For generations, becoming good at professional work involved doing quite a lot of work you weren’t yet very good at.
You didn’t start with the most ambiguous client problem, the politically sensitive organisational issue or the decision where getting it wrong could cost millions.
You started smaller.
You researched. You drafted. You checked. You reconciled. You reviewed documents. You prepared reports. You sat in meetings where someone more experienced noticed things you didn’t. You made relatively low-consequence mistakes and had them corrected.
Then you did it again.
Over time, the work became more difficult. You encountered exceptions. You started recognising patterns. You learned which details mattered and which were noise. Eventually, you developed something much harder to teach than technical knowledge.
Judgement.
Now AI can perform a growing amount of the work at the bottom of that developmental ladder.
That is a genuine productivity opportunity.
But it raises another question:
What happens to the ladder?
We may be automating more than tasks
Much of the discussion about AI and work begins with a sensible question:
Which tasks can AI do?
Organisations are identifying work that is structured, repeatable, time-consuming and relatively easy to specify. First drafts. Basic research. Summaries. Routine analysis. Document review. Administrative processing.
There is good reason to do this.
Jobs and Skills Australia’s major study of generative AI concluded that the technology is currently more likely to augment Australian jobs than replace them, with significant potential to improve productivity and change how work is performed. Importantly, it found no clear evidence so far that GenAI has caused a decline in Australian entry-level roles.
So this is not an argument that graduate jobs are about to disappear.
The more interesting issue is what happens inside those jobs.
Because some of the work AI can perform most easily is also work through which people historically learned their profession.
That gives us a second question to put beside the productivity question:
What did people learn by doing the task we are about to automate?
Those are not the same question.
The productive function and the developmental function
Consider a junior lawyer reviewing documents, conducting basic research and drafting routine correspondence.
Look only at the output, and much of that work may appear ripe for automation.
AI can search faster. It can summarise vast amounts of material. It can produce a competent first draft in seconds.
But the junior lawyer was never only producing the document.
They were encountering legal language repeatedly. Seeing how principles applied across slightly different situations. Learning what experienced lawyers considered significant. Discovering exceptions. Building familiarity with precedent, clients and context.
They were beginning to develop a mental library.
Something similar happens in accounting.
Reconciliations, routine audit work and straightforward financial analysis may not be the most intellectually exciting parts of the profession. But working through them can help someone learn what normal looks like — which eventually helps them notice what isn’t normal.
The same applies in consulting. A junior consultant gathering data, analysing interview material and preparing the first cut of a presentation is not merely producing slides. They are learning how evidence becomes an argument, what gets challenged by a client and why a beautifully constructed analysis can still fail when it ignores organisational reality.
And in HR, you do not learn to navigate a politically complex restructuring simply by completing a course on change management. Judgement develops through exposure: employee matters, difficult conversations, competing accounts of events, imperfect policies, managers who tell you half the story, organisational history, power dynamics and decisions where the technically correct answer is not necessarily the organisationally workable one.
The value of these experiences was never fully captured by the economic value of their outputs.
A task can have both a productive function and a developmental function.
AI may make the productive function substantially more efficient — or unnecessary — while leaving the developmental need intact.
The emerging Judgement Gap
I think this creates a potential organisational problem worth naming.
I call it the Judgement Gap.
I’m using that as a Shaping Change framing rather than suggesting it is an established academic construct. It describes a risk organisations will increasingly need to consider:
We may automate the work through which people historically acquired the experience required to exercise judgement.
This becomes particularly important when we consider what AI is likely to leave behind.
As routine and predictable elements of knowledge work become easier to automate or augment, the work requiring significant human involvement may become increasingly concentrated around situations that are ambiguous, exceptional, relational, politically sensitive, context-dependent, ethically difficult or high consequence.
In other words, we may remove some of the easier work while increasing the proportion of hard work.
That creates a paradox.
We could need more sophisticated human judgement at precisely the same time that we are removing some of the experiences through which that judgement used to develop.
This isn’t merely hypothetical concern about technology.
The World Economic Forum’s 2026 work on entry-level employment identifies talent pipelines as one of the areas organisations need to reconsider as AI changes early-career work. More than one in three young workers globally are already employed in occupations considered to have medium-to-high exposure to AI-driven task change.
Jobs and Skills Australia makes a similarly careful point. While it has not yet found evidence of an AI-driven collapse in Australian entry-level employment, it specifically identifies those early roles as valuable because they provide foundational career experiences.
The problem, then, is not simply whether those jobs continue to exist.
It is whether they continue to develop expertise.
Complexity doesn’t disappear. It moves.
There is another layer to this.
I think of it as Complexity Transfer.
When AI removes routine work, the complexity that remains with humans becomes more concentrated.
Imagine that 40 per cent of someone’s role consists of relatively straightforward work and 60 per cent requires deeper interpretation, relationships, exceptions and judgement.
Automating much of the straightforward work does not necessarily make the remaining job easier.
It may make it harder.
The person now spends a greater proportion of their day dealing with exceptions, ambiguity and complexity.
From an Organisational Ecology perspective, automation may reduce Friction and increase apparent Capacity. People can process information faster, produce documents faster and remove hours of repetitive activity.
But capacity isn’t simply the number of tasks someone can complete.
If the remaining workload requires more cognitive effort, more judgement and more consequential decisions, the organisation may have removed one kind of demand while intensifying another.
