Many readers who use generative AI in their daily work have likely felt a vague anxiety: "When will this task be replaced by AI?" This concern runs especially deep for routine, procedure-based work like document preparation or data entry, where fears of being displaced by AI remain persistent. On July 23, 2026, Google published a paper titled "AI & Economy ATLAS v1.0," which drew on 14,653,926 real Gemini conversation logs to present a seemingly reassuring figure: "conversations intended for full automation account for less than 10% of the total." Yet buried within the same paper is a contradictory figure—when limited to routine cognitive tasks, automation intent reaches over 25%. Whether this is reassuring or not depends entirely on whether your own job is routine or non-routine.
AI Has Penetrated 68% of Occupations, but the Median Task Actually Performed Is Just 21%
"AI & Economy ATLAS v1.0" was compiled by 18 researchers from Google and Google DeepMind (including Zanna Iscenko and Scott Strand), with Diane Coyle (University of Cambridge) and David Autor (MIT) credited as peer review contributors. The analysis is based not on surveys or Google Trends search patterns, but on anonymized conversation logs from the Gemini App, Google AI Mode, and the Gemini API themselves—14,653,926 conversations collected over a two-week period from April 6 to April 19, 2026. Given that it directly observes actual usage behavior at scale, this represents an unusually rich primary dataset for this type of labor market analysis.
According to this data, AI usage was confirmed in 68% of occupations, based on the U.S. Bureau of Labor Statistics' SOC (Standard Occupational Classification) and O*NET's detailed task taxonomy. These occupations account for just over 88% of U.S. employment, with footnote 12 of the paper specifying the figure as 88.4%. Some domestic media outlets reported this as "approaching 90% of employment," but the paper's own figure stays in the high-80s percentage range. Conversely, for the roughly 12% of occupations not covered, no Gemini usage was detected at all during this two-week observation window. Within the SOC and O*NET subclassifications, there remain areas where AI has yet to penetrate at all.
In stark contrast to this broad penetration, when looking only at occupations where AI usage was observed, the median task saturation rate—the proportion of O*NET-defined tasks within an occupation that AI is capable of performing—is a mere 21%. Breaking this down: 29% of occupations show zero task saturation, 30% show 25% or higher, 11% show 50% or higher, and only 3% reach 75% or above. Occupations at the top of the saturation ranking include software quality assurance analysts, human resources specialists, and document management specialists. The paper explicitly states that these results do not support any of three claims: large-scale automation or displacement of white-collar workers, AI's irrelevance to blue-collar labor, or the notion that AI's purpose is exclusively automation. While the breadth of penetration alone could fuel pessimistic narratives, the low task-saturation rate—found in the very same paper—pushes back against that pessimism.
An Asymmetry: Over 25% Automation Intent for Routine Tasks, Under 10% for Non-Routine
For non-routine cognitive tasks, the figure stays below 10%, while for routine cognitive tasks it exceeds 25%. This gap emerges when the share of conversations intended for full automation is broken down by task nature. Simply dividing 25 by 10 yields a ratio exceeding 2.5—a figure that concretely underscores the scale of the asymmetry Google points to.
This figure aligns with the top-ranked occupations by saturation mentioned earlier. Software quality assurance analysts and document management specialists are occupations whose work consists largely of procedural, routine cognitive tasks. The finding that automation intent skews more than 2.5 times higher toward routine cognitive tasks compared to non-routine ones supports the idea that demand for full automation is especially strong in precisely these top-ranked occupations.
The very distinction between routine and non-routine cognitive tasks follows the task classification framework that peer review contributor David Autor has established in economics. Routine cognitive tasks refer to work with clear, rule-based procedures, while non-routine cognitive tasks refer to work involving judgment, negotiation, or creative problem-solving that resists easy proceduralization. Viewed through this lens, the finding that automation intent skews toward the routine side reads as a natural consequence of AI first displacing "jobs that are easy to reduce to rules."
For readers whose occupations center on non-routine cognitive tasks, the sub-10% figure is genuinely reassuring. But for occupations where routine cognitive tasks dominate, the over-25% figure hits closer to lived experience. Even under the identical banner of "using AI," the implications are not symmetric, and a summary like "automation intent is under 10% overall" fails to capture the underlying reality. What the paper measures is, after all, only the intent toward automation as expressed in conversation—not the actual outcome of job automation. It counts the frequency with which users engaged in dialogue seeking to have AI "handle this entire task," and a gap still remains between intent and actual displacement.
