AI and Entry-Level Jobs: What the Evidence Means for Hiring and Training

Both questions belong in one decision. AI may reduce the work that justifies a junior role today. It may also change how a beginner learns the business. A useful hiring plan has to account for the work that remains and the experience your company needs to develop.

The employment research gives us reasons to pay attention. It does not provide a staffing formula for your company.

What the employment figures actually tell us

When researchers call an occupation “exposed to AI,” they generally mean that AI could perform or help with some of its tasks. Exposure is not a count of jobs already replaced. Anthropic’s research explicitly distinguishes possible uses from uses observed in its own products. Anthropic’s labor-market study

Three points might be especially important.

Young workers are showing a concerning employment gap. An August 2026 Stanford study used ADP payroll data through June. Employment among workers aged 22–25 in more AI-exposed occupations was 19% below where it would have been if it had kept pace with their less-exposed peers. The gap developed mainly due to weaker hiring, rather than more people leaving/losing jobs.

That does not mean AI fired 19% of junior employees. The authors describe a pattern, not a proven causal effect. Some differences predate widespread generative AI, and accounting for education reduces the estimated gap. The 22-25 age group also does not capture everyone entering a new occupation. Stanford’s revised study

Other research finds tentative signs of slower hiring. Anthropic’s March 2026 analysis found no clear increase in unemployment associated with higher AI exposure, but some evidence that people aged 22–25 were less likely to start jobs in the most exposed occupations. The hiring finding was uncertain, and the study’s measure partly reflects use of Claude, rather than all AI products. Anthropic’s findings

The broader picture is less clear. In its September 15 update, Yale’s Budget Lab reported that its measures had not identified a clear AI-related change in the national labor market. Its analysis did not find a connection between measures of AI use and changes in employment or unemployment. Yale’s labor-market tracking

These studies use different data and methods. A problem concentrated among people entering particular occupations could coexist with little visible change across the economy. Equally, a gap between occupations does not tell us how much AI caused it.

I think that the warning deserves attention, especially around hiring beginners. It is too early to turn these findings into a blanket instruction to stop hiring them.

Look inside the job you are trying to fill

For a construction or manufacturing company, “entry-level work” covers very different responsibilities. An apprentice learning a trade, an assistant estimator, and a new purchasing clerk should not be treated in the same manner just because they are all beginners.

Start with the proposed position’s actual work.

Consider a hypothetical junior purchasing role. Part of the day might involve copying information from supplier emails into a system. Another part might involve checking quantities, clarifying delivery dates, and finding out why a supervisor requested a different item.

Suppose AI reliably handles working with the email - not a big supposition today. You still need to establish how much time that saves, who checks the result, and whether the remaining work requires more or less judgment than before.

If routine work shrinks substantially, redesigning the position or postponing an additional hire may make sense. If it frees a beginner to investigate discrepancies and work the phones, the same technology may make that employee more useful.

Those are alternative business outcomes to test. A demonstration of one automated task cannot decide between them.

AI can help beginners perform better

There is evidence for that possibility. A study published in The Quarterly Journal of Economics followed the introduction of an AI assistant for 5,172 customer-support workers. It found a 15% average increase in issues resolved per hour, with larger gains for less experienced and lower-skilled workers. The researchers also found evidence of learning that persisted when the assistant was temporarily unavailable. Generative AI at Work

This was one company’s customer-support setting. The study did not establish whether AI would increase or reduce overall employment, and its productivity figures should not become a savings promise for a construction office.

It does give an owner a reason to test a different proposition: could a well-supported beginner become productive sooner?

That might be valuable where a senior employee spends much of the day answering recurring questions. The trial should establish whether AI provides useful help and whether the beginner can recognize when a question needs the senior person’s attention.

Finishing the task and learning the task are different outcomes

A smaller study illustrates the other side. In a randomized experiment involving 52 mostly junior software engineers learning an unfamiliar programming tool, the AI-assisted group averaged 50% on a subsequent quiz, compared with 67% for the group without AI assistance. The time saving was small and not statistically significant. Anthropic’s skill-formation study

The experiment measured understanding shortly after a coding exercise. It does not establish the long-term effect on every kind of employee. Its practical warning is narrower: completing a task with assistance does not necessarily mean the person learned how to do it.

For a business, this translates into a training question: what must this employee understand well enough to check, explain, and eventually handle independently?

Take a practice stock receipt. A new employee may need to recognize that a supplier lists cases while your inventory system counts individual units. If the software enters a plausible quantity, the employee still needs to understand why that quantity is right.

I would build training around a few completed examples, including mistakes. Ask the employee to explain the source information, check the units, and identify what would make them stop and ask for help. Let them use AI to explore an explanation, then check it against the company’s actual records and procedures.

Occasionally, use a short practice exercise without AI. This is a way to see what has been learned before assigning more responsibility. The purpose is to preserve judgment, without requiring people to keep doing every repetitive task manually.

Give beginners a deliberate route to experience

If you remove the routine tasks from a junior job, decide what replaces the learning those tasks used to provide.

For example, an assistant estimator might spend less time assembling a first draft and more time comparing it with the source documents alongside an experienced colleague. A purchasing assistant might review why a proposed substitute was rejected. A project coordinator might follow a discrepancy from the field report through to its resolution.

The point is to give beginners exposure to decisions and their consequences. Assigning them only the unusual cases may also be a mistake: the hardest exceptions are a poor starting place for someone who has not yet learned what normal work looks like. (As an aside: do you remember what normal work looks like? :-) )

A manager should be able to answer three questions: what can the employee handle now, what are they learning next, and who has time to teach and review it?

If nobody has that time, adding an AI is not a training program.

Before changing headcount, measure the whole job

Choose a workflow and compare ordinary work before and after introducing AI. Include busy periods and awkward cases, not just the clean examples used in a demonstration.

I would track four things:

Include software costs and the time spent maintaining the process. Then use those results to decide whether to change the role, fill the vacancy, or postpone an additional hire.

For recruiting, I would use a short, realistic work sample with the same AI access the role will have. Ask the candidate to explain the result, identify something that needs checking, and respond to a changed requirement. Keep the exercise appropriate to a beginner. You are looking for a person who can learn and take responsibility for the work they submit.

There will be cases where fewer people are needed for the same workload. There will also be cases where AI makes hiring and training a beginner more practical. Your measurements should help you recognize which situation you face.

The question I would bring to a staffing meeting is: After we automate this work, what will our new employee do, and how will they learn to become the experienced person we need?

Next in the series: AI Safety for Business: What to Require Before Deployment. Coming next.

Previous: Rogue AI Agents: What the Hugging Face Incident Means for Your Business · Return to the AI risk series guide