Google’s 15 Million Gemini Interactions Show Broad AI Use, Limited Automation

Google's first ATLAS study of 14.65 million Gemini interactions from April 2026 found AI use associated with 68% of occupations but only about 21% of tasks in the median occupation, with fewer than 10% of non-routine cognitive work interactions seeking end-to-end automation, complicating forecasts of rapid job replacement.

Google’s 15 Million Gemini Interactions Show Broad AI Use, Limited Automation - Google analyzed 14.65 million interactions from the Gemini app, AI Mode and Gemini API during a two-week period in April 2026. 1 https://ai.google/static/documents/GoogleATLASv1.pdf 2 https://blog.google/innovation-and-ai/technology/research/understanding-the-ai-economy/ - Google found some AI use associated with 68% of occupations, but in the median occupation with observed use, it touched only about 21% of tasks. 1 https://ai.google/static/documents/GoogleATLASv1.pdf 2 https://blog.google/innovation-and-ai/technology/research/understanding-the-ai-economy/ - Fewer than 10% of interactions involving non-routine cognitive work appeared to seek end-to-end automation. That is a measure of user intent, not successful automation. 1 https://ai.google/static/documents/GoogleATLASv1.pdf - ATLAS measures activity on selected Google products, not changes in employment, wages, hours worked or productivity. 1 https://ai.google/static/documents/GoogleATLASv1.pdf 5 https://www.brookings.edu/articles/research-on-ai-and-the-labor-market-is-still-in-the-first-inning/ Google’s first large-scale ATLAS study of real-world artificial intelligence use found that Gemini has reached a wide range of occupations without being used across most of the tasks inside them, complicating forecasts of rapid, economy-wide job replacement. The company’s AI & Economy ATLAS v1.0 examined 14,653,926 de-identified interactions from the Gemini app, AI Mode in Search and the Gemini API between April 6 and April 19, 2026. Google mapped the conversations across more than 4,000 tasks and roughly 800 occupations, covering more than 150 countries and 140 languages. 1 https://ai.google/static/documents/GoogleATLASv1.pdf 2 https://blog.google/innovation-and-ai/technology/research/understanding-the-ai-economy/ The resulting picture is broad but shallow. Google detected work-related AI use associated with 68% of occupations, collectively representing 88.4% of U.S. employment. Within the median occupation where any use appeared, however, Gemini interactions touched only about 21% of its constituent tasks. 1 https://ai.google/static/documents/GoogleATLASv1.pdf Assistance dominates the observed work conversations ATLAS uses Google DeepMind’s Observation Clustering and Taxonomy Organisation system, known as OCTO, to summarize and classify unstructured conversations. Automated systems separated work from non-work activity and mapped work interactions to occupational tasks. Non-work conversations were compared with categories from the Bureau of Labor Statistics’ American Time Use Survey. 1 https://ai.google/static/documents/GoogleATLASv1.pdf 3 https://www.axios.com/2026/07/23/google-ai-adoption-work-atlas The report groups work into categories including routine and non-routine cognitive and manual tasks. Non-routine cognitive work, including creative design, analysis and hypothesis testing, accounted for 65% of work interactions, compared with about 35% of tasks in the broader economy. 1 https://ai.google/static/documents/GoogleATLASv1.pdf 2 https://blog.google/innovation-and-ai/technology/research/understanding-the-ai-economy/ Across workplace interactions, people most often sought help with research, drafting, troubleshooting, learning, strategy or iteration rather than asking Gemini to complete a task without further human involvement. Fewer than 10% of conversations involving non-routine cognitive work appeared to seek end-to-end automation. 1 https://ai.google/static/documents/GoogleATLASv1.pdf 2 https://blog.google/innovation-and-ai/technology/research/understanding-the-ai-economy/ That figure is narrower than a general workplace automation rate. It describes the apparent intent of conversations in one task category, not whether Gemini completed the work successfully. The report also found a higher automation-intent share for routine cognitive tasks, making the type of work being measured material to the result. 1 https://ai.google/static/documents/GoogleATLASv1.pdf The data captured use around physical occupations as well. Automotive technicians and industrial mechanics used image-capable models to interpret test results, inspect equipment and troubleshoot wiring. When people in predominantly manual or technical occupations used Gemini for work, they were twice as likely to use multimodal features as other workplace users. 2 https://blog.google/innovation-and-ai/technology/research/understanding-the-ai-economy/ A map of usage, not a measure of job losses ATLAS cannot show that jobs are safe from AI. It observes what people asked selected Google products to do during two weeks; it does not establish whether an answer was accurate, adopted by a worker, incorporated into production or followed by changes in staffing. The report contains no direct measures of output, wages, working hours, hiring or layoffs. 