A Microsoft Research study offers a grounded look at where generative AI is already showing up in everyday work. Instead of ranking jobs by speculation, the researchers examined how people actually used Bing Copilot and compared those conversations with standard job activities.
The result is a picture of uneven impact. Generative AI appears most useful in knowledge work, communication, writing, advising, and sales. It has far less reach in jobs that depend on physical action, real-world presence, or hands-on repair.
How the study measured AI at work
The report, called Working with AI, analyzed 200,000 anonymized Bing Copilot conversations. The researchers mapped what users were trying to do, and what the AI did in response, to the O*NET database, which classifies U.S. jobs by their core activities.
That method matters because it separates a person’s intent from the AI’s contribution. The study draws a line between the user goal and the AI action. A person may be trying to gather information, while the AI’s role is to provide information.
This distinction helps explain why generative AI often changes the shape of a task without simply replacing the person doing it. In the source example, gathering information can be part of work done by journalists or scientists, while providing information resembles work associated with librarians or customer service representatives.
In 40 percent of cases, the user’s goal and the AI’s activity involved different sets of tasks. That finding supports the study’s view of AI as a coach or advisor in many work situations, rather than only as a direct substitute for a worker.
The tasks where generative AI fits best
The strongest pattern is clear: generative AI is most effective when the work involves language, information, and communication. The most common AI-supported tasks in the study included collecting information, writing and editing, and communicating ideas.
Those are also the areas where the system performed best by the study’s measures of user satisfaction and completion rates. For many workers, that means AI is most relevant in the parts of a job that involve drafting, summarizing, explaining, researching, or shaping written communication.
The study also identifies where current systems are weaker. Generative AI was less effective for data analysis or visual design. It also struggled with tasks requiring real-world interaction or physical work, including shopping or repairs.
The source notes that OpenAI’s ChatGPT Agent is designed to gradually close this gap. Even so, the study’s findings describe the impact of current usage patterns in Microsoft Copilot, not a universal forecast for every tool or workplace.
Which professions scored highest
To compare professions, the researchers created an AI Applicability Score. This score combines how often AI is used for a task, how successful it is, and how fully it can handle that task.
The highest scores went to roles centered on language, knowledge, and communication. The source names translators, historians, writers, media professionals, customer advisors, and salespeople. Technical roles such as CNC programmers and data scientists also ranked high.
These results do not mean every task in those jobs can be handed over to AI. They mean the daily activities associated with those professions overlap strongly with what language models are already being asked to do and can often complete successfully.
At the other end of the scale are jobs where the work is physical or rooted in direct action. The study found lower impact for caregivers, tradespeople, cleaners, and machine operators. In those roles, current generative AI tools have little effect because the work depends less on text-based information exchange and more on physical execution.
Why impact is not the same as replacement
The authors caution against treating AI capability as a simple prediction of automation or job loss. A tool can affect a task without eliminating the worker who performs it. In many cases, the study describes AI as an assistant that supports and enhances human work.
The source points to ATMs in banking as an example. Automation changed how people worked, but it also created new roles in the industry. The comparison is used to show that a technical capability can reshape a profession in more than one direction.
The study also found almost no correlation between a job’s AI suitability and its pay or education level. Roles requiring a bachelor’s degree were slightly more affected, but the difference was minor. That makes the findings less about status and more about task content.
The practical takeaway is that generative AI exposure depends on what a job asks people to do each day. Work built around written knowledge, communication, advising, and information handling is more exposed. Work built around physical presence and hands-on activity is less exposed, at least in the usage analyzed here.
The limits of the findings
The study has important boundaries. Its findings are based entirely on Microsoft Copilot usage in the U.S. They may not apply in the same way to other platforms or countries.
The analysis also excludes informal work and tasks outside standard job categories, including household labor. That means the study is best read as a structured view of formal job tasks reflected in Copilot conversations, not as a complete map of all work affected by generative AI.
Still, the research adds useful clarity to a debate often dominated by broad claims. It shows that the near-term effect of generative AI is not evenly distributed across the labor market. It is strongest where work is already expressed through language, information, and communication.