AI Reaches 90% of Workers, 21% of Their Tasks

The number that undercut a year of headlines

The finding: Scott Strand, an economist on Google's technology and society team, put a figure on a debate that has run on prediction for two years. In research Google published from its ATLAS work, drawn from 15 million de-identified Gemini interactions, adoption of AI at work is, in his framing, very broad but very shallow. The tool now shows up across 68 percent of occupations, covering 90 percent of US employment - and in a typical job it is used for only about 21 percent of tasks.

The second number is the one that matters for anyone planning a budget. Fewer than 10 percent of workplace AI interactions fully automate a task; the vast majority are collaborative - ideation, drafting, information retrieval, troubleshooting and learning. The company with the strongest commercial reason to claim AI replaces workers measured its own product doing something narrower and more useful: helping people work, not standing in for them.

Broad is not the same as deep

Why it matters: the gap between reach and depth is exactly where planning goes wrong. It is easy to read that 90 percent of employment is touched and conclude the workforce is being automated; the same dataset says the average worker hands AI a fifth of their tasks and keeps the rest. Breadth measures curiosity and access. Depth measures substitution, and only the second one justifies cutting a role.

Yes, but: the shallow figure is not a ceiling forever, and some tasks move faster than others - creative design and hypothesis testing show up in AI work at nearly twice their baseline rate. The point is not that automation will never deepen; it is that today, on the measurer's own data, it has not, and a decision made now should price the present, not a promised future.

The decision this should change

The bottom line: the most expensive mistake available to an owner right now is restructuring headcount against an automation rate that does not yet exist. When a vendor's pitch and the vendor's own measurement disagree, the measurement is the base rate - and the base rate here is a fifth of tasks assisted, not jobs removed. A hiring freeze or a layoff justified by "AI will handle it" is a bet against the numbers the seller just published.

This is a base-rate discipline, not a mood. The right frame is the one economists like Autor and Coyle bring: ask what the evidence shows the technology doing today, in aggregate, and size the decision to that - not to the most confident vendor slide or the most anxious headline. The augmentation case is real and worth funding; the replacement case, on current data, is not yet a plan.

What to do with a 21 percent number

Do this: treat 21 percent as a working ceiling for task-level automation and build your next 12 months on augmentation - give teams the tools, measure which tasks actually move, and let evidence, not the roadmap, trigger any structural change. Fund training before you fund redundancy, because the study's real signal is that value comes from people using AI well, not from removing them.

For European operators there is a second reason to go slow on headcount: works councils, collective-agreement rules and the EU AI Act's workplace-transparency duties mean a restructuring justified by automation you cannot yet demonstrate is both a financial and a legal risk. Document what the tools actually do in your own operation, revisit the number each quarter, and act when your data - not Google's, and not a competitor's - shows the depth has changed.

Read next: Google Has 60 Days to Let You Bill Outside Its Store  |  Google Already Sold $514bn of Cloud Capacity