20% of U.S. workers offload their work to AI, and what is being replaced are tasks rather than positions.

How much of your work do you need to outsource, such as delegating it to colleagues or interns?

On August 6, Epoch AI, an AI research institute, joined forces with Ipsos, a polling firm, to release a workplace survey report:

One in five U.S. employed people said AI has now largely or entirely taken over at least one task that they used to delegate to colleagues or contractors.

The distribution of tasks that used to be delegated to colleagues or contractors but are now handled by AI: Data analysis ranks first at 7.1%, and 19.9% of respondents reported at least one such task.

This figure is easily misinterpreted as: AI has replaced one-fifth of jobs in the United States.

But what this report really means is that AI first replaces tasks, not positions.

The position still exists, except that those tasks that used to be outsourced are now handed over to AI.

In other words, work tasks are being redistributed between humans and AI.

This survey was conducted from July 10 to 19. After screening by employment status, 1106 valid respondents were obtained. The original question of the questionnaire is as follows:

What tasks that you or your team used to delegate to contractors or colleagues are now mainly or entirely handled by AI?

Epoch gave a very restrained qualitative description in the report: Although task-level substitution has been observed, it does not necessarily mean the complete substitution of human workers.

The 20% figure represents the redistribution of tasks, not the disappearance of jobs.

AI has touched on all ten types of tasks, but none of them are fully taken over by AI

The survey selected 10 common knowledge work activities from the O*NET occupational database of the U.S. Department of Labor, sorted by the proportion of national employment.

Among people who take a certain task as their daily work, the lowest AI usage rate is 25% for maintaining business records; the highest is 57% for designing computer or information systems and software applications; data analysis accounts for 46%, and reading work documents accounts for 39%.

All ten types of tasks have AI users.

But another set of figures shows that even in software design, where AI penetration is the deepest, the proportion of tasks where AI completes most or all of the work is only 10%. The remaining nine tasks all have a proportion lower than 7%.

AI usage of the ten tasks. The dark color represents "AI handles most or all of the tasks", and the light color represents "AI only participates in part of the tasks". The usage rate ranges from 25% for record maintenance to 57% for software design.

AI is widely applied, covering almost all types of knowledge work, but its penetration is shallow, and few tasks can be fully completed by AI alone.

Half of the tasks save time, and one-sixth take more time

When AI only helps with part of the work, 37% of the tasks are reported to take less time. When AI completes most or all of the tasks, this proportion jumps to 53%.

The conclusion seems obvious: the more work AI does, the more time you save.

However, Epoch AI's analysis shows that this is only a correlation, not a causal relationship, and it gives three explanations:

It may be that AI taking over more work does save time; it may also be that workers want to improve efficiency in the first place, so they actively delegate more links to AI; or it may just be that these tasks happen to be the ones that AI is particularly good at.

When AI only helps with part of the work, 37% of tasks take less time; the proportion rises to 53% when AI handles most or all of the tasks. In both groups, about one-sixth of the tasks take more time than before.

The same set of data also has a less mentioned figure: about one-sixth of AI-assisted tasks now take more time than in the past.

And no matter whether AI only assists or takes over most of the work, this proportion is roughly the same.

The reason may be that the back-and-forth communication with AI itself takes time, or that after AI frees up human hands, people tend to do more detailed and more work on the same task.

In either case, the statement that "using AI will immediately boost efficiency" needs to be viewed with a grain of salt.

66% of AI outputs are accepted as they are, which does not mean 66% of them are correct

Across all tasks involving AI, 66% of the outputs are used as-is or with only minor modifications.

Breaking down this data makes it clearer: only 5.8% of outputs require no modification at all, 59.9% require minor modifications, 27.1% require major revisions, and 4.7% require extensive rework or almost complete remaking.

Distribution of modification extent for AI outputs: 5.8% no modification, 59.9% minor modification, 27.1% major revision, 4.7% extensive rework.

Epoch mentioned in the report: The amount of modification is not a direct indicator to measure the output quality. In addition, there is no stable corresponding relationship between time saving and the amount of modification.

This survey does not assess factual accuracy or statistical error rates, it only looks at whether employees are willing to use the outputs directly.

