There has been a lot of discussion around a decline in entry-level hiring over the past two years and the follow-on challenges this could present over time. More often than not, the cause given is: AI can now do the work an entry-level hire used to be brought in for. On the surface, this sounds like machines are coming for our jobs. People are less valuable in a world where AI can automate more than ever before. I don’t think this is the whole story. It’s an opportunity to discuss an underlying issue that isn’t new. How do you effectively develop people?

AI does not replace someone without experience, but it can uplift their ability to contribute more quickly. Erik Brynjolfsson, Danielle Li and Lindsey Raymond studied the rollout of an AI assistant to more than 5,000 customer support agents, published in the Quarterly Journal of Economics in 2025. As a result, productivity rose 14 percent on average. For the newest and least skilled workers it rose 34 percent, with almost no effect on the experienced ones.[1] The conclusion: a person in their first year can now take on work that would once have taken three or four years to grow into. An entry-level hire has never been more capable on day one.

The default nobody chose

This is not to say organizations aren’t trying to build good onboarding programs and make sure new hires are well trained. But often the devil is in the details. Tools now let each of us accomplish more on our own. So it becomes easy to default to working in a way that leaves people productive and disconnected at the same time.

The number everyone is citing

A study earlier this year shed some light on what is happening with entry-level hiring, and it’s worth noting. In August 2026, Erik Brynjolfsson, Bharat Chandar and Ruyu Chen at the Stanford Digital Economy Lab published an updated analysis. It draws on ADP payroll records running through June 2026. They found that employment among workers aged 22 to 25 in highly AI-exposed occupations sits about 19 percent below where it should be. The comparison is against similarly aged workers in less exposed occupations. That gap was 15 percent a year earlier, so it is widening.[2]

If you only look at the high level numbers, two details get overlooked. The first is that the adjustment runs through reduced hiring rather than through layoffs. The second is the authors’ own caution. They describe their findings as descriptive patterns rather than causal estimates. They note that the gaps shrink once education is accounted for.

The counter argument almost nobody is referencing

In May 2026, Richard Audoly, Miles Guerin and Giorgio Topa at the Federal Reserve Bank of New York ran a different test. Instead of payroll records they used job postings, matched against an occupation-level AI exposure measure. Specifically, they asked whether the slowdown was concentrated in entry-level hiring.

It was not. They found that labor demand for entry-level and senior roles within highly exposed occupations is moving broadly in parallel. They concluded that AI may be contributing, but it is not the main driver of the slowdown in hiring.[3]

If AI is not the main driver, what is? A second New York Fed study, published a few weeks later, points at the hiring decision itself. Natalia Emanuel, Emma Harrington and Amanda Pallais conclude that employers may not want to hire fresh graduates onto distributed teams. That is because it is more difficult to teach them the requisite skills from afar. By their own back-of-the-envelope estimate, that alone explains 64 percent of the rise in unemployment among young college graduates. The comparison runs from 2017 to 2019 against 2022 to 2024.[4]

Next comes the detail that matters. At the firm they studied, for instance, hiring of inexperienced workers fell while the offices were closed and recovered once they reopened. For positions on distributed teams, though, the firm kept hiring experienced people even after reopening. It was never the building. Nothing on those teams had replaced the training that sitting together used to deliver by accident.

What was being delivered

The same three authors have a second study, in the Quarterly Journal of Economics, that puts a number on the training itself. They followed 1,055 software engineers at a Fortune 500 online retailer from 2019 to 2024. The office closures of 2020 and the reopening that followed gave them two natural experiments in who sat near whom. While the offices were open, engineers sitting in one building with their whole team received 23.9 percent more comments on their programs. Specifically, the comparison is against engineers whose teams were spread across several buildings. Everyone was in an office, but only some of them were together.

Once everyone was remote, that gap shrank to 7.1 percent. For the engineers early in their careers, who had been getting the most feedback of anyone, it disappeared. Indeed, the paper describes their feedback declining and converging to a uniform lower level, regardless of tenure. Ultimately, nobody was getting less than anyone else. Everyone was getting less than before.[5]

This highlights what happens when teams rely on ambient feedback. When everyone sits together it works. It’s not culture, not belonging, not a value on a wall. Remove the proximity for any reason, a second office, a new hire in another city, a team that goes remote, and the feedback goes with it. Nothing was holding it up except the seating. Without deliberate intention, the team suffers, and the newest people suffer first.

What deliberate has to mean

If ambient feedback was a byproduct, the question for entry-level hiring is whether it can be produced on purpose. Socialization research suggests that it can.

Talya Bauer has been measuring this for two decades. In 2007, she and colleagues pooled 70 studies of 12,279 newcomers in the Journal of Applied Psychology. Structured onboarding was associated with a new hire’s confidence in the role at r = .42 and their clarity about it at .27. Both of those also tracked performance, at .35 and .29.[6] In 2025 she updated the work in the Journal of Management, covering 256 studies, 183 of which had enough data to pool. Mentoring and support showed the strongest direct relationship with a newcomer’s social acceptance, at r = .36. Social acceptance in turn related to performance at r = .35, and task mastery to performance at r = .38. However, the largest single effect in the analysis was negative. Being undermined by a coworker or a manager related to social acceptance at r = -.42.[7]

All of these drivers are relational. They make a solid case for designing entry-level hiring around deliberate support that’s more than a weekly 1:1. Still, nobody has yet measured that on a distributed team specifically.

So ask the better question

The question is not whether AI has made entry-level hiring unnecessary. It is whether anyone at your company is responsible for the feedback a new hire needs. Whether that person has also been given the time it takes. And whether anybody would notice if they stopped.

If the honest answer is no, it’s time to look at how to be more intentional about the feedback that makes entry-level hiring pay off. The good news is, it’s a design problem, and design problems have owners. Own it, and the entry-level hire becomes the best-leveraged person on the team.