AI and early-career development: making room for better on-the-job learning.
AI gives us a chance to rethink how people gain experience at work. I'd like to see junior colleagues spend less time preparing routine output and more time learning how to make the decisions behind it.
That won't happen automatically. Some of the work we can now automate was also how people learnt. The opportunity is to build a better route into experience as we change the work itself.
What People Learn Through Work
Consider a junior consultant preparing an analysis. The client gets the analysis. The junior gets a chance to learn which assumptions survive contact with a client. A supervisor's review might explain why a technically correct recommendation would fail in practice.
The work has a second purpose: helping someone make a better decision next time. That learning is easy to miss when we count the hours AI could save.
If AI handles the first pass, I'd use some of the time saved for practice in the decisions that follow. That practice needs experienced supervision.
Make Practice Part of the Job
Nathaniel Whittemore has proposed making senior judgement available through collaboration and capturing it in examples or evaluation criteria. I'd build on that with practice in the decisions people need to learn. The AI Daily Brief, 18 August.
Take a junior recruiter. Preparing a shortlist might have taught them to distinguish relevant experience from impressive wording. If AI prepares the list, the junior still needs to judge whether a borderline candidate deserves an interview.
AI could help turn everyday reviews into reusable examples of how experienced colleagues reached a decision, with those colleagues checking the record. With staff aware of what is being reviewed, it could also help supervisors compare selected cases over time against agreed criteria, reducing the paperwork involved in tracking progress.
How I'd Check Whether Judgement Is Improving
I'd protect time for those reviews, then check whether people make fewer substantive errors or need fewer corrective interventions on unfamiliar cases of comparable difficulty. For the recruiter, I'd test this sequence:
- Set a baseline. Agree the criteria for a sound decision, then ask them to assess a candidate. Record decision quality and where they need help.
- Review an AI first pass. Ask them to challenge a shortlist and identify assumptions that need checking.
- Practise with feedback. Vary the evidence in a borderline case. Have them decide before discussing the evidence with a senior.
- Try an unfamiliar case. Assess decision quality against the same criteria and record the support needed. Ask for an explanation without AI supplying it.
- Return to it later. Check whether the learning lasts. Keep tools comparable so an improved model doesn't masquerade as an improved person.
A searchable record preserves the business's knowledge. It doesn't establish the learner's competence. Education research supports testing whether learning transfers. This particular application still needs testing. Learning and transfer.
Fewer requests for help aren't automatically progress. Recognising insufficient evidence and asking for advice may show better judgement than confidently carrying on.
Course completion is easy to count. Competence is inconvenient like that.
What the Evidence Supports
Anthropic employees describe using AI to attempt work beyond their previous expertise. Wider exposure is encouraging, though their account doesn't establish deeper competence. Anthropic's workplace research.
In Anthropic's randomised study of developers learning an unfamiliar Python library, the AI-assisted group scored lower on a subsequent comprehension test. The largest gap was on debugging. This small study of immediate learning cannot predict a career. It does give us a reason to check understanding separately from completed work. Anthropic's study.
Datacom's 2026 State of AI Index reports that 91% of surveyed organisations use AI, while 4% say it is transforming core operations. A third report returns exceeding costs, so useful gains don't have to wait for wholesale redesign. Datacom's report.
The employment evidence is mixed too. Stanford's August update found employment among 22 to 25 year olds in highly AI-exposed occupations about 19% below the level implied by keeping pace with less-exposed peers. Weaker hiring accounted for most of the adjustment. The researchers do not claim a causal estimate of AI's effect. Stanford's findings.
Australia's Department of Employment and Workplace Relations found no broad AI-driven labour-market upheaval in its July report. It noted slower growth in more exposed occupations, while stressing the uncertainty. Meanwhile, Ramp and Revelio found entry-level headcount grew 12% over two years among the heaviest AI adopters in their US sample. Those businesses were already faster-growing, so that doesn't prove AI created the jobs. DEWR's findings, Ramp's research.
I don't read these findings as a reason to delay useful AI adoption. I read them as a reason to assess learning alongside output, rather than assume one proves the other.
Why the Learning Needs a Plan
Suppose a firm stops hiring graduates for two years because AI can handle their research. Experienced staff keep delivery going. Later, the firm needs more people who can take responsibility for client work. The graduates it would normally have trained aren't there.
A new intake won't arrive with the experience of the people it never trained. That delay belongs in the business case, alongside the savings.
New Zealand's public-service plans make this a timely question. In May, Nicola Willis set a target of no more than 55,000 full-time equivalent public servants by July 2029, 8,700 fewer than in December 2025. The wider reforms promise $2.4 billion in savings over four years, with AI among the measures. By July, Christopher Luxon said the Government was "well behind" on AI and other technologies. Government announcement, Willis's speech, Luxon interview.
That doesn't make these 8,700 AI redundancies. We don't know how many junior roles are involved. The Government has also announced a public-service academy and talent-development plans. I'd judge those efforts by whether capability keeps pace with the savings.
The savings have a timetable. The learning needs one too.
Who Takes Responsibility for the Next Intake?
A business can build practice into its own work, but it also draws on people trained elsewhere. We all benefit from that experience. ELab does too. Hiring someone another firm trained doesn't relieve us of responsibility for how the next person gets started.
Some of the judgement once developed in junior roles should become an explicit part of professional study. Employers have something to contribute here, particularly cases drawn from the work students will eventually do.
A curriculum can prepare someone for responsibility. Employers still have to provide opportunities to exercise it.
I would like to hear from employers and educators already changing how people gain experience. Where have you replaced routine work with a better way to learn, and what tells you it is working?
Build Capability Alongside Efficiency
This brings me back to the argument I made in April. Cutting execution before fixing orchestration could remove the people who held work together. Here, the question is how the next generation learns to take that responsibility. The AI Efficiency Trap.
At ELab, our work in AI enablement and governance includes discovery and knowledge capture. That gives us a starting point for identifying what people need to learn when their work changes.
I'd start with one workflow AI has changed and agree what people need to learn. Then put supervised practice into that work, with someone responsible for checking progress.
The opportunity I care about is a team that can take on more demanding work and show how its judgement has improved.
Tushar Singh is Head of AI and Data at ELab. ELab advises organisations across New Zealand, Australia and Hong Kong on AI enablement and governance.
Related: ELab's AI training for business leaders focuses on owners and senior managers developing their ability to evaluate and apply AI.
