Automation is often evaluated by one measure: how much output it saves.

That measure is necessary. It is not enough.

Many forms of work produce more than their visible artifact. Writing a report develops judgment. Building a model creates an understanding of its assumptions. Researching a case teaches a professional where uncertainty lives. Drafting a proposal forces someone to decide what they believe.

When AI produces the artifact instantly, organisations need to ask what else the old effort was producing—and whether they intend to preserve it.

Output is not the whole product

Consider a junior analyst asked to prepare a market assessment. The visible output is a slide deck. But the work may also create:

  • Familiarity with the source data
  • Pattern recognition
  • Knowledge of relevant stakeholders
  • An intuition for uncertainty and exceptions
  • The ability to defend a conclusion in a meeting
  • A record of where the analyst’s judgment needs support

If AI produces the deck without involving the analyst in the reasoning, the organisation may save time today while weakening the learning process that produces capable senior analysts tomorrow.

The apprenticeship problem

This matters especially for entry-level work. Many professional careers are built through tasks that are repetitive, structured, and increasingly automatable:

  • First-pass research
  • Drafting routine documents
  • Spreadsheet preparation
  • Basic code implementation
  • Summarising cases
  • Reviewing standard records

Removing these tasks can create genuine productivity gains. But it can also remove the practice field through which people learned the deeper parts of the profession.

The relevant question is not whether juniors should continue doing work a machine can do. It is whether the organisation has designed a replacement path for acquiring judgment.

Keep humans in the learning loop

A good AI-enabled workflow does not require people to reproduce every machine-generated artifact manually. It does require deliberate opportunities to inspect, challenge, correct, and explain the work.

Useful patterns include:

  • Ask the employee to review the evidence before accepting the recommendation
  • Require short explanations of why an AI output is accepted or rejected
  • Rotate people through exception-handling queues
  • Preserve opportunities for supervised end-to-end ownership
  • Use AI to generate alternatives, not only answers
  • Measure learning and judgment development alongside throughput

The aim is not friction for its own sake. It is to retain the activities through which expertise forms.

A better division of labour

The most productive division of labour is often:

AI handles abundance.
Humans handle judgment, exception, context, relationship, and accountability.

But that slogan is incomplete. Humans cannot handle judgment well if they never encounter the evidence, patterns, and mistakes from which judgment is built.

The design challenge is therefore twofold:

  • Use AI to remove low-value repetition
  • Preserve enough engagement with the work that people continue to develop real capability

The practical test

Before automating a workflow, ask:

If we remove this task, what capability will a person no longer have the chance to build?

If the answer is “none,” automate aggressively. If the answer is central to future judgment, redesign the learning path before removing the work.

The future of work is not only about what AI can produce. It is also about what people must still practice in order to remain capable of standing behind decisions.

Adapted from the book Cheap Thinking: What AI Makes Abundant, What It Makes Scarce, and Who Captures the Difference by Daoyuan Li, PhD.