AI can improve a novice's work while leaving important learning unfinished. It can also help people understand unfamiliar tasks. The effect depends on the work, the assistance and what the learner actually does. A better deliverable does not establish that someone is becoming ready for greater responsibility.
For a manager redesigning junior work, the useful question is what an assignment used to teach and where that learning will happen now. Keep worthwhile decisions, feedback and exposure to the outcome in the job. You can change the task while preserving its developmental purpose.
Here, apprenticeship means supported learning through work, including ordinary junior roles. This guide focuses on assignment design. The wider people and change guide covers the support around adoption.
Separate assisted performance from lasting learning
Research gives reasons to examine the learning process, rather than assume either benefit or harm.
In Shen and Tamkin's 2026 coding experiment, 52 Python users learned an unfamiliar programming library with or without a chat-based AI assistant. The AI group scored lower on an immediate quiz without AI; the average completion-time difference was not statistically significant. Participants were new to that library, rather than all being junior programmers. The short session did not establish what they would retain months later.
A contrasting October 2026 working paper by Brosch and colleagues studied 673 final-year IT apprentices at twelve German vocational schools. Randomly assigned AI access improved performance on tasks within and beyond their curriculum, and average immediate comprehension also improved. The comparison was AI plus internet access against a no-internet condition. Four questions after each task cannot establish lasting transfer to another job or task.
Longer observation also needs careful interpretation. In Autor and colleagues' patent-drafting study, 133 practising lawyers at eleven US firms took part in a three-month randomized trial. AI access improved assisted benchmark work, with larger gains among junior lawyers. On a final exercise without AI, the average advantage was concentrated among senior lawyers. Juniors showed no average gain, with more varied outcomes. This Google-funded working paper concerns a particular drafting assistant and profession; it does not establish that every junior lost skill.
There is evidence of beneficial learning at work too. The 2025 customer-support study by Brynjolfsson, Li and Raymond followed a staggered rollout involving 5,172 agents at one business-software firm. During AI outages, previously exposed workers continued handling chats faster than their earlier baseline, consistent with learning. The outage estimates were noisy and could reflect different chat types. They did not establish promotion readiness or the future supply of experts.
These studies test different tools, people and outcomes. They support keeping assisted delivery and independent understanding visible separately. They cannot be pooled into a single estimate of what AI does to an apprenticeship pipeline.
Find the decisions inside the old assignment
Some repetitive work consumes time without teaching much. Other assignments make a learner choose a source, discover a mismatch or see why an experienced colleague takes a simpler approach. Identify that contribution before deciding what should disappear.
In a September 2026 LinkedIn article, consulting leader Louise Keely recalls building a financial model as a junior at Bain. Her manager rebuilt a much simpler version the next morning. Comparing the two helped her see which detail the manager considered unnecessary. Her own attempt gave her a basis for understanding the correction.
Keely now describes Quantum Rise's relatively small, senior advisory teams and explicitly leaves their long-term talent model open. Her account illustrates a learning mechanism, rather than proving that AI removes it. The important question is whether the junior still gets to make and examine a meaningful attempt.
When reviewing an assignment, look for:
- A decision to make. What must the learner choose, explain or investigate?
- Evidence to inspect. Which records, observations or customer details make the choice defensible?
- Feedback to encounter. Who explains a correction, and can the learner see what happens afterward?
- A next responsibility. What more demanding work should this experience prepare them to attempt?
If you cannot name any of these, do not defend the task merely because previous employees had to do it. If you can name them, decide where they will sit after automation.
Redesign the assignment before raising the target
Suppose Sam Taylor manages purchasing at a furniture manufacturer. Jamie Lee, a junior assistant, compares freight invoices with carrier quotations and delivery records. An approved AI tool can extract the charges and draft a discrepancy summary.
Jamie notices an additional collection charge in an invoice that does not appear in the quotation. The job is to identify the mismatch and prepare a question for Sam, who decides how to raise it with the carrier. Jamie does not approve payment or decide that a charge is invalid merely because AI flagged it.
