Google published ATLAS on Thursday — the Activity, Task, Landscape and Adoption Study — built from roughly 15 million aggregated, de-identified human-AI interactions across Gemini products serving more than a billion monthly users. It is the largest first-party dataset yet on how AI is actually used at work, and its central finding is that adoption is wide and shallow.
The spread
The sample covers more than 150 countries, 140 languages, 800 occupations and 4,000 tasks, drawn from a two-week window in April. AI appears in 68% of detailed occupations, which together account for about 90% of total US employment. Every industry sector is represented. English accounts for only about a third of global AI conversations.
The depth
Inside those jobs, usage is narrow. In a typical occupation AI is applied to roughly 21% of tasks. And of the work interactions Google measured, fewer than 10% fully automate a task — the rest assist, draft or advise. More than 86% of all interactions in the dataset happen outside work altogether.
What people bring to it
The work that reaches AI is skewed toward the non-routine. Tasks Google classes as non-routine cognitive — creative design, hypothesis testing — make up 65% of AI work interactions against 35% of the economy as a whole. Google economist Scott Strand put the conclusion plainly: "Just because you're using AI doesn't mean it's going to automate your job."
Read it with the source in mind
This is Google measuring Google's own products, and the company has an obvious interest in a story about augmentation rather than replacement. It is also a snapshot of a fast-moving thing: a two-week sample from April 2026, taken before the current wave of long-running agent products shipped. What it does establish is a baseline — and a much larger one than the survey data the debate has been running on.