Key Takeaways
- Computer programmers, customer service representatives, and data entry keyers are the jobs most at risk from AI, with task coverage as high as 75%
- The most exposed workers are not low-wage earners- they are higher-paid, more educated, and predominantly in knowledge-based roles
- AI is not eliminating these jobs outright, it is compressing teams. Meaning fewer people are needed to do the same volume of work
Anthropic researchers’ most recent research provides the first substantial evidence to indicate which jobs are most likely to be impacted by artificial intelligence (AI). This study, which uses actual usage data from millions of conversations with Claude. It has been able to demonstrate that the jobs that will be impacted by AI. They are not just those identified in theory but have been impacted by AI in practice.
The findings from the research are subsequently highly specific, as well as numerically supported, and also in certain cases, unexpectedly.
What makes a job “at risk” from AI in the first place?
Not all AI exposure is the same.
Anthropic researchers differentiated between AI that augments a human’s job and AI that completely automates a job.
A job with an AI used to completely replace it would rank higher on the risk scale than one where AI is only used to speed up the job. The researchers call their measure of job risk “observed exposure,” which they define as “the amount of an AI task that AI is being used for, the theoretical possibilities of what an AI can do, and how much of that is automated.”
Observed exposure provides a grounded measure to previous studies, which tended to overstate risk since they measured what an AI could theoretically do without accounting for what was actually done with AI.
How did Anthropic identify the most exposed occupations?
Using legitimate Claude usage statistics, researchers compared the usage of Claude against a US government database – O*NET (Occupational Network), which lists out the tasks associated with approximately 800 different types of occupations.
They evaluated each task using three criteria:
If a language learning model has the potential to complete the task at least twice as fast as humans (in theory), the task appears to be listed in Claude usage statistics, and the task is an automated process vs a supplement to an automated process.
So far, the results indicated that fully automated tasks, when included in work-related contexts, earned the most points. The points earned indicated how much work AI is currently doing for each occupation at this time, based upon what AI could theoretically accomplish.
The top 10 jobs most at risk from AI right now
Here are the ten occupations with the highest observed AI exposure, according to Anthropic’s research:
According to the outcome of the research, the top-ranked profession that uses AI tools the most is that of computer programmer with a 75% task coverage; ongoing studies support this finding, as programming remains one of the highest-volume professional uses of AI tools.
Representatives who work in customer service account for the second-highest volume of AI use cases due, in part, to businesses using APIs to automate customer interactions via the Claude or similar systems. In addition, data entry clerks completed 67% of their tasks using an automated system. Reading and entering structured information are core data entry tasks and, therefore, are ideal candidates for automation.
Why are high-paying, educated jobs at the top of the list?
The data contradicts an assumption that most people hold regarding AI’s threat to low-skilled, low-paid occupations as being the first occupation characterized by an AI threat. Anthropic found the opposite.
The average earnings for workers in occupation groups most subject to AI exposure is 47% more than for workers in occupations least exposed to AI. They are also more likely to hold graduate degrees. The most exposed group is 16 percentage points more likely to be female and nearly twice as likely to be Asian compared to the zero-exposure group.
When you examine what AI is good at performing there is logic to this trend. AI has great ability at processing text, analyzing data; creating written word documents and conducting organized information tasks which comprise the core functions of knowledge-based job functions as oposed to physical labor job functions.
Which jobs are safe from AI exposure?
In 2023, approximately 30% of U.S. employees have no measurable exposure to artificial intelligence or any task that could be completed by AI. These job types have been identified as having no evidence of any AI application for their job functions:
- Cooks/Chefs
- Motorcycle Mechanics
- Lifeguards
- Bartenders
- Dishwashers
- Dressing Room Attendants
Because these occupations require physical presence, sensory judgment, and real-time interaction with humans, they are considered free from competition from AI systems. For example, an AI cannot identify how to fix a motorcycle, determine the safety of a swimming pool. Or gauge the mood of a tavern by observing the activity in it the way a human bartender can. It can give you solutions but it can’t physically fix it.
What do the top 10 jobs have in common?
There are three common traits between the highest-exposed professions: all three have an output of text (or structured) data. They are predominantly performed using a screen and have already had all routines digitised. And the bulk of the activities they require are repetitive, rule-based tasks.
Artificial intelligence is at its best when there is both a pattern to follow and repetition to perform that job. All jobs listed above involve activities with both of these criteria. If your job includes these traits, you are probably more exposed to AI than the average exposure that is suggested by the data.
Is automation the only risk, or is something subtler happening?
In the near future, most of these jobs will not become fully automated. However, they will be subject to task erosion.
The first drafts, routine queries, or initial analyses handled via AI do not remove the role being done. They shrink the existing role. One worker can do what it previously took three suitable workers to do. Or a team of ten employees has been reduced down to only four staff. The job title remains in the organization; however, the employee who does it has been removed.
This lack of empirical evidence regarding drastic reductions in employment is known as a quiet compression. In addition to having lower rates of unemployment, there’s an indirect measure of a quiet compression. It is the low rates of hiring, especially for new employees. The data currently indicates that there are fewer entry-level jobs available for those with less than two years’ working experience.
What should workers in at-risk jobs do with this information?
Does the information currently available point toward a current state of crisis based on employment? No, there are changes happening in specific occupations due to the introduction of AI. But those occupations that will have an affect from this will not cease to exist overnight. However; the type of work within the affected occupations will change because of AI.
For workers in high risk jobs due to the use of AI, the main message to take away is this: AI works on the repetitive, predictable tasks. There are many tasks that require human intuition, relationship with the client/customer, and the ability to make tough decisions. All of which AI cannot accomplish. Focusing on these types of tasks is where the value of human employee remains strong and viable.
Wrapping it up
The jobs most at risk from AI are not waiting for some future version of the technology. The exposure is happening now, measured in real usage data, not predictions. What makes this moment different from previous automation waves is the speed at which knowledge work is being affected. Physical jobs took decades to automate. Text-based, screen-based, rule-driven work is compressing in years. Workers in these roles do not need to panic, but they do need to pay attention. The question is no longer whether AI will touch your job. For many knowledge workers, it already has.
