One of the central questions around workplace AI is whether the technology is spreading across entire jobs or remaining concentrated in selected tasks.
Google’s first AI & Economy ATLAS report suggests a mixed pattern: Gemini activity appeared across 68% of the detailed occupations in its employment analysis, covering jobs that account for just over 88% of U.S. employment, but the median occupation with recorded use crossed Google’s threshold for only 21% of its O*NET (Occupational Information Network) tasks.
The findings are based on nearly 15 million aggregated and de-identified interactions across the Gemini app, Google AI Mode and the Gemini API. Google collected the sample from April 6 to 19, 2026 and mapped the interactions to more than 800 occupations and 4,000 tasks using Bureau of Labor Statistics occupation codes and the O*NET task database.
Broad occupation reach but limited task penetration
The activity spread across many occupations but crossed Google’s usage threshold for only a minority of listed tasks in the median occupation. Only 3% of occupations recorded Gemini activity in more than 75% of their constituent tasks.
Software quality assurance analysts and testers, human resources specialists and document management specialists were among the small group above the threshold.
High adoption in nonroutine cognitive and technical tasks
Workplace use was concentrated in nonroutine cognitive activity. Such tasks represent about 35% of tasks in O*NET but accounted for almost 65% of work-related interactions in ATLAS. Google’s preliminary intent classifier categorized less than 10% of Gemini conversations mapped to nonroutine cognitive work as end-to-end task automation.
Partial drafting, review, idea development and information retrieval accounted for most of the remainder. More than one-quarter of routine cognitive interactions, by contrast, targeted automation.
The activity extended into technical and manual occupations, although nearly one-third of heavily physical occupations showed no observed use. Gemini interactions mapped to automotive service technicians and industrial machinery mechanics included test-result interpretation, electrical troubleshooting and machinery inspection.
Multimodal conversations among automotive service technicians and mechanics occurred at more than twice the rate recorded across workplace interactions overall.
Census data shows augmentation outpaces employment changes
Firm-level context comes from a Center for Economic Studies working paper using nationally representative data from the Census Bureau’s 2026 Business Trends and Outlook Survey AI supplement.
During the supplement’s November 2025 to January 2026 reference period, 18% of firms reported using AI in a business function during the previous two weeks. The employment-weighted estimate was 32%. Separate questions covering the previous six months found that, among firms reporting any effect on worker tasks, 66% reported augmentation alone. Separate U.S. small-business polling has also found employers reporting changes to roles, performance expectations and hiring decisions.
A larger share of AI-using firms reported replacing software or equipment than changing employment. Sixteen percent reported replacing existing software or equipment with AI-integrated systems, while about 5% reported any employment change, split nearly evenly between increases and decreases. Recent Federal Reserve surveys and labor-market trackers have also pointed to altered hiring plans, with limited evidence of widespread job losses.
The working paper’s regression summaries found no statistically significant relationship between the breadth of worker-task use and headcount reductions after accounting for functional deployment and operational investment.
Broader use across business functions and greater implementation investment were associated with employment decreases, although the authors specified that the analysis carried “no causal interpretation.”
Anthropic dataset highlights contrasting automation patterns
A separate dataset from AI developer Anthropic recorded different patterns by access channel. Anthropic’s November 2025 data classified 52% of Claude.ai conversations as augmentation and 45% as automation. Among its first-party API customers, 75% of records were classified as automation and 74% as work-related.
The percentages are not directly comparable with ATLAS because Anthropic used a different automation taxonomy and analyzed different interaction formats and task populations. Anthropic has separately used Claude activity data to examine occupational exposure and labor-market outcomes.
Study exclusions and limitations on productivity claims
ATLAS does not include task-level content from paid Gemini API use, including enterprise use through Google Cloud. It also excludes Google Workspace, Gemini Enterprise and several other Google products. The report states that enterprise professional use may consequently be underrepresented.
Google also cautioned that ATLAS measures “behavioral interactions, not definitive productivity outcomes.” A completed conversation does not establish whether a task succeeded, saved time or produced economic value. ATLAS records where Gemini appeared in work, while the Census findings capture firms’ reported deployment and workforce effects. Neither study establishes that AI caused productivity gains or employment changes.