AI can remove the work that once built experience. The challenge is redesigning work so productivity gains don’t weaken future judgment and expertise.

AI can perform more of the junior-level work that once helped employees develop the experience required to become senior, creating a new challenge for businesses using AI to improve productivity.

The issue is not whether businesses should preserve inefficient junior work, but whether they understand what capability that work was also developing before they automate it.

Businesses need to redesign roles, so AI handles work where it creates leverage while employees continue gaining the exposure, judgment and feedback required to take on greater decision responsibility.

AI is creating an unusual structural problem inside established businesses.

The immediate benefit is easy to see. Research becomes faster. Analysis can be prepared in minutes. Proposals require less manual work. Reports assemble more quickly. Routine customer enquiries can be handled with less intervention.

The business requires fewer human hours to produce the same output.

That is real productivity.

But the operating model contains a second system that rarely appears on a process map: the work that produces today’s output has also been producing tomorrow’s expertise.

Junior employees did not simply complete lower-level tasks until somebody decided they were senior.

They accumulated exposure. They encountered variation. They saw exceptions to standard processes. They watched customers behave differently from expectations. They made small decisions, received correction and gradually learned which signals deserved attention.

Production and capability development happened together.

AI is beginning to separate them.

That is the structural fragility.

Most teams misdiagnose it because they analyse AI at task level. They identify an activity, estimate how much time AI could remove, calculate the productivity gain and redesign the workflow around the faster method.

The hidden assumption is that the task’s only contribution to the business was its output.

Sometimes that assumption is correct. Administrative repetition that produces little additional understanding should disappear when a reliable, economical alternative exists.

But other work has been performing two functions simultaneously. It creates an operational output while exposing employees to the cases from which expertise develops.

Remove the production requirement and that exposure may disappear with it.

The mistake is not automation. The mistake is failing to recognise the second function before redesigning the first.

This creates an increasingly important operational tension. AI can allow inexperienced employees to produce more sophisticated work while making the sophistication of that work a weaker indicator of the employee’s underlying capability.

A polished proposal does not establish commercial judgment. A strong analysis does not establish that the analyst can recognise a faulty assumption.

A credible customer response does not establish that the employee understands when the standard response should not apply.

AI can raise output capability faster than judgment capability.

PwC’s 2026 AI Jobs Barometer gives this tension a labour-market dimension. The most AI-exposed junior roles are seven times more likely than the least exposed to require traditionally senior capabilities such as leadership and strategic thinking.

PwC also found that AI-exposed entry-level roles requiring more senior capabilities grew while other entry-level roles declined.

The business consequence can remain hidden because conventional AI measurement captures the visible gain. Hours fall. Output increases. Cost per activity improves.

Yet senior managers continue reviewing unusual cases. Pricing exceptions still move upward. Difficult customer situations still reach the owner.

The proposal takes 30 minutes instead of two hours. The pricing question still lands on the same manager’s desk. The work became faster; the decision architecture didn’t.

The business has improved production without materially changing where scarce judgment resides.

Eventually that matters to growth, throughput and management capacity. More work can reach the same decision boundary faster. Senior people remain difficult to replace.

Delegation stalls because managers trust the output but not necessarily the reasoning behind it.

The architectural principle is therefore straightforward:

When AI removes the work, identify what else the work was producing.

The answer will sometimes be nothing worth preserving. Automate it.

Where the work was also creating exposure, pattern recognition, contextual understanding, feedback or opportunities to exercise judgment, preserve the capability-development requirement—not necessarily the work itself.

The goal is not to maintain the old junior-to-senior pathway. Nor is it to maximise automation.

It is to redesign work so productive effort can fall while human capability continues to deepen.

The evidence of that redesign should eventually appear in responsibility. People should recognise more exceptions, exercise judgment across wider boundaries and independently make decisions that previously travelled upward.

If output capability rises while judgment capability and decision authority remain unchanged, the organisation has gained productivity without fully redesigning capability.

That is not a reason to slow AI adoption.

It is a reason to design beyond the task.

AI Is Breaking the Traditional Junior-to-Senior Path

The traditional career path worked because responsibility increased after experience accumulated. AI can compress the work that produced that experience while increasing demand for judgment earlier.

Traditionally, people started with lower-consequence work. They researched before recommending. Prepared before presenting. Analysed before deciding. Escalated before acting.

Over time, they encountered variation.

A straightforward customer became difficult. A normal quote became commercially unusual. A campaign failed for reasons the standard metrics did not explain.

A supplier problem created consequences somewhere else in operations.

Eventually they stopped seeing isolated tasks and started recognising patterns.

