HOW IT WORKS

How AI is used at work: assistance is more common than full automation

Google's new ATLAS report is based on 15 million interactions with the company's AI services. In work-related conversations, AI is mainly used as support in parts of a task, but the basis does not describe the entire labor market.

Publicerad 24 July 2026, 20.10

AI generated illustration of a teacher, technician, office worker and small business owner using AI support for selected tasks while other tasks stay with the human.

Google has published the first version of ATLAS, a study of how people use the company's AI services. The dataset consists of 15 million aggregated and de-identified interactions in the Gemini app, AI Mode and Gemini API. The study covers more than 150 countries, 140 languages, 800 professions and 4,000 tasks.

In work-related conversations, AI is seen in many kinds of professions, but the use is selective. Google states that AI appears in about 21 percent of the tasks in a typical profession in the material. This does not mean that 21 percent of the job has disappeared. The measure describes which data occurs in the use of Google's tools.

According to Google, most work-related interactions are about ideation, strategy, information seeking and learning. Less than 10 percent were classified as fully automating a task. It's the difference between getting help in one step and handing over the entire workflow.

ATLAS also shows examples from technical and manual occupations. Automotive technicians and industrial mechanics use AI for troubleshooting, interpreting test results and learning in the moment, among other things. When these groups use Google's tools, multimodal AI, which can work with images or video, is more common than in other work-related conversations.

The report is an early image from Google's own products. It does not cover all AI use, and Google itself mentions that, among other things, Workspace, Translate, AI Overviews and several business services are outside. The results therefore do not show how common AI is in Sweden and cannot be used to predict Swedish jobs.

The International Labor Organization ILO comes from another direction. The ILO measures which job tasks can be exposed to generative AI, not actual use in Google's services. The organization's global assessment is that tasks in many occupations can change, but that the need for human participation makes transformation more likely than full replacement. Both documents point to the work assignment being a better level of analysis than the entire professional title.

Därför spelar det roll

When a business asks if AI will take a job, the answer is easily too crude. A profession consists of many tasks with different demands on judgement, responsibility and contact with people. A better mapping shows where AI can provide support, where a human must check the result and which parts should not be automated.

Det här kan du göra

  1. List five recurring tasks in a role rather than assessing the entire occupation at once.
  2. Highlight where AI can provide a draft, find information or suggest next steps.
  3. Maintain human accountability for decisions, sensitive tasks and actions with real consequences.