Google's 15-million-conversation ATLAS study found AI touches 68 percent of jobs but automates less than 10 percent of tasks inside them, which means most rollouts are scoped to replace a role that the data says almost never gets fully replaced.
Picture the kickoff meeting. Someone on the leadership team says the quiet part out loud: we are going to replace the SDR team, or the first-draft content function, or the tier-one support desk, with an AI agent. The budget gets approved on the promise of a line item disappearing from payroll. Nine months later the project is either quietly shelved or limping along at a fraction of the promised savings, and nobody can quite explain why, because the agent works fine on the parts it was actually good at. The problem was never the agent. It was the job description used to scope it. By the end of this you should be able to look at your next AI project and tell whether it is aimed at a task or aimed at a job, and why that difference decides whether it ships.
Google just handed the industry the data to prove the point. This week it published Understanding the AI economy, the first release of ATLAS, a study built from 15 million aggregated, de-identified conversations across the Gemini app, AI Mode, and the Gemini API, spanning more than 150 countries, 140 languages, 800 occupations, and 4,000 tasks. It is the largest look yet at what people actually do with AI at work, as opposed to what vendors say they do with it.
The headline number is the one worth pinning to a wall in every planning meeting. AI use touches 68 percent of occupations, which collectively cover 90 percent of total US employment. That is the number executives hear and translate into "AI is coming for the whole job." But read the next line in the same report: within a typical job, AI is used for only about 21 percent of tasks. Adoption is broad. Depth, inside any given role, is shallow. And of the AI interactions that happen at work, the report is blunt that "task automation is uncommon" so far, with less than 10 percent of those interactions fully automating a task outright. The rest is collaboration: ideation, drafting, research, troubleshooting, the stuff that speeds a person up without removing them from the loop.
That gap between "68 percent of jobs" and "10 percent of tasks fully automated inside them" is exactly where AI rollout budgets go to die. A team scopes a project around eliminating a role. The role turns out to be twenty or thirty distinct tasks, most of which the model can meaningfully assist but not finish alone, credit-checking a proposal, remembering an edge case, judging a borderline ticket. The project either quietly redefines success as "we made this person faster" (true, and worth funding, but not what got approved) or it grinds against the 90 percent of tasks that were never going to fully automate this year, and the whole thing gets written off as an AI failure. It was not an AI failure. It was a scoping failure that pointed the model at a job title instead of a task list.
This shows up the same way across very different roles. A content team scopes "replace the freelance writer bench" and discovers the model nails first drafts but cannot yet judge which client's tone needs a harder edit, so the freelance line item never actually zeroes out. A sales team scopes "replace the SDR who qualifies inbound leads" and finds the model drafts a clean qualification note every time but still cannot make the judgment call on a lead that does not fit the pattern, which turns out to be a third of the volume. Even outside white-collar work, ATLAS found manual and technical trades using AI the same way, auto technicians and industrial mechanics leaning on conversational AI as a live collaborator for diagnostics and troubleshooting, twice as likely to use image and video input as the average worker, but nobody is pitching "replace the mechanic." The augmentation is the product. The full replacement almost never is.
Here is what pointing it at the task list actually looks like, and it costs nothing beyond the AI subscription you already pay for. Take one role you were tempted to "replace" and pull up two weeks of what that person actually did, not the job description, the real task log. Paste it into Gemini or Claude with a prompt like this: "Here is a list of tasks a [role] performed over two weeks. For each one, tell me how often it recurs, whether a model could complete it today without a human checking the output, and whether a model could meaningfully speed up a human doing it." You will get back a task list sorted by two axes: frequency and automatability. That sort is the project plan. The high-frequency, fully-automatable slice, usually a handful of tasks, is what you build first. The high-frequency, assist-only slice is where you buy time back for the person, not replace them. The low-frequency stuff is not worth automating this quarter regardless of how it feels in the room.
Look at what a well-scoped tool in this space actually promises, for comparison. This week's tool coverage flagged Profound, which watches how ChatGPT, Gemini, Perplexity and other answer engines describe a brand. It does not claim to replace a marketing team, a research analyst, or an SEO function. It does one task, checking what the models say about you, and does it continuously instead of by hand. That is the shape ATLAS is describing at scale: the fundable, shippable AI product is almost always a task, not a headcount line.
The pitfall shows up loudest in the pitch, not the build. "We are automating the support team" gets budget approved faster than "we are automating password resets and shipping status lookups," even though the second one is the honest, deliverable version of the first. The first framing sets an expectation the data says will not be met inside a normal planning cycle, ATLAS's own numbers put full task automation at under 10 percent of AI work interactions industry-wide. The second framing sets an expectation that ships in weeks and actually holds up in the retro. Watch for the tell in your own planning docs: if the success metric is "eliminate the role" instead of "eliminate these five recurring tasks," you have scoped a project against a job that, per the largest usage study available right now, barely exists as a fully-automatable unit.
Here is what to do with this before your next AI planning meeting. Pick one role your team has discussed automating and pull the last two weeks of what that person actually worked on. Run it through the prompt above in whatever model you already pay for and get the frequency-versus-automatability sort back. Take only the high-frequency, fully-automatable tasks into your next project brief, and write the assist-only tasks down separately as a "make this person faster" initiative instead of folding them into the same pitch. Bring both lists to the meeting where someone wants to talk about "replacing" the role.
AI did not shrink the job. It carved out the 10 percent of it that was always mechanical and left the other 90 percent exactly as human as it was before ATLAS ran a single query.