Understanding the AI Economy
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- Google launched ATLAS v1.0, built from 15 million de-identified human-AI interactions across the Gemini App, AI Mode, and Gemini API — products used by more than 1 billion people monthly.
- The dataset spans 140 languages, 800 occupations, and 4,000 tasks across more than 150 countries, making it the most comprehensive look at real-world AI usage to date, powered by Google DeepMind's OCTO clustering tool.
- Workplace AI adoption is broad but shallow: it spans 68% of occupations representing 90% of U.S. employment, but within a typical job covers only ~21% of tasks.
- Collaborative uses dominate the workplace: non-routine cognitive tasks (ideation, strategy, learning) account for 65% of AI work interactions versus 35% in the broader economy, and less than 10% of interactions fully automate tasks.
- AI use extends into manual and technical trades — auto technicians and industrial mechanics use conversational AI for real-time diagnostics and troubleshooting, and are 2x more likely to use multimodal AI than other workers.
- Over 86% of AI interactions happen outside of work, covering household tasks and high-friction administrative work like navigating taxes, licensing, and government services — value typically missed in standard economic metrics.
- Global adoption tracks GDP per capita, raising digital divide concerns, though some middle-income countries in South America and the Middle East are adopting AI at rates comparable to higher-income countries.
Why it matters: Google is using its own 15-million-interaction scale to publish the first empirical map of how AI is actually used, and the data subverts the automation narrative: 86% of interactions happen outside work, and less than 10% of workplace uses fully automate tasks. Researchers and policymakers now have a real baseline to argue from, with academic contributors including Dame Diane Coyle (Cambridge) and Dr. David Autor (MIT).
