Compression

A junior analyst opens a blank document. Ten minutes later, AI has produced a respectable market summary, the outline of a model and a first recommendation.

The language is crisp. The headings are in the right place. It looks like work that once took several days and at least one mildly depressing evening.

Management sees speed. The junior feels ten feet tall. The senior receives something polished enough to review quickly.

The document contains very little visible evidence of what the junior understands.

I call this the compression of expertise. AI compresses the time and visible skill needed to produce plausible expert work. The judgment behind that work still grows through context, mistakes, feedback and responsibility.

A junior can now produce a senior-looking draft before they can recognise a senior-level mistake.

Assume AI can produce 70 or 80 per cent of the visible first pass of many tasks. The exact share will vary by profession and by task. The structural issue begins well before that number.

When the surface arrives almost finished, we lose one of the signals we used to judge competence. We may also remove the work through which competence was built.

The career ladder was a training system

Knowledge-work careers were built on a fairly simple bargain.

Juniors gathered information, cleaned data, read contracts, built models, wrote first drafts and made a large number of small mistakes. Seniors checked the work, corrected the mistakes and gradually handed over more responsibility.

Some of this was useful practice. Some of it was formatting 70 PowerPoint pages at one in the morning because somebody three levels above had discovered a new shade of blue. We should have no nostalgia for that.

The useful work created repetition and exposure.

An analyst saw enough companies to notice when the numbers felt wrong. A lawyer read enough clauses to understand how one innocent sentence could move a surprising amount of risk. A developer debugged enough code to recognise the smell of a problem before finding its source.

The task produced two things: today’s output and tomorrow’s judgment.

Companies know how to put a price on the output. The judgment was hidden inside the process, so nobody had to budget for it separately.

AI changes that calculation. A company may no longer need to pay a junior for three days to produce a first version. Fair enough. The training function remains, now with no obvious budget or owner.

Remove enough junior work without rebuilding the learning around it, and the career ladder starts losing its lower rungs.

The company may hire fewer graduates and ask experienced employees to supervise AI instead. The quarterly numbers could look excellent. Several years later, the same company may discover that fewer people are ready to become experienced employees.

One company can hire people trained elsewhere. An entire profession eventually runs out of elsewhere.

The first 80 per cent is a misleading idea

Work rarely moves along a neat conveyor belt from easy to difficult.

Experienced judgment appears before the first draft, while the draft is being produced and after it arrives. Someone has to define the problem, set the boundaries, choose the evidence, challenge the assumptions, integrate the answer and decide whether it is safe to use.

A job is also a bundle of tasks, relationships, exceptions and responsibility. Automating 80 per cent of a first pass tells us very little about automating 80 per cent of a job.

The remaining work can also carry a disproportionate amount of risk.

An AI-generated contract clause can read beautifully and still be wrong for the client. Code can pass every available test while violating an assumption nobody thought to test. A market model can calculate flawlessly using two incompatible definitions of the market.

The work looks complete because the formatting is complete.

That is where the gap between assisted performance and independent competence becomes dangerous. Someone may be able to produce good work with AI while lacking the mental model needed to recognise when the AI has gone off the road.

Polish has therefore become a weaker signal of competence.

Review the thinking, not only the document

Companies often assessed junior employees through their output. A good memo suggested a good analyst. Clean code suggested a good developer. A persuasive deck suggested a consultant who might soon be allowed near the client.

That proxy now tells us less.

With rough work, reviewers can often see the path. A polished AI draft can hide which parts were understood, assumed, borrowed or invented.

Managers will need to spend less time correcting sentences and more time examining reasoning. Ask the junior:

• Why did you define the problem this way?
• Which assumption matters most?
• What did the AI produce, and what did you change?
• Which source did you verify?
• What evidence would change your recommendation?
• Where could this fail without anybody noticing?
• At what point would you escalate the decision?

Someone who can answer those questions has used AI as part of the work. Someone who cannot has allowed the output to outrun their expertise.

This also changes hiring. A beautiful take-home assignment proves less when anyone can generate one over breakfast. Live cases, oral defences and follow-up questions become more useful because they reveal how a person thinks, responds to uncertainty and changes their position.

