Google's research arm published ATLAS v1.0 (Activity, Task, Landscape and Adoption Study), a first slice of what will be an ongoing program: take 15 million de-identified conversations from the Gemini App, Google AI Mode, and the Gemini API, and classify them onto the skeletons economists already use: BLS occupation codes, O*NET task lists, and the American Time Use Survey, plus country and language. The claimed coverage is large: more than 800 occupations, 4,000 tasks, 300 household activities, 150 countries, and 140 languages.
It is the same genre as Anthropic's Economic Index, a vendor mapping its own chat logs onto labor statistics, but on a much bigger canvas. Diane Coyle and David Autor advised on it, which is worth something; they do not usually lend their names to thin work. I read the methodology and limitations sections as carefully as the headline findings. Below are the numbers I trust, the ones I would discount, three interactive explainers for the core ideas, and an editorial at the end.
What the data actually shows
| Finding | Number | My read |
|---|---|---|
| Breadth of work adoption | 68% of occupations, covering ~88% of US employment | Credible; matches other vendor datasets |
| Depth in the median adopting occupation | AI touches only 21% of tasks; just 3% of occupations exceed 75% | The most honest number in the paper |
| Automation vs collaboration | End-to-end automation under 10% of non-routine cognitive work conversations | Replicates Anthropic's augmentation-first result independently |
| Wage gradient | +1% occupational median earnings → +2.5% usage intensity | Persists after education controls; AI complements expensive labor so far |
| High-friction admin arbitrage | Government/civic queries ~20× over-represented vs time spent; ~half of med/legal/fin/gov consultations happen outside 9-to-5 | Genuinely new empirical nugget; GDP barely sees this value |
| Trades usage | Multimodal use by mechanics/technicians at >2× the work baseline | Underreported elsewhere; cameras plus models fit hands-on work |
| Language stickiness | English ≈ one-third of global volume; non-native sessions run longer and heavier | Real product and cost implications; not a US-only story |
Numbers I would discount
- The "$100 billion" household figure. Explicitly conditional: if time savings average 30 minutes per week, then unpaid productivity could be worth about $100B in the US. The paper never measures the 30 minutes. It is an illustrative plug number multiplied out, and it will escape into headlines stripped of its condition.
- "98% of Americans' waking hours." Technically true because their activity taxonomy covers nearly everything people do. Rhetorically empty: it says a classification scheme exists, not that AI matters there.
- All depth figures. They are conditional on being a Gemini user who chose to ask a work question. Non-users do not appear anywhere in the sample, so "21% of tasks" is a ceiling observed among adopters, not workforce depth. The authors say this plainly in limitation three.
The structural caveats
Three things bound every claim in this paper. First, it is one company's funnel: no paid enterprise API content (country counts only), no Workspace, no AI Overviews, no Maps or Translate — so enterprise use, arguably where the real economics live, is under-represented. Second, every occupation, task, and activity label comes from probabilistic classifiers over conversational text; validation against human raters was reasonable, but fine-grained slices carry real uncertainty, which the authors concede. Third, a completed conversation proves intent, not value. The whole edifice measures behavior, and limitation two admits it. None of this makes the paper weak. It defines what kind of object it is: a diffusion map, not an impact study.
Three ways to feel the findings
Static tables flatten the interesting part of ATLAS: the gap between breadth and depth. So I built three small interactives. Figures quoted in them come from the paper. Two components also visualize interpolations or stylized shapes, and each one says so on the component itself.
1 · Breadth is wide, depth is shallow
Each row is an occupation archetype. Click through to see how much of its task set AI currently touches. The 21% median and the >75% tier are from the paper; the per-archetype depths and the collaboration/automation bars are my interpolations for illustration.
Broad sectors, shallow penetration. The typical pattern: drafting, review, ideation, lookup —
with a human still owning the task start to finish.
*Paper reports <10% for non-routine cognitive work overall; per-archetype values are my interpolations.
2 · A diffusion simulator with Google's elasticities
ATLAS reports +1% GDP per capita → +0.9% per-capita usage across countries, and a bottom quintile holding just 2% of conversations. Drag wealth to see how uneven the resulting map stays. This is my illustrative model of those elasticities, not a reproduction of their country-level data.
At this wealth level, my model puts this group near the top of global usage — while the poorest fifth of the world stays nearly invisible, consistent with the paper's 2% share for the bottom quintile.
3 · The out-of-hours arbitrage
Nearly half of medical, legal, financial, and government consultations in the sample happen outside 9-to-5. Press play: institutions sleep; the model does not. The value created in that gap is the kind national accounts struggle to record. (The dot pattern below is a stylized illustration, not plotted data.)
Watch the query dots cluster in the evenings and weekends when the office is closed. ATLAS finds government and civic topics roughly twenty times over-represented relative to the time people actually spend on them.
What holds up
Two vendors with different funnels — Anthropic's API-heavy sample, Google's consumer-heavy one — now agree on the shape: adoption is broad but shallow, and use skews collaborative rather than automating. When rivals measuring different populations land on compatible pictures, that convergence is the closest thing to ground truth this genre has produced. The wage gradient and the trades finding both survive my skepticism too. The out-of-hours administrative pattern deserves more discussion than it is getting, for reasons the editorial below takes up.
Editorial
The polite paradox
The paper opens with Solow's 1987 quip about computers being visible everywhere except the productivity statistics, and positions ATLAS as the instrument that will find where AI went. I think that framing deserves pushback. ATLAS does not resolve the Solow paradox; it documents the demand side of it. Fifteen million conversations tell you where people hope the machine helps. They cannot tell you whether it did, and the authors admit as much in their second limitation. A completed chat is a revealed preference for trying, not evidence of output.
There is also an unescapable conflict of interest, and the paper manages it better than most corporate research but cannot dissolve it. Google benefits from the conclusion "AI is everywhere, broadly adopted, quietly valuable at home." Every framing choice — the MAU counts in the introduction, the 98%-of-waking-hours line, the $100B conditional — pushes in the direction of scale and significance. To the authors' credit, the limitations section is candid enough that a careful reader can recompute the honest story. Most readers are not careful readers. The number that escapes into the world will be "$100 billion of unseen household productivity," and it will have been manufactured from an assumption dressed as a finding.
What genuinely interests me is why this paper exists at all. Anthropic built its index; OpenAI leaks usage claims through demos; Google responded with a public-goods research program advised by Coyle and Autor. I read that as narrative infrastructure rather than altruism. Whoever owns the measurement layer of AI adoption owns the policy conversation around it: which jobs are "exposed," which countries "lagging," whether automation or augmentation dominates. ATLAS is Google claiming a seat at that table before regulators build the table without them. Seen that way, the paper's real product is not the fifteen million conversations. It is the vocabulary.
My verdict: read it as the best behavioral map we currently have, treat every dollar figure as decoration, and keep one uncomfortable question in view. If AI really is collaborative, shallow, and concentrated among high earners, as both major vendor datasets now suggest, then either the productivity revolution has not started yet, or it will look nothing like the substitution story everyone is priced for. Both readings should make you humble about any forecast, including mine.
Source: Iscenko et al., Google's AI & Economy ATLAS v1.0, arXiv:2608.00038, July 2026. Quoted statistics come from the paper's main text; the adoption simulator is my own illustrative parameterization, and the per-archetype bars in Interactive 1 are interpolations, each labeled where it appears.