ATLAS: what Google just taught us about how people really use AI

Google has just released a report that should interest anyone who, like me, follows closely how artificial intelligence is actually settling into work and everyday life. It's the first ATLAS report (Activity, Task, Landscape, and Adoption Study), a large-scale study of how people really use Google's AI tools — the Gemini app, AI Mode, and the Gemini API.

What makes this report interesting isn't the usual AI marketing talk, but the raw material behind it: 15 million anonymized, aggregated human-AI interactions, spanning more than 150 countries, 140 languages, 800 occupations, and 4,000 tasks. It's probably the most accurate snapshot available today of how people actually use AI, far from the fantasy of AI fully replacing human work.

Here are the findings that struck me as most useful, seen through my e-commerce and data lens.

1. AI use is broad, but shallow

The report shows that workplace AI adoption spans nearly every sector and touches 68% of occupations, representing 90% of U.S. employment. But dig deeper and usage is selective: within a typical job, AI is used for only about 21% of tasks.

In other words: AI is everywhere, but it isn't replacing entire jobs — it's slotting into specific tasks. That matches what we see in e-commerce: nobody hands over the whole strategy to a model, but everyone uses it to speed up an analysis, draft a brief, or test a hypothesis faster.

2. We collaborate with AI, we don't (yet) automate with it

Another striking point: the vast majority of workplace AI use is collaborative — ideation, strategy, information retrieval, learning — rather than full automation. Less than 10% of workplace interactions fully automate a task end to end.

So-called "non-routine cognitive" tasks (creative design, hypothesis testing) are notably overrepresented in AI usage compared to their actual weight in the economy (65% vs 35%). That's no coincidence: it's exactly the kind of task where an AI assistant adds the most immediate value, complementing human judgment rather than replacing it.

3. Manual trades are in on it too — and visually so

One finding I didn't expect: workers in manual and technical trades (auto technicians, industrial mechanics) use conversational AI as a real-time collaborator for diagnostics, troubleshooting, and on-the-fly learning. And these workers are twice as likely to use multimodal AI (images, video) as the average user — for example, interpreting a test result, debugging wiring, or inspecting a part for wear.

That breaks the cliché that "AI is just for white-collar work."

4. A lot of the value is created outside of work

Over 86% of interactions with Google's AI tools happen outside the workplace. Purchase research, help using an appliance, but also high-friction administrative tasks — taxes, government paperwork, various formalities. That's real economic value, largely invisible to standard macroeconomic indicators — an interesting blind spot for anyone interested in data and impact measurement.

5. Global adoption tracks wealth, with a few surprises

Unsurprisingly, per-capita AI usage closely tracks a country's wealth (GDP per capita), raising concerns about a persistent digital divide. But some middle-income countries, in South America and the Middle East, are adopting AI at rates comparable to much wealthier nations — proof the trajectory isn't set in stone.

Another detail that stood out to me: English accounts for only about a third of AI conversations worldwide. Users don't systematically switch to English for complex tasks — they stick with their own language. An encouraging signal for every non-English-speaking market, Poland included.

How ATLAS was built

Technically, ATLAS relies on an internal Google DeepMind tool called OCTO (Observation Clustering and Taxonomy Organisation), designed to turn massive volumes of unstructured text — AI conversations — into organized, actionable categories. On data protection, Google states it added several layers of protection beyond standard anonymization: removing references to sensitive information, removing any link between ATLAS data and original user logs, summarizing the text, then aggregating summaries into groups.

My takeaway

This report confirms an intuition many of us share on the ground: AI isn't replacing work, it's redistributing it task by task, very unevenly across jobs and use cases. For data-driven e-commerce and marketing, the message is clear — real value won't come from instant, total automation, but from fine-grained, task-by-task integration into existing workflows. That lines up well with my own approach: start from the data, test, measure, and only scale an AI use case once it's proven itself on a small scope.

This is just the first version of ATLAS — Google says it plans to turn it into an ongoing research program, with academic collaborations. Definitely a report to keep an eye on for anyone who wants to understand AI beyond the promotional talk.

Source: Understanding the AI economy, Google Keyword Blog, July 23, 2026