Companies must design those assessments carefully. Otherwise, they will replace one weak proxy with confidence, performance and a nice accent. We already have enough executives selected that way.

Apprenticeship must become deliberate

Traditional apprenticeship often happened by accident. The junior attempted the work, the senior corrected it and both moved on to the next deadline.

AI removes some of that natural repetition. Companies now need to design the learning loop on purpose.

A good junior role should cover a complete small problem. Let the junior receive the messy brief, identify the missing information, direct the AI, verify the sources, make a recommendation and see what happened afterwards.

Give them bounded consequences. They can begin by observing, then recommend, decide on reversible matters, own a small outcome and gradually take on larger ones.

Before opening the AI answer, ask them to write down their own framing and prediction. Afterwards, ask them to compare the two. Sometimes the AI will be better. That teaches humility. Sometimes it will be confidently wrong. That lesson is useful too, provided the mistake gets caught before the client does.

Senior people must also make their reasoning visible. A silent correction produces a better document and teaches the junior almost nothing.

Why did the senior reject the first answer? What made them suspicious? Which trade-off did they accept? What information came from experience rather than the source material?

These conversations take time. Companies should count that time as part of the senior employee’s job and promotion case. Mentoring cannot survive as unpaid organisational charity squeezed between client calls.

AI can make apprenticeship better. It can provide immediate explanations, generate practice cases and let young employees explore areas beyond the narrow work their manager happens to assign. Its value as a teacher depends on a real feedback loop, where someone checks the reasoning and the junior sees the consequences.

Ten generated drafts do not equal five times the learning.

Junior roles will become broader and more demanding

A good junior employee can now contribute sooner.

A new developer can inspect an unfamiliar codebase and prepare a useful first change. A young lawyer can compare clauses before speaking to a supervisor. An analyst can test several ways of structuring a problem rather than spending the whole afternoon moving numbers between spreadsheets.

That is a genuine opportunity. Curious people can reach interesting problems earlier and develop a wider view of the work.

There is also a catch. Employers may expect judgment sooner while offering fewer chances to acquire it. That is a lousy bargain.

Junior employees need access to the full context behind a task. Turning them into prompt operators will produce people who can polish an answer without knowing whether the answer belongs in the room.

Prompt speed will matter less than intellectual discipline. Strong juniors will form their own view, use AI to attack it, trace important claims back to evidence, show uncertainty and ask for help before confidence turns into damage.

“The system said so” is not a professional standard. It is the digital version of blaming the intern, except the intern now writes excellent prose.

Seniority moves towards responsibility

Seniority should increasingly describe the amount of ambiguity, consequence and responsibility someone can carry without close supervision.

A senior person can define the problem before asking for an answer. They know what evidence is sufficient and which missing fact could overturn the entire recommendation. They understand the organisation well enough to see that two technically correct answers may have very different consequences.

They know when to trust the output, when to investigate and when to stop the process altogether.

They also take responsibility afterwards.

These abilities have always mattered. AI makes several old signals of seniority cheaper: recall, specialist vocabulary, blank-page speed and polished presentation. A senior role built mainly around producing decks, summaries and first drafts now has a fairly obvious problem.

Promotion criteria need to move accordingly. Output volume will tell us less. Companies should look at the ambiguity someone can handle, the consequences they can carry, the errors they catch, the quality of their escalation and the way they respond when they are wrong.

Seniority should also include the ability to develop judgment in other people. A senior employee who produces excellent work while everybody around them remains dependent is a high-performing bottleneck.

Teaching belongs inside the definition of seniority and inside the senior employee’s workload.

Redesign the ladder while there is still time

Every organisation adopting AI should rewrite three things at the same time.

First, junior roles should provide bounded, end-to-end responsibility rather than disconnected fragments of production.

Second, apprenticeship needs its own time, budget and accountable managers. Some manual practice should remain where a plausible error could cause material damage.

Third, promotion should reflect increasing ambiguity, consequence, ownership and the ability to develop other people.

AI gives us the chance to remove waste and give young people better work earlier. We should take it. Companies still need to replace the learning that was hidden inside the work they automate.

For every junior task removed, ask what the task used to teach and where that learning will happen now.

If the answer is “on the job”, point to the exact part of the job that still teaches it.