Who’s most at risk? AI shows highest impact on white-collar and sales jobs

With the rapid proliferation of generative artificial intelligence in homes and workplaces, concerns surrounding its impact on jobs and productivity are growing increasingly urgent. Now, a team of researchers from Microsoft Research has delivered one of the most comprehensive empirical analyses to date, using actual user behavior to map how AI is influencing work in real time.

In a study titled “Working with AI: Measuring the Occupational Implications of Generative AI”, submitted on arXiv, the researchers examined over 200,000 anonymized user–AI interactions from Microsoft Bing Copilot (now known as Microsoft Copilot). The dataset, spanning nine months of U.S.-based usage, offers a rare lens into how people actually use generative AI on the job and what it reveals about the occupations and tasks most susceptible to AI transformation.

What work activities are being transformed by generative AI?

The study identifies the types of work activities users are seeking AI assistance for and how well the AI performs them. Using a classifier trained to recognize 24 distinct work activities aligned with the U.S. Department of Labor’s O*NET taxonomy, the team analyzed both user prompts and AI outputs. The findings reveal that the most commonly requested AI support involves information gathering and writing tasks. These requests mirror activities traditionally associated with cognitive, white-collar work.

In terms of what the AI is most often performing, the top categories include providing information, writing, teaching, and advising. These functions correspond to activities that require knowledge synthesis, communication, and decision support, roles historically filled by human professionals in fields such as education, law, customer service, and marketing.

To evaluate how well AI performs across these activities, the researchers measured two core dimensions: success (based on AI’s relevance and usefulness in its responses) and scope of impact (the breadth of occupations each activity affects). Combining these dimensions, the team computed an AI applicability score for each work activity and, subsequently, for each occupation that depends on those activities.

The results suggest that tasks involving creativity, open-ended thinking, and structured knowledge are particularly well-suited for AI augmentation. This is especially true when the tasks don’t require real-world physical manipulation or nuanced human interaction. However, the study also found that many users attempt to use AI for tasks it performs poorly, such as complex decision-making, long-form reasoning, or managing context over time, highlighting gaps between AI capabilities and user expectations.

Which occupations are most affected by current AI capabilities?

By linking work activities to more than 900 occupations in the U.S. workforce, the researchers determined which jobs are most exposed to generative AI. Occupations in computer and mathematical fields, office and administrative support, and sales and related professions ranked highest in AI applicability. This is largely due to the fact that these roles rely heavily on tasks involving information processing, document drafting, and customer interaction, areas where AI currently excels.

For example, administrative assistants often perform scheduling, email composition, and data entry, tasks that generative AI can rapidly automate or support. Similarly, roles in sales frequently involve generating persuasive content, answering product queries, and tailoring communication, tasks well within the capabilities of modern language models.

Interestingly, the study reveals that even within these high-exposure occupations, the nature of the tasks determines AI’s actual impact. Routine and well-defined tasks tend to be more automatable, while tasks involving high levels of discretion or emotional intelligence remain largely resistant to AI disruption.

The study also compared their findings with earlier theoretical predictions about AI’s impact. While previous forecasts often used task-based exposure models without real-world usage data, this study grounds its estimates in observed behavior, revealing both overlaps and divergences. Some occupations thought to be safe in theory, such as certain education or consulting roles, show considerable AI use in practice, particularly in tasks related to teaching or advising.

What does this mean for the future of jobs, skills, and inequality?

Apart from identifying AI's applications, the researchers examined how AI applicability correlates with wages, education levels, and employment patterns. They found that occupations with higher education and wage requirements tend to have higher AI applicability scores. This raises important questions about inequality and skill-biased technological change.

Rather than replacing only low-wage or routine jobs, generative AI appears to be complementing or transforming a wide range of high-skill, high-income roles, potentially amplifying productivity for those already well-positioned in the labor market. This contradicts earlier narratives that predicted AI would primarily threaten blue-collar or low-wage employment.

The study also sheds light on how users interact with AI in the real world, including the mismatch between what people expect from AI and what it can actually deliver. In many cases, users asked AI to perform tasks that required judgment or contextual understanding beyond its current capabilities. This gap underscores the need for better user education, system transparency, and the development of AI systems that are more aligned with human needs and expectations.

Looking ahead, the researchers call for continued study using real-world usage data, rather than speculative forecasting, to understand AI's evolving role in the economy.