AI has spread across 68% of occupations globally. The median worker uses it for 21% of their tasks.
That is the central finding of Google's AI & Economy ATLAS v1.0, published July 23, 2026. The study, built from 14,653,926 de-identified interactions across the Gemini App, AI Mode, and the Gemini API, is the first systematic, large-scale audit of how artificial intelligence actually penetrates the global economy. It draws a hard line between adoption theater and genuine integration. The market has not priced in the difference.
The 21% Ceiling
Google mapped usage against over 800 occupations and 4,000 tasks. The sampling window ran from April 6 to April 19, 2026, capturing queries across 143 languages that cover 93% of first-language speakers among the world's 200 most spoken tongues. Usage appears in 217 countries and territories representing 99% of the world's population. Occupations touched by AI account for over 88% of US civilian employment.
But the headline number conceals the real story. Within any given occupation where meaningful usage appears, the median share of constituent tasks performed with Gemini is 21%. AI is everywhere, and it is thin.
"Adoption of AI is very, very broad in that it touches a huge range of occupations," Google economist Scott Strand told Axios. "AI use is also 'very shallow,' with the average worker using it for only 21% of tasks."
For 29% of detailed occupations, the study observed zero task saturation. No single task drew at least 25 users. That group includes stockers and order fillers, food preparation workers, fast food cooks, refuse collectors, and police and sheriff's patrol officers. These jobs are structurally resistant to current AI. The work is physical, regulated, or both.
The Ideation Trap
At work, AI is not automating. It is brainstorming.
Less than 10% of Gemini interactions are geared at automating non-routine cognitive tasks. The dominant use case is ideation and strategy. Workers use AI as a thought partner, not a replacement. In manual and technical trades, the pattern is even more pronounced: AI functions as a live collaborator for real-time troubleshooting.
This explains the shallow penetration. Companies deployed AI broadly but never reengineered workflows around it. A tool that helps a lawyer brainstorm a brief is not the same as a tool that drafts, reviews, and files it. The ATLAS data reveals a gap between access and integration that most enterprise AI strategies have not closed.
Why 21% Is a Ceiling, Not a Floor
The consensus interpretation treats the 21% figure as a starting point: AI is new, adoption takes time, depth will come. That is wrong.
The 21% median task saturation is not a floor. It is a hard ceiling at the current technological frontier. AI has already seized the low-hanging cognitive fruit: ideation, strategy, analysis, text generation. The remaining 79% of tasks require physical embodiment, regulatory permission, or both. Better language models will not move the needle on fast food cooking. Better robotics and a different liability framework might.
The real bottleneck is not intelligence. It is integration.
English queries represent only about a third of global volume, a detail that punctures the Silicon Valley assumption that AI adoption looks the same everywhere. It does not. But even in wealthy, English-dominant economies, the 21% ceiling holds. Meanwhile, 86% of all AI interactions in ATLAS occur outside of work, spanning activities that make up about 98% of Americans' non-sleep time. The consumer AI economy is already deep. The enterprise AI economy is not. That asymmetry will define the next phase.
The Pivot from Reach to Depth
Within 12 to 24 months, the market will pivot from celebrating AI's reach to scrutinizing its depth. The ATLAS report provides the baseline, and the baseline is damning for companies that confused deployment with transformation.
Here is the mechanism. Firms that spent the last two years rolling out broad, splashy AI features will face a board-level reckoning. The question shifts from "how many employees have access?" to "what did it actually replace?" The 68% adoption number becomes a liability when the follow-up is a margin call. Investors will demand proof of task-level ROI, and generalist chatbots that draft emails for 68% of occupations but automate nothing in any of them will be exposed as commodities. The market will reward specificity.
The winners will be specialized AI firms that target the 21% of tasks with proven saturation. Vertical tools for legal document review, medical coding, insurance claims processing, logistics routing. These are narrow, unglamorous, and measurable. A startup that can prove it automates 40% of the tasks in a single occupation will be more valuable than a platform that touches 68% of jobs at 5% depth.
That valuation gap triggers the second-order effect: an M&A wave. Buying depth is faster than building it. Expect Big Tech to acquire task-specific AI startups aggressively over the next 18 months. Google, which now has a map of exactly which tasks are untouched and the distribution network to reach them, will not leave that territory uncontested. It will double down on vertical AI tools for the 29% of occupations with zero saturation, particularly manual trades where real-time collaboration is already happening.
The third-order consequence is labor market bifurcation. Jobs with deep AI integration will see higher productivity and higher wages. Jobs with zero saturation will see stagnant wages and persistent labor shortages, not because AI replaced the workers, but because it never reached them to make them more productive. The gap between the integrated and the untouched will widen. If this prediction is wrong, we will see the 21% figure rise above 30% within 18 months without a corresponding wave of vertical tool acquisitions. That would falsify the depth thesis.
What Operators Must Do Now
For CEOs and CTOs, the mandate is clear. Audit your own task saturation. Identify the specific tasks where AI actually works inside your organization and double down on those. Ignore the 68% metric. It is a vanity number.
If you cannot name the three tasks where AI has measurably changed output per worker, you have not integrated it. You have only adopted it. The difference will show up in margins.
For investors, the ATLAS report provides a filtering mechanism. Look for startups that can prove task-level replacement in specific occupations. Ask for the saturation number. If a founder cannot tell you what percentage of tasks their product automates within a given job category, they are selling reach, not depth. Reach is cheap.
For policymakers, the 29% of zero-saturation jobs demands attention. These are the occupations least likely to be automated and most vulnerable to being left behind. They are also the jobs where labor shortages are already acute. Policy that treats AI as a uniform threat to employment misses the point. The real problem is uneven distribution of productivity gains.
A Mirror, Not a Failure
Scott Strand called the ATLAS report "just an early sketch" and "the beginning of a long-term body of work" on LinkedIn. James Manyika, Google's SVP of Research, Technology and Society, echoed the sentiment: "ATLAS v1.0 is a beginning."
They are right that the dataset is a snapshot. But it captures something structural. The 21% ceiling is not a failure of AI. It is a mirror held up to the economy, showing exactly where AI can reach and where it cannot.
The question for the next 24 months is not how many jobs touch AI. It is how many tasks AI actually owns. The ATLAS report is the baseline. The real story begins when the market starts measuring depth, not breadth.
