Table of Contents
- AI Is Changing Work Faster Than Employment
- The Entry-Level Squeeze
- Watch Hiring, Not Just Layoffs
- Automation and Augmentation Are Different
- Productivity Does Not Automatically Mean Fewer Jobs
- The Hiring Market Has an AI Arms Race
- The Signals Employers May Value More
- Why the AI Jobs Numbers Are Hard to Read
- Automating Work and Redesigning Work Are Different Decisions
Subscribe to our weekly newsletter
AI Is Changing Work Faster Than Employment
The simplest way I know to describe the evidence so far is that AI is changing work faster than it is changing employment. Broad measures still do not show mass technological unemployment. At the same time, AI is already widespread: observed usage appears in occupations representing 88.4% of U.S. employment, and about 55% of workers say they use AI for at least one activity. But reach is not the same as depth. In the median occupation with observed usage, AI touches about 21% of tasks, and only around 3% of occupations show usage across more than three-quarters of their tasks. That helps explain how AI can feel ubiquitous at work without yet producing an economy-wide collapse in jobs.
What I would monitor is what is happening underneath those averages. Many workers who use AI are already saving meaningful amounts of time, but that extra capacity can become more output, faster turnaround, or better work rather than fewer employees. At the same time, the employment weakness that does show up is concentrated. Call centers are well below historical employment trends in several countries, unemployment claims look more concerning among some college-educated workers in highly exposed jobs in California, and young workers in South Korea have lost ground in AI-exposed industries. I don’t think we can neatly attribute all of this to AI because higher interest rates, the post-pandemic reset, remote work, and the technology slowdown are tangled up in the same period. But the unevenness itself looks important. The early labor-market effect of AI may be less about jobs disappearing everywhere and more about the economics of certain kinds of work changing first.
The Entry-Level Squeeze
The clearest place to see that unevenness may be at the beginning of the career ladder. Among U.S. workers ages 22 to 25, employment in the most AI-exposed occupations is now roughly 19% below where it would be if it had kept pace with less-exposed work. The pattern is not limited to the U.S. In South Korea, most of the decline in youth employment since 2022 has occurred in highly AI-exposed industries, even as employment among workers in their 50s increased in many of the same sectors. India offers another version of the same split, with IT firms much more likely to report declining entry-level hiring than declining senior hiring. I would be careful about assigning all of this to AI. Junior hiring was already weakening before ChatGPT, and remote work, demographics, and the broader slowdown in white-collar hiring are all mixed into these numbers.
What interests me more is why the pressure might fall disproportionately on younger workers. Entry-level professional jobs tend to involve more formal, teachable work such as research, synthesis, first drafts, and routine analysis. Experienced employees bring more context, judgment, knowledge of exceptions, and an understanding of what actually matters. That makes it plausible that AI can substitute for some junior work while making an experienced employee more productive. But there is a catch for companies. Junior work is also how people become experienced. If AI removes part of that learning path, companies will have to replace it with something else. EY is already experimenting with longer residencies that move beginners more quickly toward judgment and real project experience. I suspect that redesigning apprenticeship will become as important as deciding which entry-level tasks to automate.
Watch Hiring, Not Just Layoffs
That suggests another place to look for AI’s employment effects: hiring rather than layoffs. The weakness among young workers in highly exposed occupations appears to come mainly from fewer people being hired, not a surge in people being fired. Canada shows a related pattern. The gap in job-finding rates between highly exposed and unexposed work has widened sharply while separation rates remain roughly similar. And companies adopting AI still report relatively few direct layoffs, with retraining existing employees a much more common response. A position that never gets opened will not appear in a layoff announcement or an unemployment claim, which makes this kind of adjustment easy to miss.
I also think this changes what we mean when we talk about AI replacing workers. The practical shift may be less about replacing one employee with one AI system and more about reorganizing a team around fewer people. Forty-five percent of surveyed hiring managers say they are pairing a senior employee with AI to handle work previously spread across several entry-level employees, and 30% say AI has reduced their need for graduate hires. Existing employees have an advantage here because they already have context, relationships, and knowledge that a company can build on through retraining. A prospective hire faces a different problem: the job they would have filled may simply never be created. The unit I would monitor, in other words, is not just the individual job. It is how many people a company decides it needs to get the work done.