AI changes this sequence because employees can reach sophisticated outputs without travelling through every stage previously required to produce them.

That does not mean AI prevents learning.

Research involving 5,179 customer-support agents found AI assistance increased productivity by 14% overall and 34% among novice and lower-skilled workers, with evidence consistent with AI helping newer workers access practices used by stronger performers.

AI can accelerate development.

But faster access to expert-like output is not the same thing as faster formation of expertise.

We do not yet know whether AI will ultimately accelerate or weaken expertise formation across most professions. Both outcomes are plausible.

What businesses can already control is whether they leave capability development to chance after changing the work that used to produce it.

The labour market is already showing signs that the pathway itself is changing.

Stanford’s 2026 analysis found no widespread economy-wide AI job displacement, but employment among workers aged 22–25 in highly AI-exposed occupations was about 19% below where it would have been had it kept pace with similarly aged workers in less-exposed occupations.

The adjustment appears primarily through reduced hiring rather than increased firing.

But whether junior employment rises or falls is not the central business question.

The important change is not whether the first rung disappears. It is whether climbing the first rung still develops what the second rung requires.

A business can retain junior employees and still weaken its development pathway if AI removes the experiences through which they previously learned. Equally, a business can automate substantial junior work and improve development if it deliberately creates better exposure, decision practice and feedback.

That is why this is a work-design problem rather than simply a workforce problem.

You can see the gap when junior employees produce stronger work faster but experienced managers still need to validate the reasoning whenever a situation departs from the standard case.

If the output has advanced further than the judgment behind it, the development system needs redesign—not the old workload restored.

It is easy to see a junior employee produce a stronger analysis with AI and assume development has accelerated with it.

Then a slightly unusual case arrives, the assumptions change, and the decision goes straight back to the experienced manager. The mistake is measuring development through the quality of the finished work instead of noticing that the decision boundary never moved.

Once you see that distinction, you stop asking whether people can produce senior-looking work and start asking whether they are becoming ready for senior responsibility.

The Work Produced More Than the Output

Every task produces an obvious output. Some tasks also produce capability, and conventional automation analysis rarely accounts for that second output.

Consider a salesperson preparing proposals.

The visible product is the proposal. But repeated exposure also teaches the salesperson how customer priorities differ, where margin becomes vulnerable, which objections matter and why apparently similar opportunities require different commercial decisions.

After enough exposure, something changes.

They notice things they were never explicitly taught.

That is the hidden value inside some junior work.

There is production value: what completing the work creates for the business now.

And there is development value: what repeatedly encountering the work develops in the person for later.

Not every task has meaningful development value. Copying information between systems may teach almost nothing after the first few repetitions. Automate it.

Other repetitive work contains meaningful variation. Reviewing lost sales reveals buying behaviour. Preparing estimates exposes where margin disappears. Resolving exceptions reveals where documented processes and operating reality diverge.

The repetition can look inefficient from outside while still producing something valuable inside the person doing it.

That does not mean employees must manually perform hundreds of tasks to earn experience. That would confuse the old mechanism with the capability it happened to create.

The better question is:

If we remove this work, what else are we removing with it?

That question belongs before automation, not six months later when managers discover that employees can operate the new system but struggle when the system encounters something unusual.

Look for this in your business: take one activity AI now performs or substantially prepares and ask what employees used to notice while doing it that they no longer see.

If salespeople no longer read original enquiries, marketers no longer encounter raw customer comments or estimators no longer see the unusual specifications buried inside straightforward jobs, the business may have removed more than labour.

The saving can still be worthwhile.

But now you know what needs redesigning.

AI Separates Productivity From Capability Development

AI allows businesses to obtain sophisticated outputs without requiring the person producing them to possess all the capability historically needed to create them.

That changes how managers should interpret work.

Historically, producing a competent market analysis suggested someone understood something about market analysis. Building a strong proposal required at least some knowledge of how proposals worked.

The relationship was imperfect, but the output provided evidence of capability.

AI weakens that signal.

Someone can now produce work substantially above their unaided capability. That is one of AI’s benefits. It expands what people can accomplish.

But it creates a management problem:

Visible output becomes a less reliable signal of invisible capability.

A polished analysis does not prove someone can identify the wrong assumption underneath it. A professional proposal does not prove commercial judgment. A persuasive customer response does not prove someone recognises when the standard response is inappropriate.

Microsoft research involving 319 knowledge workers and 936 real-world examples found that higher confidence in generative AI was associated with less critical-thinking activity.

It also found that the nature of critical thinking changes when AI is involved, shifting toward activities such as verification, integration and task stewardship.

Employees may spend less time constructing the first answer and more time deciding whether the answer deserves to be trusted.

