Published at: 2026-05-05T21:02:31+05:30
2026-03-13 — The future of work — The ladder is breaking; rebuild work around leverage, not titles
Thesis
Entry-level "learning jobs" are being compressed by generative AI and agents, not because companies no longer need junior talent, but because the work that used to justify junior roles is being modularized, automated, and supervised differently. The career ladder is not disappearing. It is being rebuilt into a different shape: fewer roles that are paid to produce, more roles that are paid to direct, and a growing premium on people who can turn ambiguous goals into reliable outputs with AI as a partner.
Context
For most of the last century, the white-collar career ladder was a machine for converting time into judgment. Juniors did the messy work. In doing it, they learned the business. Seniors reviewed it. Eventually, those juniors became seniors.
Generative AI changes the economics of that apprenticeship.
When a model can draft, summarize, transform formats, propose options, write code, and check itself against a rubric, the old division of labor (juniors produce, seniors decide) stops being the default.
At the macro level, forecasters disagree on how fast productivity will show up and whether it will translate into layoffs. But there is enough convergence on two points to take seriously:
A lot of tasks are exposed, especially in knowledge work.
The pressure hits entry-level first, because entry-level is where organizations historically bought cheap labor for routine cognitive tasks.
A CEO can claim that half of entry-level work may be wiped out in a few years, and even if that estimate is wrong, it points at a real structural vulnerability: the part of the labor market that depended on routine knowledge work as a training ground.
Some of the most useful evidence so far is not a grand theory. It is micro-level measurement.
In controlled settings, generative AI has produced large improvements in speed and quality for writing tasks, with the biggest gains among lower-performing participants.
If that pattern generalizes, the distributional impact matters as much as the average: the tool does not just make good people better. It pulls more people toward “good enough,” shrinking the economic value of the middle layers of work that used to be performed by juniors.
At the same time, institutions like the World Economic Forum project both displacement and creation, with a large net positive number of roles over the decade, and employers reporting that they expect to reduce headcount where tasks can be automated.
Those two things can be true simultaneously. Work can be abundant while specific job pathways collapse.
Key ideas
1. The first thing AI breaks is not jobs. It is job bundles
A “job” is a bundle of tasks. Organizations historically hired bundles because unbundling was expensive.
A junior analyst’s bundle might include:
Pulling data
Cleaning it
Making charts
Writing a first draft memo
Formatting slides
Taking meeting notes
Following up on action items
In the agent era, those tasks are separable. A team can buy the output of several of them as software.
This is why you will see seemingly contradictory signals:
Companies say they are hiring.
People say it is harder to find entry-level roles.
The job market is not uniformly shrinking. It is re-pricing bundles. Some bundles lose value. New bundles appear.
2. Apprenticeship gets more expensive when the work gets cheaper
The brutal truth about many entry-level roles is that the organization was not paying for learning. The organization was paying for output that happened to teach.
When AI makes that output cheap, the subsidy disappears.
This is consistent with evidence that entry-level hiring declines first in roles that are highly exposed to generative AI, and that regions with higher demand for AI skills have lower employment in AI-vulnerable occupations over time.
If the ladder used to be “do rote work until you earn discretion,” then automation removes the rungs.
Organizations can respond in three ways:
Cut the ladder and hope seniors can do the work.
Rebuild the ladder by funding training explicitly.
Outsource the ladder by recruiting only people who already did the training elsewhere.
Most firms will try option one first, then discover that seniors are too expensive to spend their time on foundational work.
3. The new entry-level is “operator,” not “assistant”
When a model drafts the memo, the entry-level role is no longer "write the memo." It becomes:
Build the prompt and the context
Define the acceptance criteria
Run the system through edge cases
Verify claims and citations
Decide what to escalate
Keep an audit trail
This is less like being an assistant and more like being an operator of a complex machine.
The operator’s value is not typing speed. It is reliability under uncertainty.
That is why the crucial skills shift:
From recall to judgment
From formatting to structure
From output to evaluation
You can see the seed of this in agent research that studies workflows: agents are not only doing tasks; they are re-creating a sequence of decisions, checkpoints, and handoffs.