Automation and Augmentation Are Different
That leads to a distinction I think is more useful than asking whether a job is “exposed” to AI. What matters is whether AI replaces part of the worker’s contribution or makes that worker more capable. The early evidence suggests those two uses can lead to very different outcomes: employment looks weaker where AI is used more as a substitute, while work that is more complementary to AI has held up better. For now, actual usage also looks much more like assistance than replacement. Less than 10% of observed conversations involving nonroutine cognitive work aim to automate the core task end to end. Drafting, reviewing, finding information, troubleshooting, and generating ideas are much more common. Of course, that line can move. Something that requires a person today may require much less supervision as the systems improve.
There is another reason I would be cautious about jumping from “AI can do this task” to “AI can replace this job.” Jobs include context, judgment, customer relationships, knowledge of exceptions, verification, and responsibility for the result. Making the first draft nearly free does not eliminate those things, and in some workflows it may make them more important. There is also a practical lesson for teams deploying AI: the model that is best at completing a task on its own is not necessarily the model that is best at helping someone complete it. In one set of experiments, the top autonomous model was beaten by another model as an assistant on five of seven tasks, and sometimes adding AI made performance worse. I would evaluate the worker, the model, and the workflow as a system rather than assuming that a more capable model automatically means either more automation or more productivity.
Productivity Does Not Automatically Mean Fewer Jobs
There is also a reason not to assume that every productivity gain eventually turns into fewer jobs. Among more than 21,000 U.S. companies, the heaviest AI adopters increased total headcount by about 10% over two years, while entry-level employment grew 12%. I would not read that as proof that AI creates jobs. Those companies were already larger, faster growing, and more likely to be venture backed. But it is a useful counterexample to the simple idea that making workers more productive necessarily means needing fewer of them. If AI lowers the cost of serving a customer or building a product, a company can use the savings to shrink its workforce. It can also lower prices, find more customers, launch new products, and expand.
We can already see some of that new demand taking shape. AI-skilled technology employment in the U.S. and Canada grew about 45% in a year, and AI-related roles now account for roughly 31% of U.S. technology listings, up from 11% in 2022. The effects also extend beyond software. Building the infrastructure behind AI creates work for technicians, construction crews, plumbers, HVAC specialists, and others. None of this tells us whether AI will create more jobs than it eliminates. Even economists who largely agree that AI will raise productivity remain divided on that question. I think the reason is straightforward: technology determines what becomes cheaper to do, but companies and customers determine what happens next. And whatever the net number turns out to be, the gains are unlikely to be distributed evenly across firms, occupations, or regions.
The Hiring Market Has an AI Arms Race
AI is not just changing the demand for workers. It is also changing the machinery that matches people with jobs. Applying has become dramatically cheaper because candidates can generate tailored resumes and applications at scale. Employers have responded with automated filters, chatbots, AI interviews, and more assessments. Recruiters now process roughly 291 applications for every hire, up from about 100 in early 2021, yet filling a position takes nearly 25% longer than it did before the pandemic. That is a strange outcome. Both sides have more automation, but the result can be more noise and a less efficient market.
The scarce resource in that environment may turn out to be trust. Almost two-thirds of applicants say they have encountered an AI interview, 38% say they have abandoned an application because of one, and only 21% of recruiters are very confident their automated systems are not screening out qualified candidates. I think this creates another problem for people entering the workforce. Someone with ten years of experience can point to a track record, references, and a network. A recent graduate has to lean much more heavily on resumes, credentials, portfolios, and assessments, precisely the signals AI makes easier to polish. I would not overstate the problem, since recent graduate placement rates have actually improved. But AI can make the job market harder to navigate even without reducing the number of jobs, simply by making it harder for employers and candidates to figure out who is actually a good match.
