How much of your job will AI take over?

Fun fact: Economists spend a lot of time writing code. They use code to organize datasets with millions of observations and make calculations. They use code to solve complex equations describing how actors in the economy behave in economic models.

Now, thanks to generative artificial intelligence, they have help.

“I spend a lot less time writing code, and a lot more time checking code. I spend less time doing algebraic derivations, especially if they are easy to verify,” said former Institute visiting scholar Lukas Mann, a professor of economics at the W. P. Carey School of Business at Arizona State University.

Economists have long sought to understand how automation of work processes affects workers. Automation isn’t new, after all. Plows transformed agricultural production; the steam engine transformed manufacturing. Today, AI is transforming a host of analytical tasks, including coding.

How will this new wave of automation affect the tasks workers do, the jobs they choose, and the wages they earn? A new Institute working paper by Mann and current Institute visiting scholar Lukas Freund, professor of economics at Boston College, provides a framework to analyze the labor market’s response to these changes.

A job is a bundle of tasks, a worker has a bundle of skills

“The current discourse around AI mostly centers on the notion of jobs disappearing,” Freund said. “But that seems to capture only part of the picture.” To study how the effects of automation ripple through the economy, Freund and Mann begin with two simple observations: Most jobs involve multiple tasks, and most workers have multiple skills.

“I teach first- and second-year Ph.D. students, and they’re sometimes frustrated with how much math they have to do,” Freund said. “If the more tedious elements of math become less important, I think that would then tend to attract more people who are very good at coming up with ideas. That skill becomes more valuable if the implementation of ideas becomes easier to do.”

The job of medical radiologist is another example of an occupation where the tasks are shifting due to AI. Nine years ago, a prominent physicist predicted the job would be entirely replaced by AI. Only time will tell, but so far, radiologists have harnessed AI to help sharpen images, identify medical abnormalities, and predict disease. Radiologists continue to talk to patients, take medical histories, write reports, and analyze medical records.

From a worker’s perspective, this change to the importance of different tasks in their job changes the importance of different skills. “The idea that people are differently skilled at different types of things is very relevant in the context of AI in terms of assessing who wins and who loses,” Mann said.

From tasks to skills to wages

Freund and Mann use an economic model to project how wages in different occupations and wages of different workers will change after automation due to large language models (LLMs), such as ChatGPT. To do this, the economists need to know what tasks workers perform, which jobs involve those tasks, and what skills are associated with those tasks.

Freund and Mann use LLMs to organize 20,000 occupation-specific tasks listed in the U.S. Department of Labor’s O*NET database into 38 “task clusters” that have similar skill requirements. Jobs differ in which tasks are required and how much time is spent on each task. Using research from other economists, Freund and Mann estimate that “processing and analyzing records” is the task most likely to be automated by LLMs. This task is a significant part of the responsibilities for occupations such as financial analysts and information and record analysts.

Next, Freund and Mann estimate the effect on wages if the task “process and analyze records” were completely automated. Their analysis finds that occupations that involve a lot of “processing and analyzing records” experience wage gains, on average. This might seem surprising: LLMs are taking over crucial tasks in these occupations and yet average wages are going up. Why?

Well, because automation frees up time for workers to spend on the other tasks that their job requires. As information-processing tasks are automated, customer-facing tasks, such as coordination, communication, and negotiation, rise in significance. Workers who are particularly skilled at these other tasks will see their productivity go up (and their wages rise with their productivity). Workers who are particularly skilled at information processing and not customer-facing tasks will switch to different jobs. And workers in other jobs who are skilled at coordination and communication but not analyzing records will move in.

As this explanation suggests, automation results in a fair amount of labor market churn. As the mix of tasks at different jobs changes, workers will move to jobs that make the best use of their skill set.

What this churn means for individuals’ wages depends on that skill set. Workers who are skilled at processing and analyzing records but not the customer-facing tasks that gain in relevance are likely to switch jobs and see their wages fall. In contrast, their former colleagues who were less skilled at processing and analyzing records are more likely to stay in their job and experience wage increases as their work content shifts. The biggest winners are workers who switch into exposed occupations because they are good at customer service and coordination, but their lack of skill at information processing had previously kept them out of the occupation.

The age of AI automation

The role of LLMs in the economy continues to evolve rapidly. As of late 2024, 40 percent of the U.S. population aged 18 to 64 reported using generative AI (which includes LLMs), and 23 percent of employed respondents used it for work, according to survey results analyzed by Alexander Bick, Adam Blandin, and David Deming. These numbers are likely to increase.

Freund and Mann hope the framework they built can be used to analyze automation events past, present, and future. “In principle, we can connect our model to many other technologies,” Freund said. “As long as we have a mapping between a technology and the set of tasks it automates, we can feed that technology into our model.”

One way in which the wave of LLM automation differs from past automation events is that the tasks that are being automated now are associated with skills that vary a lot from person to person. The skill to analyze records is more dispersed in the population than skill with routine physical tasks, for instance, a conclusion Freund and Mann draw from an analysis of worker wages. This means that LLM automation could affect the labor market differently than previous waves. In particular, Mann said, “People may move around occupations a lot more because the returns to moving are generally higher” when the automated skill varies so much person to person.

One reason information-processing skills vary so much across people is they generally take years of education, training, or on-the-job experience to do well. This is another much-commented-on feature of this wave of LLM automation: It is affecting workers who have invested a lot—of time, of money—to gain their expertise.

What this means for worker welfare is yet to be seen. On the one hand, specialized skills may not easily transfer to a different job. On the other hand, Freund said, “If we think that people select into these jobs partly because they involve a lot of learning and they’re good at learning, maybe for them it’s not so hard to transition.”

Ultimately, where automation leads may depend on how we use the time we get back. It’s a decision Mann and Freund face themselves. “I can basically feel the paper playing out in my own life in real time,” Mann said.