On 23 July 2026, Google published the first edition of ATLAS, a study built on around 15 million anonymised human-AI interactions, covering more than 150 countries, 140 languages, 800 occupations and 4,000 tasks. The headline finding runs against the prevailing narrative: AI is used in 70% of occupations, but each worker only uses it for around 21% of their tasks, and fewer than 10% of interactions actually automate something.
There is a lot of talk about what AI could do at work, and rather little about what people actually do with it. That is precisely the gap Google is trying to fill with ATLAS, short for Activity, Task, Landscape and Adoption Study. The scale is unprecedented, and the results deserve close attention, including for what they do not show.
The method, and its scale
The study draws on around 14.65 million anonymised and aggregated interactions, collected over two weeks in April 2026, across the Gemini app, AI Mode in Google Search and the Gemini API – products with more than a billion monthly users in total.
Automated systems classified each interaction as work-related or not, then linked the professional activity to 4,000 tasks spread across 800 occupations, in 150 countries and 140 languages. Personal uses were mapped to the categories of the American Time Use Survey. The taxonomy was produced by an analytical framework developed by Google DeepMind. Google says ATLAS will become a multi-year research programme.
The central finding: broad but shallow
Here is the conclusion that matters, and it is a nuanced one. Adoption is very broad: AI use touches around 70% of occupations, which represent roughly 90% of US employment. In other words, almost no occupation remains entirely untouched.
But that same adoption is very shallow. On average, a worker only uses AI for around 21% of their core tasks. And crucially, fewer than 10% of work-related interactions fully automate a task. The vast majority fall under collaboration: information retrieval, idea generation, drafting, iteration, troubleshooting, learning.
| Indicator | ATLAS figure | What it means |
|---|---|---|
| Occupations touched | ~70% | Adoption is near-universal |
| Tasks per worker | ~21% | But usage remains sporadic |
| Full automation | fewer than 10% | Replacement is rare |
| Non-routine cognitive tasks | 65% of AI uses vs 35% in the economy | AI serves mostly creative and complex tasks |
That last figure may be the most counter-intuitive. The assumption was that AI would first absorb repetitive, mechanical tasks. The opposite is happening: tasks classified as non-routine cognitive, such as creative design or hypothesis formulation, appear almost twice as often in AI uses as they do in the economy at large.
First, professional use of AI is not reserved for white-collar workers: the study notes real assistance in predominantly physical and manual occupations, on the administrative and adjacent tasks that accompany the core work. Second, on geography: per-capita use broadly tracks national income levels, but several middle-income economies in South America and the Middle East show adoption rates comparable to those of rich countries. The access gap is not a mechanical inevitability.
The limitations that need naming
Three caveats apply, and the first is obvious: it is Google studying its own products. The study measures usage of Gemini, not ChatGPT, Claude or specialised business tools. The same user may well reserve Gemini for quick questions and hand heavier tasks to another tool.
Second caveat, and it is a significant one: the dataset excludes enterprise telemetry, notably Google Workspace, AI Overviews, Gemini for Google Cloud and Gemini Enterprise. Yet it is precisely in enterprise deployments that automation would be most visible. Measuring automation while excluding enterprise usage is like looking for your keys where the light is on.
Third caveat: two weeks in April 2026 is a snapshot. In a field where capabilities shift every month, a photograph is not a trajectory. Google acknowledges this, presenting the report as an early view of a rapidly moving landscape.
Worth noting: these findings align with earlier work from Anthropic and OpenAI, which strengthens them, even if Anthropic for its part reported a higher level of automation.
What this changes in the jobs debate
This study lands in a context of high tension. In recent weeks we have covered 8,000 job cuts at Meta and 1,800 at Allianz, both justified by AI efficiency gains. How do we reconcile those announcements with a study claiming automation remains rare?
The two realities can coexist, and that is the most honest reading. AI does not need to fully automate an occupation to eliminate jobs. If it makes a team 20% more productive on part of its tasks, management can decide to do the same work with fewer people. Full automation is rare, but partial productivity gains are enough to justify reorganisations. This is exactly what we noted about Meta: AI acts less as a frontal replacement than as an accelerator, and sometimes as a justification.
Still, this report provides a useful counterweight to the catastrophist narrative. Across 15 million real conversations, people overwhelmingly use AI as a collaborator, not a replacement. They ask it to help them think, learn, get started. That is a less spectacular use than the promised great replacement, but probably more revealing of what is actually happening in offices. For now.