And if it simultaneously removes the lower-complexity work through which people learned to manage that complexity, the immediate productivity gain may conceal a longer-term capability cost.
“They can concentrate on higher-value work” isn’t quite enough
This is one of the phrases I hear repeatedly in discussions about AI:
AI will free people to concentrate on higher-value work.
Perhaps.
But there is an assumption buried inside that sentence.
That people are already capable of doing the higher-value work.
What if they aren’t yet?
If AI writes the first draft, who learns what makes a good first draft?
If AI performs the initial analysis, how does a novice develop the pattern recognition to notice when the analysis is subtly wrong?
If AI handles straightforward cases, where does a developing practitioner encounter enough straightforward cases to recognise an unusual one?
And if tomorrow’s professionals spend their early careers supervising AI-generated work they have never learned to perform themselves, on what basis do they evaluate it?
Research on knowledge workers already suggests this deserves attention. A 2025 study involving 319 professionals found that higher confidence in GenAI was associated with less critical thinking, while workers’ own expertise and confidence influenced how actively they evaluated AI outputs. The researchers found that AI shifts critical thinking towards activities such as verification, integration and oversight.
That matters because effective verification itself requires knowledge.
You need to know enough to recognise that something is wrong.
This doesn’t mean preserving bad work
None of this is an argument for making young professionals spend three years doing tedious work because everyone before them had to.
Some traditional junior work was developmental.
Some was simply cheap labour.
Repetition is not automatically learning, and “paying your dues” is not a development strategy. Research on deliberate practice has long distinguished purposeful practice involving attention, feedback and improvement from mere rote repetition.
So preserving inefficient work for nostalgic reasons would be the wrong response.
The opportunity is better than that.
When we automate a task, we can ask what developmental function it once served and then decide whether that function still matters.
If it doesn’t, remove it.
If it does, redesign it.
Apprenticeship may need to become much more deliberate
For many professions, development has historically happened partly through an informal apprenticeship:
routine work → variation → feedback → increasing complexity → pattern recognition → judgement.
AI disrupts that sequence.
The answer may be to replace accidental apprenticeship with something much more intentional.
That could include simulation, case-based learning, mentoring, shadowing, rotations, deliberate practice, peer review and progressive exposure to increasingly complex situations.
But perhaps the most interesting possibility is learning with AI.
There is evidence that AI can support development rather than simply bypass it. A large field study of customer-support workers found the biggest productivity improvements among novice and lower-skilled employees, with evidence suggesting that AI helped disseminate practices associated with more capable workers.
That suggests the choice isn’t between “learn the old way without AI” and “let AI do everything.”
We can design a third option.
Imagine a developing practitioner asking AI to produce an analysis — but their job does not end when the answer appears.
They must identify its assumptions.
Challenge the evidence.
Find weaknesses.
Generate an alternative explanation.
Identify what information is missing.
Explain what evidence would cause them to change their conclusion.
Compare the AI recommendation with their own.
And finally defend, modify or reject it.
That is a very different developmental experience from typing a prompt and accepting the result.
It begins to develop what we might call Diagnostic Judgement: the capacity not merely to produce an answer, but to understand the situation well enough to evaluate whether an apparently plausible answer deserves to be trusted.
We need to stop calling these “soft skills”
There is another simplification creeping into the AI conversation.
Machines will do the technical work, the argument goes, while humans concentrate on soft skills.
I think that badly understates what humans may actually be required to do.
The capabilities becoming more important include judgement, pattern recognition, contextual interpretation, systems thinking, political awareness, ethical reasoning and the ability to make defensible decisions under ambiguity.
They include knowing when a technically accurate answer is inappropriate for the context.
Recognising the weak signal everyone else has dismissed.
Understanding why two apparently similar situations require different responses.
Knowing when the data is incomplete.
And having sufficient expertise to challenge an AI system that sounds extremely confident.
Those are not “soft” capabilities.
They are sophisticated professional capabilities.
And many of them have traditionally been developed through accumulated experience.
Which means the organisational challenge ahead is bigger than reskilling people to use AI.
We may need to redesign how experience itself is created.
Two questions before we automate
There is enormous value available from AI. I use it myself. I have no interest in preserving inefficient processes simply because they are familiar. But productivity is only one part of organisational capability. Before automating a task, leaders may need to ask two different questions:
What value does this task produce?
And:
What does someone learn by doing it?
If AI can provide the first more quickly, cheaply or effectively, excellent. Then we need to decide deliberately how we will provide the second. Because the question isn’t whether we should preserve junior work that technology can do better.
We shouldn’t.
The question is whether we understand what that work was teaching people before we remove it.
If AI removes part of the pathway to expertise, we need to design a new pathway. We cannot simply remove the ladder and assume people will somehow arrive at the top.
About Ros
Rosalind Cardinal is the Founder and Managing Director of Shaping Change, an Australian leadership and organisational development consultancy.
With a career in organisational development, executive coaching and leadership consulting that began in the late 80’s, Ros helps leaders and organisations navigate the hidden dynamics that shape culture, influence, decision-making and performance.
She is the creator of the Recognition Pathway™, Political Intelligence Compass™, Organisational Hidden Ecology™ and the Women’s Leader Archetypes™ framework.
Book a chat with Ros.
Read Next: You’re Leading the Change. You’re Also Living Through It.