How Much of Your Job Can AI Actually Do? About a Fifth
Google published something last week that’s worth your ten minutes, because it answers a question most surveys dodge: not are people using AI at work, but which parts of their work.
The answer is broad and shallow. AI activity turned up in 68% of detailed occupations — jobs representing 88.4% of employed US civilian workers. So yes, it has reached your job. But in the median occupation, it touched just 21% of the tasks in it. And only 3% of occupations showed AI use across more than three-quarters of their tasks.
Everywhere, an inch deep. That combination is the whole story, and it cuts against both of the loud narratives.
What 21% looks like on a Tuesday
The study is called Google’s Activity, Task, Landscape and Adoption Study. Researchers took 14,653,926 de-identified AI conversations from a two-week window in April 2026 and mapped them onto a standard catalogue of more than 800 occupations and 4,000 individual work tasks — so instead of asking people what they do with AI, they matched what people actually asked for against the tasks their job is made of. , and the mapping is what makes it useful. Most adoption research tells you a percentage of workers. This tells you a percentage of work.
A fifth of your tasks is not a small thing. It’s also not your job. In practice it looks like the parts of the week that are text-shaped and low-stakes: drafting, summarising, reformatting, looking something up, having a concept explained, getting unstuck on a first sentence. The parts that were never the reason anyone hired you.
What it doesn’t look like is the other four-fifths — the judgment, the relationships, the sitting in a room reading how people are reacting, the knowing which of two defensible options is the right one here.
”Median” is the word doing the work
Be careful with that 21%, because a median hides a spread.
Usage isn’t evenly distributed, and the direction may surprise you: a 1% increase in an occupation’s median earnings correlated with more than a 2.5% increase in usage intensity. Control for education and the coefficient is still 1.86. The higher-paid your work, the more AI you’re likely using.
That’s the opposite of the story where AI arrives to do the low-status work first. It’s showing up fastest in exactly the jobs people assumed were safest, and it’s being used there as an accelerant rather than a replacement.
Where it stops being augmentation
Here’s the part I’d resist skipping, because it’s where the comfortable reading breaks down.
The study separated tasks into routine and Work that can’t be reduced to a set of structured instructions — the study’s examples are hypothesis testing and creative design. Routine cognitive work, by contrast, is anything fully specifiable as rules and steps. The distinction matters because it predicts what people are trying to get AI to do: help, or finish. . Then it looked at what people were trying to do in each.
In non-routine cognitive work, attempting to hand the task over end-to-end was the intent of fewer than 10% of conversations. Almost everyone was collaborating.
In routine cognitive work, over 25% of conversations were aimed at full automation. Two and a half times the rate.
So “AI is a collaborator, not a replacement” is true on average and much weaker if your work is the kind that can be written down as a procedure. The average is doing a lot of reassuring here, and whether it applies to you depends on which side of that line your tasks sit. Most people’s work is a mix. It’s worth knowing your own ratio.
One more figure in the same direction: non-routine cognitive work is 35% of all catalogued tasks but attracted 65% of the interactions. People reach for AI most where the work is least mechanical — and least automatable.
What this study can’t tell you
I’d rather hand you the limitations than let you over-read the numbers, and Google is refreshingly direct about them. Their chief economist called these “early observations,” and the paper says the findings “should not be considered definitive.”
Specifically:
- It measures asking, not achieving. The data “cannot show whether the intended output was achieved or how effective the exchange was.” Someone attempting to automate a task is not someone who succeeded.
- It’s Gemini, and only some of Gemini. It excludes paid API usage, Google Workspace, AI Overviews, Translate, and Maps. So it misses much of the enterprise traffic where the serious workplace automation would live.
- It’s current use, not potential. It shows what people are doing with these tools, not what the tools could do.
- The task labels are probabilistic — assigned by an AI classifier, which the authors call “a preliminary attempt.”
- Two weeks in April 2026. A snapshot, in a fast-moving year.
And the detail that reframes everything above: over 86% of the interactions in the sample happened outside formal work at all. People are mostly using this stuff for their own lives — purchases, appliances, taxes, licences, fines. The workplace story is a slice of a much bigger domestic one.
The takeaway
Two things are true at once, and you need both. AI has arrived in essentially every occupation — and in the typical one it’s doing about a fifth of the tasks, mostly alongside a person rather than instead of one.
So the useful question isn’t “will this reach my job.” It has. It’s which fifth, and you’re better placed to answer that than any study: the routine, specifiable, text-shaped parts of your week are where it lands first and hardest. Go find the two or three tasks in your own job that fit that description and hand those over deliberately, before somebody else decides on your behalf which fifth it should be.
Then notice what’s left. The four-fifths that don’t compress is a reasonable description of what your job is actually for.
Sources: Google’s AI & Economy ATLAS v1.0 (14,653,926 de-identified interactions sampled 6–19 April 2026, mapped to 800+ occupations and 4,000 tasks), reported by Axios, PPC Land, and Implicator.ai. All figures above are Google’s. The argument that you should choose your own fifth before it’s chosen for you is mine, not theirs — the study describes behaviour and makes no recommendations.
Related: AI Role Redesign: Every Job Is Changing, Almost No One Is Rewriting Them is the management-side version of this problem — what happens when the fifth changes but the job description doesn’t. AI Productivity Gains: Everyone’s Using It, Almost No One Is Saving Time explains why a fifth of your tasks getting faster doesn’t automatically show up as a fifth of your week back.