How 14.65 Million Conversations Are Mapped to O*NET Occupational Categories
The task of converting over 14.65 million conversations into O*NET occupation and task categories is handled by an automated pipeline. Conversation clusters are mapped to the Bureau of Labor Statistics' SOC occupational classifications and O*NET's detailed task data, and classification accuracy is verified by measuring agreement between synthetically generated ground-truth data and labels assigned by human evaluators. In simplified terms, this is close to applying the content-classification techniques used to automatically sort large volumes of text into categories, adapted to the thousands of detailed tasks defined by O*NET. For example, a conversation in which an HR staffer asks about vacation policy and one in which an engineer asks about a code bug are both chat-format exchanges, yet they get assigned to entirely different occupation and task labels.
Within this classification, "cognitive tasks" account for 86% of work-related conversations. Compared to the roughly 50% share of cognitive tasks in O*NET's overall task composition, this proportion jumps significantly in conversational data. Since the conversational format itself is inherently better suited to cognitive tasks than to physical tasks requiring bodily action, it's a natural result that cognitive tasks make up a higher share of work conversations than they do of the overall task composition. Separately, another finding shows that over 86% of all conversations (combining work and non-work conversations) occur outside of business hours. The former concerns the composition of conversation content; the latter concerns the timing of when conversations occur—these are entirely different metrics pointing at different things.
The Wage Correlation and the $100 Billion Unpaid-Labor Estimate—and the Fragility of Its Assumptions
The paper reports another correlation: a 1% increase in an occupation's median wage is associated with more than a 2.5% increase in AI usage intensity. The conversation-weighted median wage across observed occupations comes to roughly $83,000. This is about $20,000 higher than the actual national employment-weighted median wage, putting the latter figure at roughly $63,000 by simple subtraction. AI usage skews more heavily toward higher-wage occupations.
Even though penetration itself spans a wide range of occupations, when measured by conversation volume, usage is relatively concentrated among higher-wage workers, and the distribution of benefits is not uniform across occupations. A further challenge is that usage data itself is thinner for lower-wage occupations, making the actual situation harder to capture. Reading the 21% median saturation rate without accounting for this wage bias risks giving an impression more optimistic than the actual distribution. The same caution applies to the 68% penetration figure—viewing these numbers through the lens of wage level changes how they appear.
Google's researchers further estimated that, assuming a hypothetical time savings of 30 minutes per week in unpaid household work, standard valuation methods would put the value of unpaid productivity gains in the U.S. at roughly $100 billion. Converted at a rate of ¥163.85 to the dollar (as of July 29, 2026), that comes to approximately ¥16.4 trillion. This is not a measured economic effect. It is merely an estimate built on the assumption of 30 minutes per week, and the resulting dollar figure would shift directly if that assumed time period changed.
Google's own acknowledged limitations should not be overlooked. This analysis does not include data from Google Workspace (3+ billion users), Google Translate (1+ billion users), AI Overviews (2.5+ billion users), Google Cloud Gemini Enterprise, or agentic coding tools. The collection window was also only two weeks, and no verification was done for seasonality or shifts in usage trends. Figures like 68%, 21%, and 25% represent a slice of Google's product ecosystem, not a picture of the entire economy. Google itself acknowledges that questions such as entry-level hiring or the domestic digital divide cannot be answered using this primary data alone.
James Pethokoukis, a senior fellow at AEI, has characterized the current situation as being in the "downward phase of the J-curve" of productivity growth. The idea is that the value of intangible assets—trained talent and restructured business processes—may not yet be fully reflected in official statistics. There remains an unfilled gap between the usage reality that primary data like conversation logs reveals and existing economic statistics such as GDP. The $100 billion unpaid-labor estimate can be read as one attempt to bridge that gap.