Removing the extraction need not remove the learning. Sam can change what Jamie owns while keeping the source comparison and follow-through within reach.
| Assignment choice | What Jamie practises | What Sam must provide |
|---|---|---|
Extract every field manually | Locating invoice and quotation details | Feedback on relevant mistakes; a reason the repetition is useful |
Use AI extraction and investigate a selected discrepancy | Comparing evidence and proposing an explanation | Access to the records and review of Jamie's reasoning |
Observe the carrier clarification and update the case | Connecting the original question with its resolution | A permitted view of the outcome and an explanation of the decision |
Receive only the finished summary | Reading a conclusion | Another assignment that supplies the missing investigation and feedback |
Start Jamie with cases where the relevant records are available and Sam can intervene before an external action. Later, vary the discrepancy and the responsibility as Jamie demonstrates understanding. A more demanding assignment might require Jamie to distinguish an unexplained charge from a charge supported by a changed collection request, then explain the difference to Sam.
This is a local design choice, not a proven training formula. Sam must allow time for the investigation and review before increasing the invoice target. Otherwise the formal learning assignment competes with a delivery expectation that rewards skipping it.
Use the role-specific learning guide to turn the chosen responsibility into demonstration and practice. The assignment decision comes first: the learner needs real opportunities to use what the exercise teaches.
Listen to uncertainty without diagnosing skill loss from it
A Reddit user describing four months of SAP development experience worried that AI helped them finish invoice-printing and debugging tasks while leaving them with little learning. They also reported encouragement from senior colleagues about their ability to explain the work.
In a clarifying reply, the developer described trying diagnostic plans, comparing working test cases and explaining the problem and their steps to a senior for confirmation. That is more active than the headline alone suggests. The employer and capability were not independently assessed.
Take that kind of concern seriously without turning it into a verdict. Ask the employee to show a decision they made and a point they still cannot explain. Ask the supervisor which responsibility they are now prepared to let the employee attempt. A productive conversation may reveal a capability gap, a lack of confidence or too few opportunities to try the next task.
Keep access to learning separate from evidence of learning. Jamie might perform poorly because the needed delivery record never reaches them, or do well because Sam supplies every interpretation. Record those conditions before judging the person. The proficiency guide explains how to sample a comparable changed task and record help.
Check whether the new work still leads somewhere
Review a small run of actual assignments with the junior employee and supervisor. Look beyond the number of completed outputs. Has the person encountered different cases, received an explanation of corrections and seen an outcome? Are they taking on a more demanding decision with appropriate support?
For Sam and Jamie, keep a short record of the discrepancy, Jamie's explanation, help supplied and what the carrier clarification established. When a later case is comparable, notice what Jamie can now investigate without Sam supplying the reasoning. Keep tool access and task differences attached to the observation. A few cases guide the next assignment; they do not prove a causal training effect.
If the new workflow leaves juniors only the hardest exceptions, restore a supported route into those cases. That could mean joining an experienced colleague's investigation, attempting a bounded part first or revisiting a resolved case before taking responsibility for a live one. Do not require a novice to carry consequential work simply to create a learning opportunity.
Questions about AI and apprenticeship
Can AI help junior employees learn?
Yes, AI can help people understand unfamiliar work, but a good assisted result alone does not show learning. The studies discussed here include both beneficial and adverse comprehension findings under different conditions. For a local decision, observe what the employee can explain and apply on a comparable task, and record the assistance available. Compare the evidence and its limits before choosing a blanket rule.
Should managers keep manual tasks that AI can already do?
Keep or replace a manual task according to what it teaches and what the job requires. Re-entering the same fields indefinitely may add little; investigating why two records disagree can expose an important decision. Identify the learning purpose, then choose a supported assignment that still supplies it. The freight-invoice comparison shows several choices without assuming that manual work is always better.
Is checking AI output enough to train a future expert?
Checking can be useful practice when the learner understands the underlying work, has evidence to inspect and receives meaningful feedback. Merely accepting finished summaries may leave important decisions invisible. Give the employee a supported way to investigate causes and see outcomes before assigning harder exceptions. Use the assignment questions to identify what the checking role actually teaches.
Does weaker junior learning mean a company should stop hiring juniors?
No such staffing conclusion follows from a learning concern alone. First examine the remaining work, demand, supervision and opportunities to develop capability in the particular role. A hiring decision needs those facts alongside relevant employment evidence. The entry-level careers guide separates that decision from short-term task performance and broad labor-market headlines.
