Google ATLAS Report: AI Collaborates, Rarely Automates Jobs

Google published the first edition of its "AI & Economy ATLAS" report on July 23, 2026, and the top-line finding pushes back on the idea that generative AI is quietly automating work out from under people. Less than 10% of workplace AI interactions in Google's dataset actually complete a task end to end. The rest cluster around ideation, strategy, information retrieval, and learning, according to the report on Google's blog.

The scale of the sample is what makes it worth reading. ATLAS v1.0 draws on 15 million aggregated, de-identified human-AI interactions across the Gemini App, AI Mode, and the Gemini API, which together have more than a billion monthly active users. Google's team mapped those interactions across 800 occupations and 4,000 tasks, using a methodology it calls Observation Clustering and Taxonomy Organisation, or OCTO. The work is co-led by Zanna Iscenko, AI & Economy Lead in the Chief Economist's Office, and Scott Strand, Head of StratOps and Special Projects in Technology & Society.

The numbers underneath the headline are where it gets interesting. AI shows up in 68% of occupations, which together cover 90% of US employment, but only touches about 21% of tasks inside a typical job. Creative design and hypothesis testing skew much higher, with roughly 65% of those interactions involving AI against a 35% baseline. Manual and technical trade workers are about twice as likely to use multimodal AI, a useful hint about where the phone-camera modality is actually landing. And 86% of all AI interactions happen outside work entirely.

The honest caveat is that this is a Google-authored report about Google's own products, released while the "AI is going to eat jobs" story is running hot. The framing that automation is "uncommon" today does not tell you what the ratio looks like once agentic workflows mature, and the report does not break its numbers down by industry, firm size, or seniority, nor does it benchmark against enterprise surfaces where task completion is often the design goal. Take the specific percentages as one company's read of its own traffic, not a settled labor-economics finding.

What is worth taking away is the shape of the data more than the exact digits. If the collaborative pattern holds even directionally, the strategic move for a leader is less about headcount math and more about identifying which slice of that 21% of tasks in your organization is already getting reshaped, and whether the people doing them are being trained to work at the top of the partnership rather than the bottom.