Platform research · Google
AI Is Used for Just 21% of Tasks in the Average Job
Google's first ATLAS study, built from 15 million anonymised AI conversations, finds AI has reached almost every kind of job but is still used for only about a fifth of the tasks within them, and it collaborates far more often than it automates.
Source published Summarised
Key findings at a glance
- 15 million aggregated, de-identified human-AI interactions in the first ATLAS dataset (v1.0)
- 1 billion monthly users of the Gemini app, AI Mode and API that the study draws on
- 150+ countries and territories where ATLAS recorded AI use, covering 99% of the world's population
- 800 occupations covered by the analysis of how people use AI
- 4,000 distinct work and life tasks mapped in the dataset
- 68% of occupations now show some workplace AI use, spanning 90% of US employment
- 21% share of tasks in a typical job where AI is actually used, so adoption is broad but shallow
- <10% of work AI interactions fully automate a task; most use is collaborative
- 65% of AI work interactions involve non-routine cognitive tasks such as design and hypothesis testing
- 35% share of non-routine cognitive tasks across the wider economy, well below their share of AI use
- 86% of all AI interactions happen outside of work, including researching purchases and admin
- 2x trades workers such as mechanics are twice as likely to use multimodal AI as knowledge workers
What the report covers
Google has published the first release of its AI and Economy ATLAS (Activity, Task, Landscape, and Adoption Study), described as the most comprehensive look so far at how people actually use AI. The v1.0 dataset is built from 15 million aggregated, de-identified interactions across the Gemini app, AI Mode and the Gemini API, tools used by more than a billion people each month. It spans over 150 countries, 140 languages, roughly 800 occupations and 4,000 tasks.
The headline is that workplace AI is broad but shallow. Adoption now touches 68% of occupations, covering the vast majority of employment, yet within any given job AI is used for only about a fifth of tasks. Most of that use is collaborative: ideation, information retrieval, drafting and learning, rather than handing whole tasks to the machine. Fewer than one in ten work interactions fully automate a task.
ATLAS also shows how much AI use sits outside the office. More than 86% of interactions happen in daily life, from researching purchases to navigating admin such as tax and licensing. That domestic use rarely shows up in economic statistics, but it matters for anyone trying to reach customers. For a UK view of how far business adoption has come, see our report on UK business AI adoption.
What this means for UK marketers
For UK marketers, ATLAS reframes the AI question. The story is no longer whether people use AI, they clearly do, it is how shallowly most of them use it. Adoption reaching 68% of occupations while touching only a fifth of tasks tells you the value now sits in depth, not access. Turning a free chatbot licence into a repeatable workflow that moves a metric is precisely where an agency or in-house team earns its keep. If you want that built into your marketing, it is the core of our digital marketing services.
The collaboration finding matters for content teams. AI shows up far more in non-routine cognitive work, 65% of work interactions versus 35% across the economy, and it rarely automates. Treat it as a drafting and ideation partner, not a replacement: the value is shifting from production to judgement, brief-writing, brand consistency and editing the model's output.
The most commercially important number may be that 86% of AI use happens outside work. People are increasingly researching purchases and solving problems by asking an assistant, not by scanning a results page. That makes being visible and citable inside AI answers a discovery channel in its own right, which is why we treat generative engine optimisation as part of our SEO work. It is also worth noting that English accounts for only about a third of global AI conversations, so brands chasing international reach cannot assume an English-first funnel will carry them. British firms weighing where to invest can sense-check this against the ONS view of AI use in UK businesses, which shows adoption wide but still thin on the ground.
How they got the data
ATLAS v1.0 draws on 15 million aggregated and de-identified human-AI interactions from the Gemini app, AI Mode and the Gemini API, which together serve more than a billion users a month. The insights span more than 150 countries and territories representing 99% of the world's population, 140 languages, around 800 occupations and 4,000 tasks. The data is organised using Google DeepMind's Observation Clustering and Taxonomy Organisation (OCTO), which distils large volumes of unstructured conversation text into grouped categories.
Google describes several privacy safeguards, including removing personally identifiable information, stripping references to sensitive topics, severing links between the ATLAS data and the underlying user logs, and aggregating summaries across multiple users. The figures are labelled official statistics in development and reflect usage of Google's own tools rather than the whole AI market, so they are best read as a directional signal rather than a definitive market share.
Platform research
Read the full report
From Google · published 23/07/2026