Capability&Consequence

Essay

The Apprenticeship Problem

A firm can improve this year's project margin and weaken the supply of judgement it will need a decade from now.

Document
CAC-008
Issue
1.0
Published
July 28, 2026
Reading time
10 minutes
Review due
January 26, 2027

TL;DR

Junior professional work produces two things at once: a client deliverable and the experience from which future judgement is built. AI can reduce the labour required for the first while quietly removing the second. The immediate project captures the saving; the capability cost appears years later and somewhere else in the firm. Treat that gap as capability depreciation. Every material automation of junior work should identify the experience being removed, fund a replacement and measure whether the replacement produces judgement rather than merely training completion.

A consulting firm can improve this year's project margin and weaken its 2036 partnership at the same time.

The pressure is already visible. HBR has examined how AI changes both the structure and entry-level hiring model of professional-services firms. Recent working papers describe the resulting risk as a judgement vacuum, a pipeline floor and a hollow pyramid.

Different names, same structural problem: the firm automates the base while continuing to depend on the judgement produced through it.

The mechanism is simple.

Junior work produces two things.

It produces the client artifact: the research, model, interview synthesis, diligence finding, workpaper or first draft.

It also produces the practitioner.

People learn by forming a view, discovering where the data break, making mistakes, receiving edits, watching a client reject a technically correct answer and seeing a senior practitioner work through ambiguity.

AI can reduce the labour required for the first output.

It can also remove the experience that produced the second.

The billable apprenticeship

The traditional leverage model did more than move work from expensive people to cheaper ones. It embedded professional development inside client delivery.

Junior practitioners researched, modelled, drafted, checked and reconciled. The work advanced the engagement. The repetition and feedback developed the practitioner. The client paid for hours that produced both outputs.

Not every hour was valuable training. Repetitive work without challenge or feedback can teach very little. Research on expertise places more weight on the quality and variety of practice than on time served alone. A 2025 Annual Review of organisational expertise emphasises active practice, varied conditions and informative feedback.

But the old model contained a subsidy: some of the cost of creating future senior judgement sat inside billable production.

AI pulls the two outputs apart.

The artifact may now require fewer junior hours. The learning does not happen automatically somewhere else.

The margin arrives first

The immediate economic case is attractive.

Reduce junior production time. Keep quality stable. Complete the engagement faster or with fewer people. Improve margin.

The capability cost has a different owner and a different clock.

It appears later, when a manager cannot identify the weak source behind a fluent answer; when a senior reviewer has to reconstruct reasoning that the team never formed; or when the firm needs a new partner with judgement built through situations that no longer reach junior staff.

The project that created the saving may be closed years before the cost appears.

Call the gap capability depreciation.

It is the future capability a firm consumes when it removes developmental experience without funding a replacement.

The term matters because "training risk" sounds like an HR issue. Capability depreciation sits inside the economics of automation. The project captures a current saving by drawing down an asset the project P&L does not record.

The wrong question

Most automation programmes ask:

How much junior work can the model remove?

The better question is:

Which experiences are required to create the judgement we expect to sell later, and where will those experiences come from now?

That question changes the automation decision.

A task that is low-value as production can still contain a useful developmental exposure. A task that consumes hundreds of junior hours can still be poor training. The work should not be protected because it is old or billable. Its developmental function should be identified before it disappears.

The capability-depreciation ledger

Every material automation of junior work should add five lines to the business case.

1. Production removed

Name the work: initial research, model construction, document review, interview synthesis, drafting, reconciliation or scenario analysis.

Estimate the hours removed. Measure the checking and rework added. Use accepted output rather than first-draft speed.

2. Experience removed

Name what the practitioner used to encounter while doing the work:

  • Forming an independent problem frame
  • Finding and assessing primary sources
  • Discovering how data fail
  • Building and testing assumptions
  • Reconciling conflicting evidence
  • Explaining a position under challenge
  • Seeing how client context changes the answer
  • Receiving detailed feedback on an attempted solution

If none of those are present, the task may be a strong automation candidate with little capability cost.

3. Replacement funded

Decide how the experience will be recreated.

Protected first-pass analysis can require a junior to form a view before seeing the model's answer. Structured source audits can preserve evidence judgement. Historical cases and seeded errors can create repeated practice. Progressive client responsibility can expose practitioners to the organisational context no simulation fully captures.

