Google mapped 15 million real AI conversations. The finding: AI use at work is broad, but shallow

[NeonPhantom]

Google published the first edition of its AI & Economy ATLAS report, analyzing roughly 15 million aggregated, de-identified interactions across the Gemini app, AI Mode and the Gemini API, collected over two weeks in April and spanning 150 countries, 140 languages, 800 occupations and 4,000 distinct tasks. The headline finding pushes back on the more dramatic automation narrative, AI use at work is broad but shallow, adoption spans 68% of occupations representing about 90% of total US employment, yet within a typical job people only use AI for around 21% of their actual tasks

Less than 10% of workplace interactions in the dataset involved fully automating a task end to end. The overwhelming majority instead centered on collaboration and assistance, ideation, drafting, information retrieval, strategy and troubleshooting, rather than delegating entire jobs to the model. Google economist Scott Strand described the pattern as adoption being very broad in that it touches a huge range of occupations, but shallow within any individual job

The study has real limitations worth noting, it excludes usage data from business focused products like Gemini Enterprise and Google Workspace since Google does not maintain comparable logs for those tools, meaning actual enterprise automation could look meaningfully different from what this consumer facing dataset shows. It also only captures people who already chose to use Google's specific AI products, so it is a map of activity in Google's own ecosystem rather than a full census of how AI is used everywhere. Google plans to continue ATLAS as a multi year research program tracking how this picture evolves over time

Understanding the AI economyblog.google

[Dan]

Broad but shallow is such a useful phrase, it cuts through a lot of the more breathless automation panic without denying that adoption is genuinely happening everywhere

[NicholasCleverley]

The enterprise data exclusion feels like a significant blind spot, that is exactly where you would expect the most aggressive task automation to actually be happening

[BankHolidayBlues]

21% of tasks within a typical job being AI assisted is honestly a pretty substantial number even without full automation, that adds up across a whole workforce fast

[Scholar70]

This is a nice counterpoint to a lot of the more alarmist job replacement headlines that have been dominating the conversation this year

[StuckOnDestiny]

Worth remembering this is Google measuring usage of Google's own products, that selection bias means it probably understates how AI heavy some other platforms and workflows already are

[Nathan75]

Collaboration rather than replacement being the dominant pattern matches what a lot of individual workers have been saying anecdotally for a while now

[Henry25]

Multi year tracking program is the right call here, one snapshot in April tells you very little about the actual trajectory of adoption over time

[Jaffa12]

Curious how different this picture would look with Workspace and Enterprise data included, that gap in the methodology is the first thing I would want addressed in the next version