That changes what managers need to observe.

The finished artefact tells you what the employee and AI were able to produce together. It tells you much less about whether the employee recognised a weak assumption, noticed missing evidence, understood the trade-off or would know when the same answer should not be used again.

The development question therefore becomes:

Output capability → Judgment capability → Decision authority

These three things should not be assumed to move together.

If AI dramatically improves someone’s output but six months later they still escalate exactly the same pricing decisions, customer exceptions and commercial judgments, their production capability has moved further than their decision capability.

That distinction matters because businesses do not scale merely by producing more sophisticated work.

They scale when more people can be trusted with meaningful responsibility.

Don’t Preserve Junior Work—Redesign How People Learn

The goal is not to protect old tasks. It is to preserve the capability-building mechanisms that still matter after those tasks change.

This is where businesses can overcorrect.

Once leaders recognise that junior work helped develop experience, they may conclude that some routine work should remain manual because “people need to learn.”

That is the wrong lesson.

If AI can reliably remove low-value administration, repetitive drafting or mechanical analysis, making employees continue doing it for tradition’s sake preserves labour, not development.

The old apprenticeship model was not necessarily well designed.

It was convenient.

Necessary work supplied cases, repetition and feedback. Learning happened alongside production without anyone deliberately engineering it.

AI creates the opportunity—and requirement—to separate the two.

Suppose AI can assemble a customer account history, identify changes in purchasing, summarise support issues and prepare an initial account analysis.

Keep the saving.

Then make the developmental work more deliberate.

Have the employee identify which signals matter. Require a recommendation before revealing the senior view. Compare two superficially similar accounts that developed differently. Examine why the AI’s interpretation should or should not be trusted.

The employee does not need four hours of preparation merely to receive four hours of developmental value.

Don’t preserve the labour. Preserve the learning loop.

The elements that matter are exposure, interpretation, judgment, consequence and feedback.

And one common behaviour deserves challenging directly: giving junior employees AI and calling the resulting productivity “upskilling” is not a development strategy.

Better tools improve capability at the point of use. They do not automatically create judgment that remains when the answer is incomplete, ambiguous or wrong.

Look for this in your business: managers spend less time correcting basic work but remain the permanent quality-control layer for anything requiring judgment.

If that continues, AI has reduced workload below the manager without reducing dependency on the manager.

Decide What AI Should Do and What People Still Need to Experience

The boundary between human and AI work should reflect not only what AI can perform, but what people still need to encounter to develop future judgment.

“Can AI do this?” is a technical question.

It is not enough to design the work.

Consider customer complaints. AI may classify them, identify recurring themes, summarise conversations and prepare responses. There is little value in requiring someone to manually categorise 500 complaints simply because previous employees did.

But a future customer leader who never encounters actual complaints may miss something important.

Summaries reveal recurring patterns. Direct exposure can reveal emotional intensity, ambiguity and the unusual case that does not fit the categories.

The same principle applies elsewhere.

AI can prepare sales proposals, but salespeople still need exposure to why customers hesitate.

AI can analyse marketing performance, but marketers still need to understand why customers notice, believe or ignore particular messages.

AI can identify operational anomalies, but managers still need experience distinguishing genuine failure from harmless variation.

This exposes a less obvious constraint.

As routine work becomes abundant and cheap, meaningful opportunities to form judgment may become scarce.

There are only so many difficult negotiations, unusual customer failures, pricing exceptions and ambiguous operating decisions inside a business.

If senior people continue handling all of them because AI handles everything routine below them, the richest learning opportunities become even more concentrated at the top.

You can see this without measuring anything.

AI handles the ordinary enquiries, straightforward quotes and standard cases. Then watch where everything unusual goes.

If it still travels directly to the owner or the same two experienced managers, the business has automated the work around its capability constraint rather than changing the constraint itself.

That arrangement may be operationally safe today.

Left unchanged, it also means the people who already possess the most judgment continue receiving the experiences from which more judgment is developed.

Redesign Junior Roles Around Exposure, Judgment and Feedback

A redesigned junior role should be defined by the capability the person is expected to develop, not by whatever basic work remains after AI.

Start with the changed state.

What should this person be able to recognise, judge and decide in 12 months that they cannot reliably handle today?

Then work backwards.

Someone expected to price non-standard work needs exposure to different pricing situations, visibility into margin consequences, opportunities to make recommendations and feedback on whether the reasoning was sound.

A marketer expected to develop strategic judgment needs exposure to customer evidence, positioning choices, campaign results and failed assumptions—not simply more AI-assisted content production.

An operations manager develops by diagnosing why reality differs from the process, not by producing more reports about the process.

