Google and AI: 800 Jobs Analyzed, Few Replacements

Google and AI: 800 Jobs Analyzed, Few Replacements

Google has analyzed nearly 15 million conversations with Gemini to determine the actual impact of artificial intelligence on work. The results are reassuring, provided one does not delve into the appendices.

For the past three years, a certain type of study has emerged in the tech sector: those predicting whether your job will survive the age of AI. Institutions like Goldman Sachs have mentioned 300 million jobs at risk, while McKinsey has provided percentages of automatable hours, and the International Monetary Fund has noted that 40% of global jobs are affected. These publications relied on a method of estimating what AI could accomplish if left to its own devices. The report published by Google on July 23 takes a different approach by focusing on what people actually do, leading to significantly different results.

Analyzing Queries to Get Closer to Reality

The document is titled ATLAS, which stands for Activity, Task, Landscape and Adoption Study. It is based on 14.65 million depersonalized interactions collected between April 6 and April 19, 2026 across three platforms: the Gemini app, the AI mode of search, and the programming interface. A classifier first sorts the exchanges between professional and personal use, then associates each work-related conversation with a database of 4,000 tasks and 800 jobs derived from the American O*NET framework. This work covers 150 countries and 140 languages, making it the largest sample of actual usage ever published by an AI provider.

Measuring adoption based on queries rather than statements is akin to assessing a city's traffic by analyzing GPS traces instead of asking drivers where they think they are going. The results are more concrete and often less spectacular.

The findings show both extensive and superficial usage. AI is present in 68% of the jobs surveyed, collectively representing 88.4% of American employment. In a typical job, AI is involved in only 21% of tasks. Less than 10% of professional conversations seek to have a task executed end-to-end by the machine. Only 3% of jobs show usage covering more than three-quarters of their tasks.

The report also confirms that AI has surpassed office environments. Nearly one-third of very physical jobs show no detectable usage, but among those that do, automotive technicians and industrial mechanics use multimodal functions (photo, video) twice as much as the average. Showing a malfunction to an assistant rather than describing it is a usage that 2023 projections did not anticipate. This observation aligns with a study published by Anthropic last June on the boundary between digital jobs and those rooted in the physical world.

The Figure That Wasn't in the Press Release

The infamous "less than 10%" made its rounds in newsrooms within forty-eight hours. It deserves careful reading, as it does not apply to all professional conversations: it pertains to non-routine cognitive work, which requires judgment, design, and interpretation. For routine cognitive work, such as data entry, formatting, and standardized processing, the tables in the report show a completely different result, with the intention of complete automation exceeding a quarter of the exchanges. This second figure does not appear anywhere in the announcement.

Moreover, the study does not cover Workspace, Google Translate, generated summaries in search results, Gemini for Google Cloud, or Gemini Enterprise, due to a lack of preserved usage logs. Yet, these are the products deployed in businesses, where the automation of a process is more likely to be organized rather than improvised by an employee. One of the project economists acknowledges this: he wishes he could leverage this data.

The most obvious and rarely mentioned point remains. Google is measuring the usage of its own products, with its own economists, on a sample it has defined, and concludes that consumer AI is not destroying jobs. The report honestly presents its results as preliminary observations and announces a multi-year program, but the position of both judge and party should be kept in mind while reading.

The French Market Tells Exactly the Same Story

French data aligns with ATLAS in an almost uncomfortable way. The study conducted by Coface with the Observatory of Threatened and Emerging Jobs estimates that about 16% of tasks are exposed to automation in France. This corresponds, according to the previously detailed projections, to five million positions by 2030. The precision accompanying this figure has circulated much less: the indicator reflects task exposure, not their destruction. 136 jobs, or 15% of the total, show less than 5% of tasks exposed, logically focusing on physical activities.

Adoption in France also remains very lagging. Figures from Bpifrance at the launch of the government plan "Dare AI" indicated that 8% of very small businesses and 13% of SMEs were equipped with a turnkey solution. Among large companies, the proportion exceeded half. With 200 million euros allocated under France 2030, this plan aims to equip all large companies, 80% of SMEs and mid-sized enterprises, and half of very small businesses by 2030. The state is thus seeking to broaden an adoption that ATLAS shows to be superficial once installed: wide, but as thin as a sheet of paper.

Regarding the question that concerns everyone, that of job losses, the counters remain stubbornly flat. The Budget Lab at Yale University reports no disruption in American employment statistics thirty-three months after the launch of ChatGPT. One exception stands out: the IT sector, where tasks once assigned to juniors are now performed by machines, without any changes in experienced staff numbers. This is not a wave of layoffs, but a door closing.

The real lesson from ATLAS therefore does not concern jobs, but tasks. No one will lose their job due to a conversation with Gemini; however, the routine part of work (meeting notes, rephrasing, filling out tables, email follow-ups) is already shifting elsewhere, without any social plan or announcement. Examine your own week and take stock: if one-fifth of your tasks resembles this list, you are not ahead of the global average; you are exactly at its level.