Which Work Changes First With AI: Tasks, Not Titles

"Will AI take my job?" is the question people ask. The researchers who try to measure it mostly ask a smaller one: which tasks inside a job could a model do, or speed up? That shift from titles to tasks is more precise, and it is more useful, because you can apply it to your own week.

A job is a bundle of tasks

Occupational data in the US is already organized this way. O*NET, developed under the sponsorship of the US Department of Labor, covers 923 data-level occupations as of September 2026 and links more than 19,000 task statements to over 2,000 detailed work activities. A single occupation contains many tasks with very different characteristics: drafting a report, calming an upset customer, checking a shipment, deciding who gets a discount.

A model can be very good at some of those and useless at others. That is why "which jobs will AI replace" produces confusing answers. Almost no occupation is made entirely of tasks a model can do, and almost none is untouched.

What the "GPTs are GPTs" paper measured

A widely cited task-level study is "GPTs are GPTs: An Early Look at the Labor Market Impact Potential of Large Language Models" by Tyna Eloundou, Sam Manning, Pamela Mishkin and Daniel Rock, first posted in March 2023. The authors took O*NET tasks and rated each one for "exposure," using both human annotators and GPT-4 as a classifier.

Their definition of exposure is specific. A task counted as directly exposed if using a language model could cut the time to complete it by at least half while keeping equivalent quality. "Equivalent quality" meant that the recipient of the output would not notice or care that a model helped. A second category covered tasks where the model alone would not halve the time, but software built on top of it could.

The headline results:

• About 80% of the US workforce could have at least 10% of their work tasks affected.
• About 19% of workers may see at least half of their tasks affected.
• With direct access to a model alone, about 15% of all worker tasks could be done significantly faster at the same quality. Counting software built on top of models, that rose to between 47% and 56%.

Two caveats from the paper matter more than the numbers. The authors say they make no predictions about the development or adoption timeline. And they define exposure "without distinguishing between labor-augmenting or labor-displacing effects." An exposed task is one where a tool could save a lot of time. It is not a forecast that the person doing it will be replaced.

The gap between 15% and roughly half is the practical lesson. In the paper's estimate, most of the potential comes from software built on top of models and wired into the systems where work already happens, not from someone pasting text into a chat window.

What the ILO found globally

The International Labour Organization ran a similar task-based analysis for the whole world. Its August 2023 working paper, Generative AI and Jobs, found clerical work the most exposed, with nearly a quarter of clerical tasks highly exposed and more than half at medium exposure. It concluded the technology is more likely to augment jobs than destroy them, by automating some tasks rather than taking over whole roles. It put the share of employment potentially exposed to automation at 5.5% in high-income countries and 0.4% in low-income countries.

A refined index published on May 20, 2025 found that one in four workers worldwide is in an occupation with some generative AI exposure, while 3.3% of global employment falls in the highest exposure category. Clerical occupations still had the highest exposure, and some highly digitized professional and technical jobs had become more exposed. The authors again judged transformation of jobs to be the most likely impact.

Exposure is not the same as automation

A task can change in two directions. Automation means the tool does the task and a person reviews the result, or nobody does. Augmentation means a person still does the task, but faster or better with the tool's help.

One large field study of augmentation comes from customer support. In "Generative AI at Work," Erik Brynjolfsson, Danielle Li and Lindsey Raymond studied 5,179 support agents who were given an AI assistant. Issues resolved per hour rose 14% on average, with a 34% improvement for novice and low-skilled workers and minimal impact on experienced, highly skilled ones. The same exposed task, answering customer questions, changed a lot for new agents and little for veterans. The authors report suggestive evidence that the tool spread the best practices of more able workers to newer ones.

The studies above do not give a formula for which way a task goes. As a working rule of thumb, a task leans toward automation rather than augmentation when:

• The inputs and outputs are digital and fairly standard.
• The task repeats often in the same form.
• The output is quick and cheap to check.
• A mistake is cheap to fix.
• Nobody needs to be personally accountable for the decision.

Remove any of these and the likely pattern shifts toward augmentation, or toward little change at all. The GPTs paper adds evidence on skills: roles that rely heavily on science and critical thinking skills showed a negative correlation with exposure, while programming and writing skills were positively associated with it. Its authors conclude that "manual work is not exposed to LLMs or even LLMs with additional systems integration for the time being."

Map your own role in five steps

You do not need a research team to apply this. You need a list and an honest hour.

1. List your tasks. Use two weeks of your calendar, sent email and finished work. You can also start from the task list for your occupation in O*NET and edit it until it matches what you actually do.
2. Estimate time share. Mark each task as a large, medium or small share of your week.
3. Answer three questions per task. Is the work mostly text, numbers or code? How hard is it to check the output? What does an error cost?
4. Label the likely pattern. Automate candidate, augment candidate, or stays human for now.
5. Test one augment candidate. Pick a frequent task with cheap errors. Time it for a week without help and a week with a tool, and ask the person who receives the output to compare quality.

Here is what a first pass might look like for a hypothetical operations coordinator:

The value is in what the last column implies. If the automate candidates add up to most of the week, the role changes a lot even though the title does not. If they are a sliver, the change is real but modest.

What changes first, and what to do about it

Across these studies, exposure concentrates in digital, text-heavy work such as clerical tasks, writing and programming, and the ILO judges transformation of jobs, not their disappearance, the most likely impact. If routine tasks shrink, the tasks left over become a larger share of the work: handling exceptions, making judgment calls, managing relationships, and reviewing machine output for errors.

That tells you where to invest. Get good at the review step for the tasks that are changing, because in both patterns someone still has to catch what the tool gets wrong. Put more of your time into the tasks in your "stays human" column. Then redo the map every six months, since tool capabilities keep moving and your labels will move with them.