The Signals Employers May Value More
So what does any of this mean for an individual worker? The most obvious implication is that knowing how to work with AI is becoming part of the baseline in many professional jobs. Nearly three-quarters of surveyed employers now describe AI skills as an advantage or requirement, and workers seem to have noticed. Mentions of AI added retroactively to old professional experience jumped more than sixfold after the release of ChatGPT. There is even survey evidence that workers who use AI more frequently have experienced fewer layoffs. I would be very careful with that last finding. It does not show that using AI protects your job, since frequent users may already have different skills, roles, or employers. But I think the broader signal is hard to miss: being able to use these tools effectively is increasingly becoming an expected part of the job rather than a specialty.
At the same time, AI may be making actual experience more valuable, not less. Among recent graduates, 81.6% of those who worked while in school were employed, compared with 40.7% of those who did not. That is especially interesting because some of the entry-level jobs that provide this experience may themselves be under pressure. I would also put more weight on evidence of what someone can do now. Recently earned credentials still predict performance, while older accomplishments lose much of their value over time, and professional profiles have become surprisingly easy to rewrite after the fact. For employers, the practical lesson is to look harder for recent, verifiable evidence of capability. For workers, I think the combination to build toward is straightforward: learn to use AI, accumulate real operating experience, and be able to demonstrate what you can actually do.
Why the AI Jobs Numbers Are Hard to Read
After looking through all of this, I think we know more about where AI is beginning to change work than about how many jobs it will ultimately create or eliminate. That distinction matters because many of the numbers being used in this debate measure different things. An AI exposure score tells us what a model might be able to do, not whether a company has deployed it or eliminated a job. Usage data tell us what people ask AI to do, but not what happens to employment afterward. Unemployment claims capture jobs that disappeared, but miss positions that were never created. Even corporate announcements are slippery. Nearly 113,000 announced U.S. job cuts have been attributed to AI this year, but executives have reasons both to emphasize AI-driven efficiency and to downplay replacement depending on who they are talking to. I would be skeptical of any precise number claiming to tell us how many jobs AI has already destroyed.
What seems clearer is the shape of the transition. Workflows are changing before aggregate employment, hiring appears to be responding before layoffs, younger workers face unusual pressure, and AI can increase demand for labor at some firms while reducing it elsewhere. How painful that transition becomes may depend as much on speed as on the eventual net job count. A gradual shift gives workers time to acquire new skills and companies time to redesign jobs. A fast, concentrated one leaves much less room to adjust. So I don’t think the evidence supports either “AI is destroying the job market” or “there is nothing to worry about.” The more useful conclusion is that the effects are arriving unevenly, through several channels at once, and our measurement systems are still catching up.
Automating Work and Redesigning Work Are Different Decisions
There is one more possibility I think is easy to miss in the usual debate over whether AI will create or destroy jobs. Employment can hold up, productivity can rise, and workers can still share less in the gains. Automating an existing task and creating new work for people to do are two different things, and historically both have mattered. One estimate suggests that roughly 40% to 50% of the human work supported by today’s U.S. labor market comes from tasks that are relatively new. That helps explain how earlier waves of automation could coexist with rising wages. Workers did not simply lose old tasks. New ones emerged around the technology. If AI proves much better at absorbing existing work than creating new things for people to do, the headline employment numbers could remain fairly stable while wage growth tells a less encouraging story.
What makes this especially relevant for companies is that the outcome is not dictated by the technology alone. There is a useful precedent in manufacturing. When automakers in Japan, South Korea, and parts of Germany introduced robots, they also redesigned jobs and created new tasks around them. U.S. manufacturers did less of that. The broader lesson is that automating work and redesigning work are separate choices. AI tells a company what has become cheaper to do. Management still decides whether to use that capability mainly to remove work, expand output, or give people new things to do. For anyone deploying AI now, I think that is the question worth adding to the usual productivity discussion: not just “What can we automate?” but “What can our people do now that they could not do before?”
