Google mapped 14,653,926 de-identified Gemini interactions to more than 800 occupations and 4,000 O\*NET tasks. AI appeared in 68% of detailed occupations โ roughly 88% of employed US civilians. Inside a typical one of those occupations, it touched 21% of the tasks (Google ATLAS v1.0, 2026).
Two-thirds of jobs. One-fifth of the work.
And where the work is non-routine and cognitive โ judgment, analysis, problem-solving, the category your expensive people are hired for โ end-to-end automation is the apparent intent of fewer than 10% of conversations. The other 90% is drafting, review, retrieval, and thinking out loud.
If your AI business case runs on role elimination, this is the number that breaks it. Not because the tools are weak, but because the arithmetic was built on the wrong unit. A fifth of a job is not a fifth of a headcount.
What ATLAS Measured, and Why It Lands Differently
Google's AI & Economy ATLAS (Activity, Task, Landscape, and Adoption Study) sampled interactions across the Gemini App, AI Mode, and the Gemini API between 6 and 19 April 2026, then mapped them onto official occupational and task taxonomies across 150 countries and 140 languages. Dame Diane Coyle (Cambridge) and Dr David Autor (MIT) contributed to the report โ Autor being the economist who spent two decades arguing that technology reshapes labour task by task rather than job by job (Google, 2026).
Most AI adoption data you have seen is self-reported. A survey asks a manager whether their team uses AI, and the manager answers the question they wish they had been asked. ATLAS is behavioural: it observes what people actually brought to the model, then classifies the task. That shift matters, because self-reported adoption and observed task coverage are not the same measurement and have never produced the same number.
The distribution is more skewed than the 21% median suggests. For 29% of detailed occupations, ATLAS observed no task saturation at all โ stockers and order fillers, food preparation workers, refuse collectors, patrol officers. At the other end, only 3% of occupations showed AI usage across three-quarters or more of their tasks: software QA analysts and testers, human resources specialists, and document management specialists (PPC Land, 2026).
Three occupations in a hundred are deeply covered. That is your automation candidate list. It is shorter than your slide says.
Google's own comparison is worth noting for calibration: Anthropic's Economic Index found roughly 36% of occupations using AI for at least a quarter of their tasks; Google's equivalent figure is around 30%, a gap the authors attribute largely to stricter privacy thresholds rather than genuinely different behaviour (PPC Land, 2026). Two independent datasets, two different vendors, same shape: wide and shallow.
Why End-to-End Automation Stays Under 10%
ATLAS classified conversation clusters into five intents: task automation, partial drafting and generation, review and refinement, ideation and strategy, and information retrieval and learning.
The result inverts the standard displacement story. Non-routine cognitive analytic tasks account for 35% of all tasks in the O\*NET taxonomy but 65% of Gemini work interactions. People reach for AI most on the hardest, least standardised parts of their job โ and it is precisely there that they least often hand the whole thing over. Fewer than 10% of those conversations show end-to-end automation intent. In routine cognitive work, more than a quarter do (Google ATLAS v1.0, 2026).
The mechanism is not mysterious. A task gets fully delegated when the output is verifiable at a glance and the cost of an error is bounded. Reformatting a dataset qualifies. Deciding which supplier to drop does not. Non-routine work is exactly the work whose correctness cannot be confirmed without redoing most of the thinking โ so the human stays in the loop, and the time saved is real but partial.
Google states the implication plainly: the data does not support claims that AI is about to cause massive automation and displacement of white-collar work. The authors also flag that the intent classification is preliminary and that behaviour may shift as capability advances. Both halves of that sentence belong in your planning assumptions.
Your Headcount Arithmetic Assumes a Depth You Do Not Have
Here is where this becomes an operations problem rather than an interesting chart.
The standard mid-market AI business case multiplies a per-task time saving by a task volume, converts hours to FTEs, and books the difference. It works arithmetically only if the saved hours are contiguous and belong to a role whose remaining work can be absorbed elsewhere. ATLAS says the saved hours are scattered across a fifth of a task portfolio, in the parts of the job that were never the constraint.
Take the median case literally. AI covers 21% of a role's tasks. Assume โ generously โ a 40% time reduction on each of those. That is an 8% reduction in the role's total task time. You have not freed a person. You have freed a fragment of everyone, distributed across the week in pieces too small to reallocate and too small to notice on a timesheet.
