AI Made the New Starter Faster. You Still Owe Them an Apprenticeship.

AI can make a new starter productive before they are ready to carry the judgement behind the work. Treat speed as borrowed capability, then rebuild apprenticeship around explanation, exposure, review and accountable human decisions.*

An AI-assisted new starter may produce a plausible customer brief, control summary or claims chronology in their first week. That surface quality does not establish that they understand which fact changes the answer, when the source is weak or why an exception needs escalation.

The answer is not to delay access to approved AI. It is to stop confusing assisted output with independent capability. Give new starters the tool, then design the work so they must expose their reasoning, encounter edge cases and receive expert feedback before any authority expands.

That is the apprenticeship leaders now owe their teams. In banking, insurance and superannuation, it should be treated as part of operating design, not an optional learning activity squeezed around production.

Why is faster not the same as ready?

The strongest field evidence cuts both ways. A 2025 study in The Quarterly Journal of Economics* examined a generative AI assistant used by 5,172 customer-support agents. Access increased issues resolved per hour by 15 per cent on average. Less-skilled and less-experienced agents increased issues resolved per hour by approximately 30 per cent, and treated agents with two months of tenure performed as well as untreated agents with more than six months of tenure on the study's productivity measure. The researchers also found evidence of learning during system outages, when suggestions were unavailable. But the setting was one firm and one type of work, so it is evidence about a mechanism, not a universal productivity promise. The study says exactly that.

There is another warning in the same result. The most skilled agents saw little productivity benefit, and the quality of their conversations declined slightly after receiving AI assistance. The tool helped transmit established practices, but stronger practitioners could be pulled towards its average. That matters because apprenticeship is not only the transfer of standard practice. It is also learning when the standard pattern does not fit, which is the same reason leaders need to protect the reps that build expertise.

In regulated professional work, the hard part is often not producing the first draft. It is noticing that a customer's circumstances do not match the template, that a control description overstates what happens in practice, or that a source supports the figure but not the conclusion. A polished answer can hide those missed judgements.

Jobs and Skills Australia's national study concludes that generative AI is more likely to augment work than replace it. Its skills analysis stresses confident, critical and responsible engagement with digital technologies, alongside the continuing importance of non-digital capabilities.

Treat early speed as borrowed capability. It is useful, but it remains borrowed until the person can explain, test and correct the work without leaning on fluency as proof.

What work must a new starter still do?

Start by defining readiness as four observable acts.

1. Notice: identify the facts, assumptions and missing information that could change the outcome.
2. Explain: state the applicable process, evidence and reasoning in their own words.
3. Challenge: find where the AI output is unsupported, incomplete or too certain.
4. Own: make or escalate the human decision within their actual authority, then record why.

These acts turn an output into evidence of capability. They also prevent a familiar management error: judging the learner by the document's surface quality when the assistant supplied much of that surface.

Build a task-to-capability map for the role. List the recurring work, the judgement hidden inside it and the evidence that would show the person can carry that judgement. For a financial-crime analyst, the task might be drafting an alert narrative. The hidden judgement is deciding which transaction pattern is material and what further evidence is needed. For an insurance team member, it might be preparing a coverage summary. The hidden judgement is identifying which policy wording and facts need a qualified practitioner to resolve.

Then assign work at three levels:

• Observe: the new starter watches an experienced practitioner handle a real, de-identified case and asks why each branch was taken.
• Construct: the new starter prepares a view before seeing the AI or expert answer, then compares the differences.
• Defend: the new starter explains a recommendation, its evidence and its limits to the reviewer who holds the authority.

Do not promote someone from observe to defend because they have completed a fixed number of cases. Move them when their evidence is consistent across an ordinary case, an ambiguous case and an exception.

Safe Work Australia identifies lack of role clarity, poor support and insufficient training as risks when AI and digital technologies are introduced. It also warns that automating routine tasks can leave workers with more complex, cognitively demanding work or with more of their work focused on reviewing AI output. Its guidance treats these hazards as workplace risks to identify, assess and control. Your apprenticeship needs to prepare people for that residual work, not merely teach them to generate routine work faster.

How do you make AI part of the apprenticeship?

Use AI as a scaffold that makes thinking visible. Do not let it become a private answer machine between the learner and the final document.

This prompt is for an explanation-first case review. The new starter must use only approved, de-identified material, and the assigned senior practitioner must review the questions, sources and final rationale before the work is used.

Act as a questioning coach for a new starter in [TEAM_OR_FUNCTION].

Context: [DE-IDENTIFIED_CASE_SUMMARY]
Approved sources: [SOURCE_LIST]
The human decision-maker is: [ROLE]

Do not decide the case or draft the final recommendation.
Ask one question at a time so the learner identifies:
1. the decision or task actually required
2. the facts that could change the outcome
3. the source supporting each material statement
4. missing information and plausible exceptions
5. the required escalation point

After the learner's answers, produce a table with: learner reasoning, evidence cited, unresolved gap, and reviewer question. Mark any claim that is not supported by the approved sources.