But exposure alone is not enough.

Experience is not accumulated repetition. It is accumulated variation interpreted through feedback.

Someone can perform the same activity hundreds of times and develop little additional judgment if nothing forces them to distinguish one situation from another, make a decision and learn from what happened next.

That distinction matters because it allows businesses to remove enormous amounts of repetitive work without necessarily removing valuable experience.

The development mechanism becomes:

Meaningful variation → interpretation → judgment → action → consequence → feedback

AI can strengthen this loop. It can surface relevant cases, retrieve precedent, challenge assumptions and make feedback faster.

That means AI does not have to weaken apprenticeship. Properly designed, it may make capability development more deliberate than the old model ever was.

But there is a critical difference between reviewing someone’s output and developing someone’s judgment.

If a manager simply fixes the answer, the artefact improves.

If the employee makes a recommendation, explains the reasoning, sees the consequence and receives feedback, the person has a chance to improve.

This changes what businesses should protect when AI redesigns a role. The scarce developmental asset is not a volume of junior tasks. It is access to sufficiently varied situations where judgment can be exercised safely and its consequences understood.

Look for this in your business: identify what someone can now decide independently that they could not decide six months ago.

If AI has made them dramatically faster but the boundary of independent responsibility has not moved, do not mistake increased production capacity for deeper capability.

The objective is not simply more productive juniors.

It is a business capable of producing its next generation of trusted decision-makers.

Imagine an estimator who once spent hours assembling straightforward quotes before a senior reviewed the difficult ones.

After AI took over much of the preparation, the business resisted filling the recovered time with more quotes and instead let the estimator analyse selected exceptions, recommend a price and compare the reasoning with the senior decision.

The work became faster, but more importantly, the employee began handling situations that previously travelled upward.

The role stopped being measured by how much preparation was completed and started developing someone trusted to make increasingly valuable decisions.

Measure the Productivity Gain Without Creating a Capability Gap

An AI initiative is incomplete if it measures what disappeared from the workload but not what changed in the business.

Hours saved matter.

But they are an intermediate measure.

If proposal preparation falls from two hours to 30 minutes, ask what happened to the released capacity.

Did response time improve? Did throughput rise? Did conversion improve? Did margin increase? Did senior employees regain capacity for higher-value work?

Then ask the second question:

What happened to capability?

Not through a vague training score.

Measure observable movement.

Are employees identifying exceptions more accurately? Are they making decisions previously escalated upward? Are senior reviewers finding fewer reasoning errors? Can employees explain why an AI recommendation should be accepted or rejected?

Most importantly:

Has decision authority moved?

Imagine AI reduces proposal preparation by 70%. Six months later, proposal volume has increased and response times have fallen.

That is value.

Now imagine the sales manager still handles every unusual pricing decision because nobody below them has developed sufficient commercial judgment.

The initiative worked.

It also left an important constraint untouched.

Possibly more exposed than before.

Because there is a second-order effect: AI can increase execution capacity faster than a business increases decision capacity.

More enquiries can be researched. More proposals can be prepared. More customer situations can be analysed. More work moves through the system.

If judgment remains concentrated in the same owner or senior managers, that additional execution capacity sends more work toward the same scarce decision points.

The report gets produced faster. The approval queue doesn’t move.

The constraint has moved from producing the work to deciding what happens next.

Production capacity ↑ → work reaches decisions faster → judgment remains concentrated → management becomes the constraint → economic value plateaus

That is why capability development is not merely an HR consideration attached to AI implementation. It affects how much AI-created productivity the operating model can actually convert into growth, margin and throughput.

Faster output alongside unchanged approvals, escalations and senior dependencies is therefore not evidence that AI failed.

It is evidence that productivity and capability are different systems and only one has been redesigned.

If execution keeps accelerating while judgment remains fixed, the business can become faster without becoming more scalable.

The real return from AI is not maximum automation.

It is improved business performance accompanied by a wider distribution of the judgment required to sustain it.

Six months after an AI rollout, the dashboards can look excellent: faster reports, quicker proposals, more output per employee.

Yet watch what happens when something unusual occurs.

If the same three experienced people still receive every difficult customer, pricing exception and ambiguous decision, the organisation may have increased productive capacity without increasing decision capacity.

The strongest AI-enabled business will not simply produce more work—it will produce more people capable of carrying responsibility.

Conclusion

AI doing junior work is not automatically a problem.

Forcing people to continue doing work a machine can perform better is not a strategy for developing talent.

But neither is removing that work and assuming expertise will somehow appear later.

The old system bundled two things together: getting work done and learning how the business works.

