What Do People Use AI For at Work? A Task-by-Task Look at the 2026 Data

The Short Answer

People do not use AI to do their jobs. They use it for a handful of repeatable tasks inside their jobs. According to the Anthropic Economic Index, published February 2025, the single largest cluster of work usage is computer and mathematical tasks like writing and fixing code, at 37.2 percent of Claude.ai usage. Writing, teaching, and administrative tasks follow well behind.

The pattern matters more than any one figure. Across occupations, most people reach for AI on a narrow set of tasks rather than across their entire role. By the January 2026 update, the ten most common tasks alone accounted for 24 percent of all sampled conversations. Work usage is concentrated, task-shaped, and far more specific than headlines about AI replacing jobs suggest.

What Do People Use AI For at Work, According to the Data?

The most reliable answer comes from task-level research rather than surveys about job titles. The Anthropic Economic Index maps anonymized conversations onto the US Department of Labor's O*NET database, which breaks every occupation into its component tasks. That method lets researchers see which tasks show up, not just which workers say they use AI.

In the original February 2025 analysis, usage clustered heavily in a few occupational task groups. Computer and mathematical tasks led at 37.2 percent. Arts, design, and media tasks came next, followed by education, office and administrative support, science, and business and financial operations. Read the table as a map of tasks, not people. The winners are all text-heavy, screen-based tasks that are easy to describe in words. That is the real common thread behind what people use AI for at work.

Which Specific Work Tasks Top the List?

Category shares are useful, but the task names are where it gets concrete. In the January 2026 report, the single most common individual task was modifying software to correct errors, plain debugging, at roughly 6 percent of usage on its own. Coursework help and educational instruction also ranked high, along with writing and copyediting tasks.

Notice what these have in common. They are bounded tasks with a clear input and a checkable output. You can hand over a broken function, a rough draft, or a confusing concept and get something back you can immediately judge. Tasks that depend on physical presence, live judgment calls, or accountability you cannot delegate barely register in the data. That is not a limitation people are working around. It is a genuine signal about where current AI is actually useful.

Is AI Doing the Whole Task or Just Helping With It?

This is where the fearmongering usually falls apart. The Anthropic Economic Index splits usage into two modes. Automation, where the person delegates a task and takes the result, and augmentation, where the person and the model go back and forth to build something together. In the February 2025 data, augmentation led at 57 percent versus 43 percent for automation.

The mix has moved over time. By November 2025, reported in the January 2026 update, augmentation still led at 52 percent against 45 percent automation, described as a broad shift back toward collaborative use. In plain terms, more than half the time people are working with AI on a task rather than handing it off completely. The dominant pattern is a person staying in the loop, steering, correcting, and deciding.

How Much of Any Real Job Actually Touches AI?

Even for the tasks that do show heavy AI use, whole jobs rarely light up. In the February 2025 report, only about 36 percent of occupations showed AI use across at least a quarter of their tasks, and just about 4 percent showed use across 75 percent or more of their tasks. By the January 2026 update, the share of jobs where AI was used for at least a quarter of tasks had risen to 49 percent, so the trend is real, but even then it describes a quarter of tasks, not entire roles.

The academic research agrees on the shape. Eloundou, Manning, Mishkin, and Rock, publishing in Science in 2024, estimated that around 80 percent of US workers could have at least 10 percent of their tasks affected by large language models, while about 19 percent could see at least half of their tasks affected. Read carefully, that is a task-exposure finding, not a job-loss forecast. Most workers have some exposed tasks. Very few have most of their work exposed. You can see where your own work lands with the free task-level check on our homepage.

What Does This Mean for Your Own Task List?

Here is the empowering part. Because AI adoption happens task by task, your personal risk and your personal opportunity both depend on your specific mix of tasks, not your job title. Two people with the same title can have very different exposure depending on what they actually do all day. Averages cannot tell you which group you are in.

The national data tells you the pattern. Your own task list tells you the plan. AI usage at work is concentrated in a narrow band of text-heavy, checkable tasks. It leans collaborative more often than not. And it touches slices of jobs rather than whole ones. Knowing your own slices is how you turn a scary headline into a to-do list.