1 https://ai.google/static/documents/GoogleATLASv1.pdf Its occupational reach also should not be read as a worker adoption rate. Finding some activity associated with 68% of occupations does not mean 68% of workers in those occupations use Gemini. More than 86% of the interactions in the dataset were classified as non-work activity. 1 https://ai.google/static/documents/GoogleATLASv1.pdf 2 https://blog.google/innovation-and-ai/technology/research/understanding-the-ai-economy/ The sample excludes Google Workspace and Gemini Enterprise, where company-managed deployments could produce different behavior, including more workflow automation. Google said the necessary logs were not available for those products. AI Overviews, Google Translate and other AI-enabled services also fell outside the study. 2 https://blog.google/innovation-and-ai/technology/research/understanding-the-ai-economy/ 3 https://www.axios.com/2026/07/23/google-ai-adoption-work-atlas Independent research cautions that platform logs and worker reports can produce different task maps. Economists Alexander Bick, Adam Blandin, David Deming and Tyler Schumacher found that chat logs tend to overclassify basic, generic requests when assigning them to the federal O NET occupational framework. Their nationally representative survey nevertheless found generative AI use in 80% of occupations and across 40% of job tasks, usually with adoption below 50%. 4 https://www.frbsf.org/research-and-insights/publications/system-research-st-louis-fed/2026/07/what-work-does-generative-ai-do/ Employment evidence is mixed and more narrowly targeted Labor economist Jed Kolko has emphasized that AI usage, potential exposure and actual labor demand are separate measures that frequently rank occupations differently. He described the current employment evidence as inconclusive and warned that findings can be sensitive to definitions and classification methods. 5 https://www.brookings.edu/articles/research-on-ai-and-the-labor-market-is-still-in-the-first-inning/ Evidence tied directly to hiring remains less dramatic at the aggregate level. Federal Reserve researchers examining Census adoption surveys and millions of online job postings through November 2025 found no reduction in total postings among firms or industries with higher AI adoption. They cautioned that losses in specific occupations could be offset by hiring for other roles inside the same companies. 6 https://www.federalreserve.gov/econres/notes/feds-notes/ai-adoption-and-firms-job-posting-behavior-20260327.html More granular payroll research has identified a possible pressure point. Stanford Digital Economy Lab researchers reported a 16% relative employment decline among workers ages 22 to 25 in the most AI-exposed occupations, even after controlling for firm-level shocks. Employment for more experienced workers in the same occupations and workers in less-exposed fields remained stable or grew. The authors presented the pattern as consistent with an AI effect, not definitive proof of causation. 7 https://digitaleconomy.stanford.edu/publication/canaries-in-the-coal-mine-six-facts-about-the-recent-employment-effects-of-artificial-intelligence/ A June review by International Labour Organization economists similarly concluded that large-scale displacement remained limited, while identifying weaker opportunities for younger workers as an emerging risk. It also found that task-level time savings had not yet translated consistently into measurable gains in output, earnings or employment. 8 https://www.ilo.org/publications/impact-genai-jobs-productivity-and-work-organization-review-empirical Google describes ATLAS as an ongoing project but has not announced a publication schedule. Repeated releases could reveal whether collaborative use develops into deeper automation. The first edition establishes where Gemini is being used; employment and firm-level data are still needed to establish what that use does to workers. Companies mentioned Further sources 1 Google, “Google’s AI & Economy ATLAS v1.0: Mapping Gemini Usage in the Economy,… ↗ https://ai.google/static/documents/GoogleATLASv1.pdf 2 Google, “Understanding the AI economy,” July 23, 2026. ↗ https://blog.google/innovation-and-ai/technology/research/understanding-the-ai-economy/ 3 Axios, “Google: AI adoption at work is broad, but automation remains limited,” … ↗ https://www.axios.com/2026/07/23/google-ai-adoption-work-atlas 4 Alexander Bick, Adam Blandin, David J. Deming and Tyler Schumacher, “What Work … ↗ https://www.frbsf.org/research-and-insights/publications/system-research-st-louis-fed/2026/07/what-work-does-generative-ai-do/ 5 Jed Kolko, “Research on AI and the labor market is still in the first inning,” … ↗ https://www.brookings.edu/articles/research-on-ai-and-the-labor-market-is-still-in-the-first-inning/ 6 Jessica Liu and Douglas Webber, “AI Adoption and Firms’ Job-Posting Behavior,” … ↗ https://www.federalreserve.gov/econres/notes/feds-notes/ai-adoption-and-firms-job-posting-behavior-20260327.html +2 more The stories that matter, in one email. Free — unsubscribe anytime.