An output with few modifications may be really good, or it may just be that no one has time to check it carefully.

The first to be eliminated is not you, but outsourcing

Looking at different tasks, data analysis ranks first, with 7.1% of respondents reporting that this task has been delegated to AI. 5.7% for reading work documents, and 5.3% for maintaining business records.

All ten types of tasks are covered, only the depth of AI application varies.

These tasks have one thing in common: clear boundaries, deliverable finished products, can be accepted after completion, and no need to sit in the office to communicate repeatedly with others to align progress.

In the past, these were exactly the types of tasks most suitable for outsourcing.

A marketing manager who originally needed to find an external data analyst to run a quarterly report can now get the result by opening a dialog box himself.

The first to be affected in this chain are contractors and piece-rate jobs on task outsourcing platforms, which are naturally individual task packages that can be packaged and sent out separately.

What AI is best at right now is precisely receiving a clearly defined task package.

NBC interviewed two researchers in a report released on the same day.

The judgment of Amreeta Das, who led the study, is as follows:

People are exploring how to distribute work between themselves and AI; so far, this does not necessarily mean that the entire job is automated, but AI is entering the workplace through the tasks in these jobs.

Aya Ibrahim, a senior researcher at the AI Now Institute, raised a more pointed question.

She believes that the role of contractors will not disappear entirely because of this, but contractors who are paid by tasks will face more direct pressure.

What she really wants people to discuss is how unstable "a job that is completely pieced together by a series of tasks" itself is.

This redistribution does not happen to everyone at the same pace

After work is split into tasks, there is a distribution of which tasks are split first.

In another survey conducted by Ipsos for Groundwork Collaborative from June 11 to 16, 1533 employed people or marginal workers were covered.

The conclusion is that AI adoption is extremely uneven: 30% of workers use AI at least once a week, but these people are highly concentrated in high-income, college-educated and white-collar groups.

Among people with an annual income of more than $100,000, people with a bachelor's degree or above, and white-collar workers, nearly half use AI at least several times a month.

For people with an annual income of less than $50,000, people with a high school education or below, and blue-collar workers, this proportion is only about a quarter or even lower.

Expectations are also divergent.

In the same survey, two-thirds of workers believe that AI will worsen their workplace experience, on the grounds that it will cut jobs and increase pressure, rather than taking over repetitive tasks.

This judgment is highly consistent across all dimensions of race, gender, education and income.

But when it comes to their own situation, the higher the education level of people, the more optimistic they are. Nearly half of college graduates believe that AI will help their work, while only one-fifth of people with a high school education or below think so.

Gallup uses a different statistical caliber, but the conclusion is consistent in direction.

As of May 2026, 52% of U.S. employees use AI at least several times a year, 30% use it at least several times a week, and 15% use it every day.

65% of people say AI has a positive impact on their productivity, but only 14% strongly agree that AI has changed the way work is done in their organizations.

AI has entered most people's work processes, but it has only changed individual tasks, not the organizational mode of work itself.

20% is a progress bar, not the end point

These 10 tasks are selected from O*NET by employment proportion, medical activities are excluded, and financial transactions, internal coordination, training, and procurement are also not included.

It describes a part of U.S. knowledge work, not all professional activities.

Therefore, the 20% figure only presents a section of these 10 tasks: who should do a job is being restructured among employees, colleagues, contractors and AI.

In this process, what is more worthy of attention is not how many people are replaced, but how much work is being split up.

If a job is just a series of tasks that can be outsourced, it has been gradually split and outsourced even before the emergence of AI. Today's AI only accelerates the speed of this splitting process.

So instead of asking "Will AI replace me?", it is better to ask a more specific question:

How many tasks in my job can be split into separate packages that can be independently accepted, inspected and assigned to others?

The more tasks that can be split out, the greater the exposure risk.

On the contrary, those parts of the work that require on-site presence, responsibility-taking, and repeated communication and alignment with others are difficult to be packaged and outsourced in the short term.

AI can draft the contract, but the signature still needs to be done by a human.

This article is from the WeChat official account "AI Era", author: ASI Revelation, published with authorization from 36Kr.