Why Google, Anthropic, OpenAI, and Goldman Sachs Numbers Don't Match
The debate over generative AI's impact on the labor market has continued since OpenAI's 2023 estimates. Anthropic and Goldman Sachs have each published their own data, and Google's real-usage logs now bring new primary data into this ongoing debate. Yet the figures surrounding AI and employment still vary widely depending on who is doing the measuring. Lining up what exactly each source measures reveals why. The table below summarizes the differences in measurement targets, key figures, and methodology.
| Organization | Nature of Data | Key Figures | Measurement Method |
|---|---|---|---|
| Google (ATLAS v1.0) | Real usage logs | Median task saturation rate 21%; automation intent under 10% for non-routine / over 25% for routine | Automatic mapping of 14,653,926 Gemini conversations to O*NET/SOC |
| Anthropic (Economic Index) | Usage-intent classification | 57% augmentation / 43% automation | Classifies Claude conversations as "augmentation" or "automation" (Google's footnote 13 notes direct comparison is not valid) |
| OpenAI (Eloundou et al. 2024) | Exposure prediction | ~80% of the workforce could be affected by LLMs in at least 10% of their work tasks | Predicts susceptibility to impact by analyzing job content with an LLM (not actual usage) |
| WEF (Future of Jobs Report 2025) | Employer-survey-based future forecast | 92 million jobs displaced, 170 million jobs created, net gain of 78 million by 2030; 40% of employers expect workforce reductions due to automation | Survey of employers worldwide |
| Goldman Sachs (via TechCrunch) | Actual employment statistics (as reported) | Reported average net decline of ~16,000 jobs/month, concentrated among Gen Z and recent graduates | Actual employment count statistics (primary source unconfirmed) |
Google's 21% figure reflects actual usage, counted from real conversations that took place. Anthropic's 57%/43% split, meanwhile, is the result of classifying conversational intent as either "augmentation" or "automation." The two differ both in unit of measurement and observation period, and Google itself explicitly states in footnote 13 that a direct comparison is not valid. OpenAI's 80% figure is not a count of actual usage either—it is merely a prediction of whether LLMs could touch a given task. The WEF figure is a forecast through 2030 based on an employer sentiment survey, not a representation of current reality.
Even the widely reported Goldman Sachs figure of a monthly net decline of 16,000 jobs is easy to misread in isolation. A survey conducted by Ramp and Revelio Labs across roughly 22,000 companies found that companies with higher AI adoption intensity increased hiring by 10.2%, and by 12% when limited to new graduate hiring. Data pointing in exactly opposite directions coexists within the same period, making the question "Is AI reducing employment?" too complex to answer with a single figure. Even usage-rate figures diverge depending on the framing: a 2026 survey on generative AI usage rates conducted jointly by Google and Ipsos shows figures of 62%/48%, but since Google itself was involved in the survey, it cannot be treated as independent third-party data.
Lining up figures without confirming the measurement methodology behind them gets us no closer to the real picture of AI's impact on employment. Real usage logs, usage intent, exposure predictions, employer surveys, and actual employment statistics are each mirrors reflecting a different facet. Reporting or commentary that fails to specify which facet it's examining should be read with a degree of skepticism.
How to Read the "AI Isn't Taking Jobs" Narrative: Implications for Japanese Companies
Google's primary data itself provides no evidence supporting large-scale automation or job losses among white-collar workers. But when limited to routine cognitive tasks, automation intent exceeds 2.5 times that of non-routine tasks, and the top of the saturation ranking is populated by proceduralized occupations like software quality assurance analysts and document management specialists. This 2.5x asymmetry adds a crucial caveat to the reassuring narrative that "AI hasn't taken jobs yet"—namely, that it all depends on the nature of the work involved.
The paper also notes that English-language conversations account for only about a third of global conversation volume. Usage in non-English-speaking regions, including Japan, is likely already substantial, though this report offers no detailed breakdown by country or language. Domestically, labor market analyses based on real usage logs of comparable scale remain rare, making ATLAS a valuable methodological precedent to reference. As Japanese companies assess the impact of their own AI adoption, it's worth adopting a similar mindset—looking beyond surface-level penetration rates to examine how task saturation rates diverge between routine and non-routine work.
Google itself does not treat ATLAS as a one-off report. It positions this as an ongoing, multi-year research project. If the collection period is extended beyond two weeks and data from Google Workspace and agentic coding tools—currently excluded—is incorporated, it will become possible to verify whether the current signal of leading automation intent in routine cognitive tasks actually progresses into real-world automation. How the 11% of occupations already exceeding 50% saturation move in the next survey will serve as an early test case. How future editions of ATLAS update the figures for routine work will determine just how long this reassuring narrative remains valid.