There is useful learning science underneath some of these choices. Research on prediction as a learning strategy finds benefits when learners generate an answer before seeing the solution. But a generic curriculum is unlikely to replace the apprenticeship of every professional-services firm.

The replacement has to match the experience removed.

4. Senior time

Somebody has to observe, challenge and teach.

AI may reduce the number of junior production hours while increasing the concentration of senior attention required to develop each person. That time is becoming more valuable at exactly the moment the firm needs more of it.

"Partners should mentor more" is not an operating model.

The business case needs the hours, staffing ratios, incentives and budget.

Accenture reported US$1 billion of learning and development investment and approximately 47 million training hours in fiscal 2025. The figures do not tell us whether a particular method produced better judgement. They show the scale of explicit investment that capability building can require.

5. Evidence of capability

Training completion is not the outcome.

The firm needs to know whether people can:

  • Identify unsupported claims and weak sources
  • Build sound assumptions before senior correction
  • Detect errors in historical or seeded cases
  • Escalate appropriately under uncertainty
  • Perform across varied case types
  • Explain and defend reasoning with a client
  • Reduce rework as responsibility increases

Confidence scores and course attendance are easy to collect. The ability to catch a consequential mistake is the thing being built.

One task, redesigned

Take a market and competitor scan.

An agent can assemble the first pass quickly. The production case measures time saved, source coverage, error rates and revision effort.

The capability question is what the junior used to learn while doing it.

If the task required decisions about source quality, contradictory evidence and ambiguous market definitions, removing the first pass removes useful practice.

The redesigned workflow could require the junior to:

  • Define the question and source hierarchy before invoking the agent
  • Form an initial hypothesis
  • Audit a sample of citations against primary sources
  • Find evidence that contradicts the synthesis
  • Defend the final inclusion and exclusion decisions to a reviewer

The agent still removes production work. The junior still performs the acts most likely to build evidence judgement.

Whether that design works should be visible in subsequent performance, not assumed from the elegance of the exercise.

The governance dependency

Lawyers using AI retain duties of competence and supervision under ABA guidance. Auditors retain duties of due professional care and scepticism under PCAOB standards.

Management consulting does not carry the same formal professional regime. The operating problem is still relevant.

If a firm expects people to challenge, supervise and stand behind machine-assisted work, those people need enough expertise to know when the machine is wrong.

A firm that automates the pathway to that expertise creates a dependency on the stock of judgement it already has.

The stock can last for years.

That is what makes the problem easy to ignore.

The practical test

Before removing a junior task, ask:

If this work disappears, which experience disappears with it?

If the answer is "none", automate it.

If the answer includes evidence judgement, ambiguity, client context or progressive responsibility, put the replacement inside the investment case before taking the saving.

The apprenticeship problem is not a reason to preserve inefficient work.

It is a reason to stop booking the production benefit while leaving the capability cost to the future.

The decision this should change

Add capability depreciation to the business case for automating junior work. Record which developmental experiences disappear, how they will be replaced, who will provide the senior teaching time and which observed outcomes will show that judgement is still being produced.

What this adds

Prevailing consensus

AI will compress junior research, drafting and analysis, allowing professional-services firms to operate with fewer people or redirect them to higher-value work.

What this challenges

Some of that production work also funded apprenticeship. Removing it can improve current delivery economics while reducing the experience from which future judgement is built.

New contribution

Capability depreciation connects each automated task to the developmental exposure removed, the replacement investment required and the evidence that capability is still forming.

What would weaken the argument

Sustained improvement in judgement, promotion quality and client outcomes after material reductions in junior analytical work, without significant replacement training or senior teaching cost, would weaken the thesis.

Sources and references

  1. HBR - How AI Is Upending How Consulting Firms Hire Talent
  2. HBR - AI Is Changing the Structure of Consulting Firms
  3. SSRN - The Judgment Vacuum
  4. SSRN - The Shape of Professional Service Firms under Generative AI
  5. SSRN - The Hollow Pyramid in Legal Services
  6. Annual Review - Experts and Expertise in Organizations
  7. Psychonomic Bulletin and Review - Predicting as a learning strategy
  8. ABA - Formal Opinion 512 on generative AI tools
  9. PCAOB - AS 1015, Due Professional Care
  10. Accenture - Fiscal 2025 financial performance