Employers often want interns who already have experience, even though an internship is supposed to be where experience begins. AI turns this old paradox into an organisational problem.

If AI can finish in five minutes what once took a junior employee two days, making the person repeat the old process makes little sense. But if the process disappears completely, where will that person learn to spot anomalies, judge quality and carry responsibility for an outcome?

CHAPTER 01AI is removing the old steps. We need to know what to keep

Many careers used to develop in a simple way. Juniors started with basic work, moved slowly, made mistakes, were questioned and had to revise. Over time, they learned what looked normal, what deserved doubt and when something felt off. Senior people often cannot fully explain why they notice a problem so quickly; that instinct usually comes from accumulated experience.

AI is now taking over many of those early steps: searching, organising, data entry, first drafts, formatting and basic analysis.

Consider a junior analyst collecting data on 100 hotels. There is little reason to preserve the copying, pasting and formatting. Yet while looking at 100 hotels, the analyst may also begin to notice why two “luxury” hotels command very different rates, why certain brands cluster in one city, or why a number can technically match the source and still feel implausible.

Those discoveries are not skills in copying and pasting, but they may once have been encountered during the process.

So the real design task is deciding what happens to friction. Mechanical work should disappear, necessary process should be compressed, and experiences that force people to compare, judge, take responsibility and receive feedback need to be rebuilt elsewhere.

CHAPTER 02As generation gets cheaper, judgment becomes the bottleneck

This becomes concrete very quickly inside companies. A junior employee who once produced one piece of work a day may now generate ten. Output scales. Judgment does not.

If the junior cannot evaluate the work, someone else still has to ask: Is the data credible? Can this logic go to a client? Is the conclusion too loose? What does this chart actually show?

Generation is getting faster. Judgment does not scale automatically.

Such an organisation has merely shifted the bottleneck from execution to judgment. AI produces, junior staff pass work along, and senior people absorb more and more review.

What companies need are people who can independently handle most ordinary decisions and escalate only the difficult ones. In the past, that capability often grew alongside operational fluency. In the future, it will depend more on whether someone can judge an output, spot what is wrong, explain why, and know which decisions they can own and which should be escalated.

CHAPTER 03Use the time saved by AI to increase “experience density”

There is a risk here. Once we admit that some friction helps people develop, it becomes easy to give unnecessary hardship a new educational story.

Three hours of manual formatting does not make a junior consultant better at consulting. Clicking the same button five hundred times does not produce taste. Much of the old pain existed because the tools were poor. We do not need to preserve it.

What matters are experiences that change judgment. If a junior analyst once spent eight hours on a report and seven of those hours were spent moving information around, AI can remove the seven. The saved time can be spent comparing more cases, explaining anomalies, questioning a client, making a controlled decision and seeing what happens.

I think of this as “experience density”: how many meaningful cycles of comparison, judgment, feedback and correction a person goes through in the same amount of time.

CHAPTER 04Perhaps we need an apprenticeship model for the AI era

The valuable part of apprenticeship is repeated comparison inside real work: what did I think, how did an expert judge it, and what happened afterwards?

That loop can survive even if the form changes. A junior does not need to manually process 100 cases; they might study 10 deeply, let AI process 90, and then search for differences and anomalies. They do not need to begin every task from a blank page; AI can propose several options, while the junior has to explain which one is better, where each fails and why.

Senior people also need to record more of the judgment they carry but rarely explain. Instead of saying only “this slide feels wrong”, they can capture criteria, counterexamples, before-and-after revisions and the trade-offs that shaped the decision.

One thing still cannot disappear: real responsibility. Juniors need a space where the cost of mistakes is controlled but the consequences are real. Judgment rarely matures through observation alone. It grows when people make a choice and then see the distance between that choice and reality.

CHAPTER 05Education faces the same question

In education, AI first changes the cost of producing a finished result. A 3,000-word assignment or a complete proposal can now become polished much more quickly. The final product may look stronger without telling us how much thinking the student actually did.

A research methods course has always been meant to train the harder parts: why a question is worth studying, why variables are defined in a certain way, whether a sample represents the target population, whether statistical significance matters in practice, and whether a conclusion can withstand challenge. AI does not make these skills newly important. It simply makes the distance between “submitting a complete answer” and “learning how to research” easier to see.

Teaching and assessment may therefore need to make more of the process visible: what evidence a student used, why they accepted or rejected an AI suggestion, where they changed their judgment, and whether they can explain their choices. The final product still matters, but it is increasingly weak evidence on its own.

That brings me back to the AI tools I help introduce at work. Saving time matters, but I also want to ask whether the removed process once carried some learning with it. If it did, that experience needs to be placed somewhere else: in comparison, explanation, feedback and a small amount of real responsibility.

Then the time AI saves has a chance to become experience instead of simply disappearing.