: Richmond (2026) finding 18% of jobs at relatively high short-term automation risk; Handa et al. (2025) finding 57% of usage suggesting augmentation against 43% suggesting automation; Lindenlaub et al. (2026) arguing existing AI-exposure metrics fail to explain observed usage and proposing a comparative-advantage approach; and "a number of researchers [who] have found no labor-market effects of AI or determined that it was too early to tell," including Massenkoff and McCrory (2026).
Relationships
- supports: AI Labor Disruption — usage-side evidence that adoption breadth and automation depth diverge sharply
- related: Remote Labor Index: Measuring AI Automation of Remote Work (Mazeika et al., 2025) — converging finding from benchmark rather than telemetry data
- related: AI and Productivity — the national-accounts gap for unpaid friction reduction
- related: Google DeepMind, AI Divides (Literacy / Occupational / Ethico-Philosophical), AI Economic Primitives