Google Study Documents Multimodal AI Use in Trades
Google released its first ATLAS study on July 23, analyzing about 15 million de-identified interactions across the Gemini app, AI Mode and Gemini API. It found AI activity across 68% of occupations representing 90% of U.S. employment, but only about 21% of tasks in a typical job. Manual and technical workers who used the tools were twice as likely to use multimodal AI.
Google released the first version of its Activity, Task, Landscape, and Adoption Study, or ATLAS, on July 23. The company analyzed about 15 million aggregated and de-identified interactions across the Gemini app, Google Search's AI Mode and the Gemini API to describe how people use its AI tools at work and outside it.
The results extend beyond office work. Google says people in manual and technical occupations, including automotive technicians and industrial mechanics, use conversational AI for diagnostics, troubleshooting and learning. When workers in those occupations use Google's tools, they are twice as likely to use multimodal AI involving images or video as the average workplace user.
Broad occupational reach, limited task depth
Google found workplace AI activity in 68% of occupations, covering occupations that represent 90% of U.S. employment. Within a typical job, however, the tools appeared in only about 21% of tasks. Fewer than 10% of workplace interactions fully automated a task; most involved assistance such as research, ideation, troubleshooting or learning.
Those measures answer different questions. Occupational reach shows how many kinds of work appear somewhere in the dataset. Task coverage shows how selectively people use AI within a job. Neither figure means 68% of workers actively use Gemini, nor does the study show that AI performs 21% of all work without human involvement.
What the dataset does not capture
Axios reports that ATLAS v1.0 classified 14.65 million interactions collected over two weeks in April across 800 occupations, 4,000 tasks, 150 countries and 140 languages. Automated systems classified the interactions and mapped work activity to occupational tasks.
The sample is large but bounded. It covers activity inside Google's observed consumer tools and API, not the entire labor market. Axios also reports that Gemini Enterprise and Google Workspace activity was excluded because Google does not retain the required logs. The study is therefore a snapshot of use in selected Google products, not a complete measure of enterprise AI adoption, productivity or job displacement.
Why the trades finding matters
For product teams, the finding suggests that multimodal interfaces can reduce the burden of translating physical evidence into text. A technician can use images of equipment, wiring or test results as part of a troubleshooting exchange. That can make AI assistance more accessible in work where the relevant information begins outside a document or dashboard.
LDS interpretation: teams building these tools should evaluate more than answer quality. Image capture conditions, grounding in approved manuals, uncertainty handling and escalation to a qualified worker determine whether a diagnostic assistant is useful in practice. ATLAS documents an emerging interaction pattern; it does not validate any particular model for safety-critical repair decisions.
Key Points
- 1ATLAS analyzed about 15 million interactions across the Gemini app, AI Mode and Gemini API, with enterprise Workspace activity excluded.
- 2Google observed AI activity across 68% of occupations but only about 21% of tasks in a typical job, while full automation represented less than 10% of workplace interactions.
- 3Manual and technical workers using Google's tools were twice as likely to use multimodal AI, pointing to image-assisted diagnostics and troubleshooting.
Scoring Rationale
ATLAS provides a large observational view of AI use across occupations and highlights multimodal workflows in manual trades. Its relevance is tempered by its short collection window, automated classification and exclusion of enterprise Workspace activity.
Sources
Primary source and supporting public references used for this report.
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.
Try 250 free problems
