Google ATLAS Report: AI Touches 68% of Jobs

Google has put a number on how widely artificial intelligence has crept into working life, and the headline is that the technology is everywhere and nowhere at once. On July 23, 2026, Google published the first edition of its AI & Economy ATLAS, a large-scale study built from 15 million de-identified interactions across the Gemini App, AI Mode in Search, and the Gemini API. The finding that drew the most attention: AI now touches 68% of occupations, but within any given job it is used for only about a fifth of the work — and it fully automates fewer than 10% of tasks.

ATLAS — short for Activity, Task, Landscape, and Adoption Study — is Google’s answer to a question that has hung over the AI boom since it began: past the hype and the capital spending, what is the technology actually being used for? The report frames itself as an ongoing, multi-year research program, and its first release is less a verdict than a baseline.

What the study measured

Google says ATLAS draws on 15 million aggregated and de-identified human-AI interactions spanning 150 countries, 140 languages, 800 occupations, and roughly 4,000 distinct tasks. The underlying products — the Gemini app, AI Mode, and the API — are used by more than 1 billion people monthly, giving Google an unusually wide lens on how AI is being put to work.

The methodology maps conversations onto occupational task frameworks, then classifies each interaction by the kind of work it supports and whether the AI is doing the task or helping a human do it. Crucially, Google stresses that the data is de-identified and aggregated: the study is designed to describe patterns across the economy, not to profile individual users.

The result is a map that reaches deep into the labor market. The 68% of occupations that show AI usage represent roughly 90% of US employment, meaning the technology has found its way into nearly every corner of the workforce — even if unevenly.

Broad reach, shallow penetration

The more revealing number is how little of each job AI actually handles. Within any given occupation, workers deploy AI tools for an average of 21% of their core responsibilities, according to the study. That is a striking gap: AI is present in two-thirds of jobs, but in each of those jobs it does only about a fifth of the work.

Google’s read is that AI is behaving far more like a collaborator than a substitute. People overwhelmingly turn to Gemini for research, drafting, iteration, troubleshooting, and learning — the connective tissue of knowledge work — rather than handing over whole roles. The report’s language is deliberate: this is augmentation, not replacement, at least for now.

That framing echoes categories popularized by rival labs. Google’s report explicitly references the augmentation-versus-automation distinction, noting that augmentation accounts for the majority of interactions — about 52% of conversations — against roughly 45% classified as pure automation. In other words, most of the time a human stays in the loop, using the model to accelerate a task rather than to complete it unattended.

The tasks people actually bring to AI

Beneath the topline, the study clusters usage into a handful of recurring activities. The dominant patterns are collaboration and ideation — brainstorming, structuring arguments, generating options — followed by retrieval (finding and synthesizing information), strategy (planning and decision support), and learning (explaining concepts, teaching skills). What people are not doing, on the whole, is delegating end-to-end responsibility for a job.

That distribution matters because it cuts against the simplest version of the automation story. If workers were using AI mainly to offload complete tasks, you would expect the automation share to dominate and the per-occupation task coverage to climb toward full jobs. Instead, ATLAS describes a workforce reaching for AI in narrow, high-frequency ways — the moments where a first draft, a quick summary, or a second opinion saves time.

Google positions this as evidence that today’s models slot into existing workflows rather than dissolving them. It is a more measured picture than either the boosters or the doom-callers tend to paint, and Google clearly wants the data — rather than the rhetoric — to define the conversation.

Geography tracks income — mostly

The report also maps adoption across borders, and the pattern is largely what economics would predict: per-capita AI utilization aligns closely with national income levels. Wealthier countries, with more knowledge work, faster connectivity, and higher AI familiarity, use the tools more intensively.

But there are exceptions worth noting. Several middle-income economies in South America and the Middle East posted adoption rates matching those of higher-income markets, suggesting the usual income gradient can be leapfrogged where demand and access line up. Google reads this as an early sign that AI diffusion may not follow the slow, wealth-gated path of some past technologies — a hopeful note for the broader debate about who benefits from the AI build-out.

The productivity question ATLAS doesn’t answer

For all its breadth, ATLAS is careful about what it does not claim. Usage is not the same as impact, and the study stops short of asserting that AI is measurably lifting output. Critics were quick to seize on that gap: if AI is present in 90% of US employment yet the macro productivity statistics remain stubbornly ordinary, the two facts sit uneasily together.

That tension is the real subtext of the report. Widespread adoption with modest per-task penetration is exactly the profile you would expect from a technology still early in its diffusion curve — powerful in specific moments, not yet reorganizing entire jobs. Whether that changes as models grow more capable, more agentic, and more trusted with unsupervised work is the question ATLAS is designed to track over time.

The framing also invites comparison with other attempts to measure the same phenomenon. Rival economic indices built on different model usage have reported a nearly even split between automation and augmentation overall, with automation dominating programmatic API traffic — where businesses hand a model context and let it run — even as consumer chats lean toward augmentation. Google’s data fits that shape: individual users collaborate; systems automate. The distinction between how humans and how software invoke these models may end up mattering more than any single headline percentage. For readers new to the underlying technology, our explainers on what an LLM is and reasoning models provide the background.

What it means

ATLAS is, above all, a positioning move. Google is one of the largest AI spenders on the planet — it recently raised its 2026 capital-expenditure forecast to as much as $205 billion, a number that has investors openly nervous — and it has an obvious interest in shaping how the world understands what all that money buys. A study showing AI woven into 90% of US employment is a useful backdrop for a company defending historic levels of infrastructure spending.

Read skeptically, the report is a reminder that reach and impact are different things. AI in 68% of occupations sounds transformative; AI doing 21% of the tasks in those occupations sounds like a very good productivity tool that has not yet remade the economy. Both can be true, and the space between them is where the next few years of the AI debate will play out.

For workers, the near-term takeaway is reassurance with an asterisk. The current pattern is augmentation — models as collaborators, not replacements — but the report’s own automation figures are large enough that the trajectory, not the snapshot, is what matters. If automation’s share keeps climbing, especially through the API channel where software calls software, the collaborator story could shift faster than the headline suggests. The same dynamic is already visible in software, where AI coding assistants have moved from autocomplete to autonomous agents in under two years.

The most important thing about ATLAS may simply be that it exists as a recurring measure. The AI economy has been debated largely in anecdotes and projections; a standing dataset that tracks real usage across hundreds of occupations, updated over years, is the kind of instrument the debate has lacked. What to watch is version two — whether the per-task penetration climbs, whether automation gains on augmentation, and whether the productivity numbers finally start to move to match the adoption ones. Until they do, ATLAS’s central image holds: AI is everywhere, and still doing a surprisingly small slice of the work.