The sequence matters. The learner commits to a view before the assistant produces a polished alternative. The senior can then review the gap between what the learner noticed and what the case required.

This second prompt is for an adversarial comparison after the new starter has drafted their own rationale. The qualified reviewer must verify every source, correction and escalation point before accepting the output.

Compare [DRAFT] with the approved sources below. Assist the learner's review, but do not make the decision.

Approved sources: [SOURCE_EXTRACTS_OR_LINKS]
Decision authority: [HUMAN_ROLE]

Return five sections:
1. claims directly supported
2. claims only partly supported
3. assumptions presented as facts
4. missing counter-evidence or exceptions
5. questions the human reviewer should ask

Quote the exact source passage for each finding. If the supplied sources do not answer a point, say "not established from supplied sources".

A June 2026 theoretical working paper from USC models this problem as a choice between AI autonomy and review intensity. Its argument is not empirical proof, but the design proposition is useful: the same technology can support learning or remove learning opportunities depending on who reviews, how closely and at what career stage. The paper frames AI deployment as a talent-development decision. Review intensity is not free, of course; it is the same budget problem as the AI review tax, spent deliberately on the people who are still building judgement.

The point is not to force new starters to complete every task without assistance. It is to preserve the moments in which they form a view, see a consequence and receive correction.

Fictional worked example: [NEWSTARTER] joins a superannuation operations team and is asked to prepare a de-identified exception brief for [CASETYPE]. The approved assistant drafts a clear chronology in minutes. Instead of accepting it as evidence of readiness, [REVIEWER] asks [NEWSTARTER] to mark every material fact, link it to an approved record and identify which gap would change the next action. [NEWSTARTER] finds that the chronology merged an event date with a processing date. The brief is corrected, the human decision remains with [AUTHORISEDROLE], and the missed distinction is added to the team's next case review.

Use this apprenticeship evidence checklist before expanding authority:

• Can the person explain the work without reading the AI output back to you?
• Can they distinguish a source fact, an inference and an unresolved gap?
• Can they identify a case where the standard workflow should stop?
• Can they show what they changed in the AI output and why?
• Can they name the decision they may make and the decision they must escalate?
• Can they repeat the performance when the assistant is unavailable?

Do this Monday

1. Pick one recurring task assigned to new starters. Write down the judgement hidden inside it and the human role that owns the final decision.
2. Select three de-identified cases: ordinary, ambiguous and exception. Confirm that each can be used within your organisation's privacy, security and records controls.
3. Require a short pre-AI rationale for each case. Two paragraphs are enough: what matters, what is missing and what the person would do next.
4. Run one explanation-first review with an experienced practitioner. Capture the correction and the cue that should have triggered it.
5. Set an authority gate for the task. State the evidence required before the person may handle it with lighter review, and retain sampling by a qualified human after the gate is passed.

Do not turn the exercise into a shadow performance score. Its first purpose is developmental: make judgement visible, direct coaching to the gap and improve the work design. If formal performance concerns arise, use your organisation's established process and obtain appropriate advice.

Bottom line

AI can shorten the path to a useful first output. It cannot relieve a leader of building the person who must recognise the exception, test the evidence and carry the responsibility. Design apprenticeship around visible reasoning and reviewed judgement, not time served or document polish. The goal is not a new starter who can use AI quickly. It is a professional who can use AI well, challenge it and still know when the answer belongs with someone else.

This article is general information and education only. It is not legal, compliance, financial or professional advice. Obligations vary by organisation and circumstance. Verify current requirements against the primary sources cited and seek advice specific to your situation.

References

1. Erik Brynjolfsson, Danielle Li and Lindsey R. Raymond, "Generative AI at Work", The Quarterly Journal of Economics, Volume 140, Issue 2, May 2025. https://academic.oup.com/qje/article/140/2/889/7990658
2. Jobs and Skills Australia, "Australia's AI Transition: Jobs, Skills and the Future of Work", final release September 2025. https://www.jobsandskills.gov.au/studies/generative-artificial-intelligence-capacity-study
3. Jobs and Skills Australia, "Our Gen AI Transition: Skills", 2025. https://www.jobsandskills.gov.au/studies/generative-artificial-intelligence-capacity-study/our-gen-ai-transition-skills
4. Safe Work Australia, "Artificial intelligence (AI) and digital technologies - Managing risks", accessed 31 July 2026. https://www.safeworkaustralia.gov.au/safety-topic/hazards/digital-technologies-ai/managing-risks
5. Georgios Petropoulos, Rungs Removed or Scaffolds Built?, USC Marshall School of Business working paper on autonomy, review and learning in early-career labour markets, 8 June 2026. https://ssrn.com/abstract=6899423