AI Is Taking Over Junior Work, So Who Becomes the Senior Later

Summarizing documents, drafting a first version, tidying data, answering standard customer questions. That used to be the new hire's job, and also how they learned. Now AI can do all of it in seconds. On productivity, that's good news. But there's a question that rarely comes up in meetings: if those practice tasks disappear, where do the seniors of five years from now learn?

The honest answer is that nobody knows for sure yet. In this article I've pulled together what has data behind it, what is still debated, and what you can do in your company now without waiting for certainty.

The data is still early, and researchers don't agree

The most-cited study comes from the Stanford Digital Economy Lab. Erik Brynjolfsson, Bharat Chandar, and Ruyu Chen used ADP payroll data from the United States. In the August 2026 update, employment of workers aged 22 to 25 in highly AI-exposed occupations now stands about 19% below where it would be if it had kept pace with similarly aged workers in less-exposed occupations. The figure was 15% in the July 2025 data. The adjustment mainly comes through reduced hiring of young workers, not more layoffs. Meanwhile, in occupations where AI mostly complements workers, employment is flat or rising, especially among more experienced workers.

The authors themselves are cautious. The gap shrinks when education is accounted for, some differing trends were already visible before generative AI was widely used, and the gap is larger in the ADP sample than in national surveys. They also stress that this study, or any single study, shouldn't be treated as definitive evidence.

A different view comes from the Yale Budget Lab. In its tracker updated September 15, 2026, it concludes that the occupational mix isn't yet changing in ways that clearly line up with the arrival of AI, and that measures of AI usage show no connection to changes in employment or unemployment.

So here's the picture: there's a signal on the hiring side for young people, but no agreed conclusion. The data is also from the United States. I haven't found a comparable payroll-based study for Indonesia, so treat this as a signal worth watching, not a forecast for your company.

Why this is a business owner's problem, not just a young people's problem

A senior is a junior who stayed and learned. A lot of expertise in a company isn't written in any SOP: how to spot a customer who will cause trouble, a gut feeling about numbers that look off, knowing when a rule can be bent. That kind of knowledge moves through working together.

There's a theory paper on exactly this. In Enrique Ide's economic model, novices pick up knowledge that isn't written down (tacit knowledge) by working alongside experts. The result: automating entry-level tasks can raise output when it's adopted, but can reduce long-run growth if it pulls novices away from the experts. It's a theoretical model, not a measurement from real companies, so I use it as a way of thinking, not as evidence.

I once built a system that let one admin handle a process that used to take four admin roles, by running the system and checking the results. The story is in How I Cut Operating Costs From Tens of Millions of Rupiah to Hundreds of Thousands. It was a good result for costs. But the question in this article applies to every company that does something similar: the routine work that disappears was also where people learned.

Learning needs difficulty, and how you use AI decides the outcome

Two recent studies looked directly at what happens to people who learn with AI's help.

Anthropic published a randomized trial on January 29, 2026. The participants were 52 software engineers, mostly junior, asked to learn a new Python library called Trio. The group that used AI averaged 50% on the quiz afterward, versus 67% for the group that wrote the code by hand, a gap equal to nearly two letter grades. The AI group finished about two minutes faster, but that time difference wasn't statistically significant. How people used AI mattered: those who handed almost everything to AI (writing code or debugging) averaged under 40%, while those who used AI to ask conceptual questions, not just to request code, scored better.

The second study, Bastani et al. in PNAS (2025), involved nearly a thousand high school students learning math with a GPT-4-based tutor. During practice, the version that looked like regular ChatGPT (GPT Base) raised practice performance by 48%. But when the AI access was taken away, their scores were 17% lower than those of students who never had it. The GPT Tutor version, designed to protect learning, raised practice performance by 127%, and most of the negative effect disappeared.

Quiz score (%), AI-assisted group vs hand-coding group, Anthropic 2026

50 vs 67

Exam score of students used to GPT Base once the AI was taken away, PNAS 2025

-17%

How much faster the AI group finished, not statistically significant

~2 min

The contexts differ from an office: one studied learning a coding library, the other high school students. But the pattern is the same. AI that hands over the answer raises performance today and can weaken your own ability, while AI that guides the learning process protects it. That doesn't mean AI is bad for learning. How you use it decides the outcome.

What you can do now

This section is purely my own suggestions, not findings from the studies above. The goal is simple: let AI do the practice tasks if it can, but don't let the practice disappear with them.

One more thing about hiring. If you've ever decided to stop taking on juniors because AI can do the work anyway, ask again where your seniors three to five years from now will come from. From outside, which means competing with other companies at market prices. From inside, which means someone is learning now. It's your call, but it's worth making deliberately. On AI literacy in the team, I also cover it in Mistakes Companies Make When Adopting AI.

Wrapping up

I'm not saying you should ban AI for new hires. What I'm saying is that if AI takes over the practice tasks, the practice needs to be redesigned on purpose. The data on the effect on employment is still early and debated. But I think the cost of redesigning how your team learns now is far smaller than the cost of running short of people who understand operations later.

If you're working out how to split work between your team and AI without shutting down the path to learning, tell me about your team's situation first.