In other words: the labor market is migrating from “who can do the task” to “who can run the process.”
4. Productivity will arrive as variance reduction
Most people imagine productivity as a linear speedup: do the same thing faster.
But one of AI’s most important effects is variance reduction:
Fewer empty documents
Fewer blank slates
Fewer “I don’t know what to write” stalls
More consistent baseline quality
In experiments on professional writing tasks, AI improved quality while reducing time, and the largest gains came from those who started out weaker.
If you are a manager, this changes what you can safely delegate.
You stop thinking, “Who is talented enough to do this?”
You start thinking, “Who can produce a dependable draft if they have the right tooling and rubric?”
Variance reduction is why the middle of the labor market feels pressure.
When the floor rises, the premium on merely being above average shrinks. The premium moves to:
Taste
Domain understanding
Accountability
Strategic selection of what matters
5. Work will re-center around accountability, not activity
When AI can generate activity, the obvious activity becomes suspect.
In the old model:
Activity implied effort.
Effort implied progress.
In the new model:
Activity can be faked.
Effort can be automated.
Progress must be measured.
This pushes organizations toward more explicit definitions of success:
What is the goal?
What is the metric?
What is the acceptable error rate?
Who signs off?
In procurement and compliance-heavy environments, that shift will be sharper. It is the same reason AI governance is increasingly treated as a supply chain and national security issue: organizations will care about who controls the system and what terms govern its use.
6. The biggest risk is not displacement. It is career deserts
Even if overall employment stays high, you can still create a society where early careers are unstable and delayed.
A career desert looks like this:
Few roles that train beginners
Higher credentialism
Longer unpaid or underpaid “portfolio building” periods
Greater reliance on networks
In that world, opportunity becomes less meritocratic.
The macro statistics can look fine while the lived experience of young professionals worsens.
This is one of the reasons why policy institutions emphasize skills and training as the critical adjustment mechanism.
But training is not only a policy problem. It is a design problem for firms.
Counterarguments
Counterargument: “AI will create more jobs than it destroys. The ladder will be fine.”
There is evidence pointing to net job creation over the decade and to broad workforce transformation rather than simple elimination.
That is encouraging. But net job creation does not guarantee stable pathways.
A labor market can add jobs and still fail beginners if:
The new jobs require experience.
The old jobs were the training pipeline.
The question is not “Will there be jobs?”
The question is “Will there be ladders?”
Counterargument: “If AI makes juniors productive, firms will hire more juniors.”
In some cases, yes.
If the junior-plus-AI combination can produce output that used to require a more expensive mid-level hire, then juniors can become attractive.
But the catch is supervision.
AI reduces drafting cost. It does not remove the need for:
Defining what “correct” means
Catching subtle errors
Aligning work to strategy
Owning the consequences
If the supervision burden stays high, firms will still prefer fewer, more experienced hires.
This is why the firms that hire more juniors will be the ones that:
Build strong rubrics
Use checklists and validation steps
Standardize processes
Treat training as a real product
Counterargument: “AI is overhyped; productivity won’t show up.”
It is true that macro productivity statistics can lag micro-level gains.
Adoption takes time. Organizations resist redesign. Many pilots fail.
But it would be a mistake to conclude that nothing is changing.
Even modest productivity gains, combined with task unbundling, can still reshape roles.
The ladder can break without a visible GDP explosion.
Takeaways
The unit of change is the task bundle, not the job.
Entry-level roles are vulnerable because routine cognitive tasks were the hidden subsidy for learning.
The new entry-level role is closer to “operator”: define, run, verify, and escalate.
AI’s most important near-term productivity effect may be variance reduction, not pure speed.
Firms will pay more for taste, accountability, and process design than for raw output.
The biggest social risk is not unemployment; it is career deserts and delayed adulthood.
If you are early-career, your edge is the ability to produce reliable outcomes with AI while being honest about uncertainty.
If you are a manager, your edge is designing workflows that turn AI into leverage without destroying the training pipeline.
Sources