This is why so many pilots report enthusiastic users and flat operating costs simultaneously. Both are true. The savings are real and the savings are unbankable, because banking them requires task consolidation โ moving the remaining 79% into a redesigned role โ and no one has been assigned to do that.
The deeper trap is the 3%. When a business case gets built off a lighthouse function โ QA testing, HR document handling, high-coverage administrative work โ the coverage rate observed there gets silently extrapolated to the rest of the org. It does not transfer. Those three occupations in a hundred are structurally unusual: high task standardisation, text-native outputs, immediate verifiability. Your finance analysts and your account managers have none of those properties.
The Expertise Inversion Nobody Priced In
The finding most likely to be wrong in your current plan concerns who benefits.
The prevailing assumption is that AI hollows out juniors and amplifies seniors. ATLAS complicates it from both ends. Usage intensity rises steeply with pay โ a 1% increase in an occupation's median earnings is associated with more than a 2.5% increase in usage intensity, falling to 1.86% once you control for educational attainment. Weighted by employment, median annual earnings in the sample are $62,252; weighted by Gemini conversations, $82,919; weighted by tokens, $86,157 (PPC Land, 2026).
So senior, well-paid people are the heavy users. But scoring each task's expertise level using the Autor and Thompson methodology โ which rates rarity and domain-specificity of the vocabulary in the task statement โ usage is most over-represented on the lowest-expertise cognitive tasks. Rewriting a report into another language. Drafting a product specification.
High-expertise workers are bringing AI their low-expertise work.
That inversion is the single most actionable thing in the study, and it points the opposite way from most deployment plans. The value is not concentrated in giving your senior people a smarter thinking partner for the hard calls; it is concentrated in stripping the low-expertise residue out of expensive calendars. If your rollout is targeted at junior roles on the theory that their work is more automatable, you are deploying against the observed gradient. Adoption follows discretion, not task simplicity โ people use AI where they have the authority to decide how their own work gets done.
Where This Evidence Stops
The honest limits, because a Head of Operations should not act on a vendor's own telemetry without pricing them in.
ATLAS is Google measuring Google. It observes Gemini surfaces only, so a shop standardised on a different assistant is outside the frame. It is a two-week sample in a fast-moving market. Enterprise platforms โ Workspace, Gemini Enterprise, agentic coding tools โ sit largely outside the granular dataset, and that is precisely where deep, workflow-embedded automation is most likely to be happening. Over 86% of the interactions are non-work; work conversations account for around 13.5% of App and AI Mode traffic, so the work-related subsample, while large, is a slice.
Most importantly, ATLAS measures behaviour, not outcomes. A completed conversation tells you nothing about whether the person got what they needed or saved a minute (Forte Group, 2026).
None of that rescues the headcount case. Under-measurement of enterprise agentic tooling would raise the ceiling on where automation is heading; it does not change what a 21% median task coverage implies about the org you are running this quarter. And the direction of the bias is worth naming: this is a model vendor publishing a study that undercuts the automation narrative its own commercial interest favours. Findings that cut against the publisher's incentive deserve more weight, not less.
The Audit That Replaces the Guess
One concrete move before your next AI budget review: stop counting tools and count tasks.
Pick the three roles carrying your largest AI investment. For each, write out the task portfolio โ 15 to 25 discrete tasks, plainly stated. Mark each one against two questions: is the output verifiable in under a minute by someone qualified, and does a wrong answer cost more than redoing it. Tasks that are cheaply verifiable and cheaply wrong are your genuine automation set. Everything else is assistance, and assistance shows up as quality and cycle time, never as headcount.
Then measure the coverage rate you actually have. If it is near 21%, you are median โ normal, not failing, and not a savings line. If a role clears 60%, you have found one of the 3%: redesign that role deliberately rather than waiting for the saving to appear on its own. If a role sits under 10% after a year of licences, the tool is not the problem and neither are the people. The work simply is not shaped for it, and the licence spend belongs somewhere else.
The uncomfortable version of this finding is also the useful one. AI is in almost every job in your company and deep in almost none of them, end-to-end automation remains the exception rather than the plan, and the returns are sitting in the low-expertise residue clogging your most expensive calendars. Go take that back. It is smaller than the headline promised and far more collectable.