Junior employees researched, drafted, reconciled, investigated and prepared because those activities were economically necessary. While doing them, they encountered the variation from which experience emerged.

AI is loosening that connection.

That gives businesses enormous freedom.

It also creates a design responsibility.

You can remove administration without removing customer exposure. You can automate preparation while increasing opportunities to exercise judgment.

You can use AI to shorten the distance between an employee making a decision and learning whether the reasoning was sound.

But only if capability development becomes part of the redesign.

Otherwise, the outcome can be deceptively comfortable.

Outputs improve. Work gets faster. Junior employees appear more capable because the artefacts they produce become more sophisticated. Senior people quietly continue handling the difficult decisions.

Then, eventually, the business discovers that the people who know how to judge the work are still the same people who knew how to judge it before AI arrived.

That is faster production sitting on top of the same capability constraint.

The better path is not complicated to state:

Automate what no longer deserves human effort. Preserve or recreate the experiences that build judgment. Then measure whether responsibility moves as capability develops.

The choice is not between protecting junior work and embracing AI.

It is between allowing yesterday’s apprenticeship system to disappear accidentally and designing tomorrow’s capability system deliberately.

When AI removes the work, ask what else the work was producing.

What replaces it should not be accidental.

Action Steps

Identify work doing two jobs

Review the activities AI is already performing or substantially preparing and identify which previously created both an output and meaningful employee exposure.

Task-level ROI misses capability that developed as a by-product of productive work. Automate freely where development value is negligible; investigate further where it is material.

Define the capability that must survive the automation

For work with development value, specify what employees were actually learning: recognising exceptions, understanding customers, evaluating trade-offs or exercising commercial judgment.

Preserving a task without defining its developmental function preserves labour rather than capability. Retain the required learning mechanism, not automatically the old workflow.

Separate output capability from judgment capability

Assess whether AI-assisted employees can explain, challenge and independently evaluate the work they now produce faster.

Better output can conceal unchanged underlying judgment. Do not expand decision authority solely because AI has improved the quality or speed of someone’s deliverables.

Protect access to meaningful exceptions

Track where unusual customer, pricing, operational and commercial situations now go after routine work is automated.

Exceptions are often where judgment develops fastest. If every meaningful exception still travels directly to senior management, deliberately redistribute controlled exposure rather than concentrating learning at the top.

Build feedback around decisions, not corrections

Have developing employees form a recommendation before receiving the senior or AI-supported answer, then compare the reasoning and consequence.

Correcting finished output improves the artefact; feedback on reasoning improves judgment. Redesign review around how decisions were reached, not simply whether the document was acceptable.

Measure whether responsibility moves

Track what employees can independently recognise and decide after the AI-enabled redesign.

Hours saved prove efficiency, not capability development. If output rises while escalations, approvals and senior dependencies remain unchanged, treat the capability architecture as unfinished.

FAQs

Bonus: Three Things AI Changes That Are Easy to Miss

The obvious AI conversation focuses on what work can disappear. That makes efficiency visible, but it can make the architecture underneath the work almost invisible.

The more useful perspective is to notice what becomes scarce after something else becomes abundant. AI makes production easier. That does not mean everything surrounding production becomes easier too.

Stop treating sophisticated output as evidence of sophisticated capability

This is what you are doing wrong: evaluating people’s development through work whose quality AI increasingly helps determine.

A strong proposal, analysis or recommendation once revealed more about the capability required to create it. That signal is weakening.

Managers increasingly need to evaluate the judgment behind the artefact, not simply the artefact.

If this doesn’t change: responsibility will remain concentrated with senior people because polished work never becomes trusted independent judgment.

The scarce resource may become opportunities to form judgment

Routine cases can increasingly be processed cheaply. The exceptions left behind become disproportionately valuable because they contain ambiguity, trade-offs and consequences—the conditions under which judgment develops.

The surprising implication is that meaningful exceptions should not automatically be protected from junior employees.

They may become some of the organisation’s most valuable developmental assets.

If this doesn’t change: your most experienced people will keep accumulating the experience your next generation needs.

Removing work can reveal what a role was really for

AI forces a useful confrontation. If 60% of a role can disappear without damaging the business, that 60% was never the role’s enduring purpose.

The opportunity is not to refill the employee’s week. It is to redefine the role around the capability and responsibility that now create greater value.

If this doesn’t change: AI will make old roles cheaper instead of helping you design better ones.

The interesting question is therefore no longer how much junior work AI can perform.

It is what kind of people your redesigned system will produce.

Other Articles

How to Prove Marketing Claims When AI Can Make Any Claim

How to Turn AI Time Savings Into Real Business Value

When AI Has the Answers, What Should Your Content Do?