Training Is Part of the Cost of Oversight

An organization can clear its review queue and still be consuming the expertise that makes review possible. Experienced staff finish today's checks. Departures and lost qualifications are less visible. If replacements do not develop, a successful month can leave the next month harder to manage.

My argument is that an automation plan should account for two obligations: checking current work and maintaining the capacity to check future work. A training promise that ignores the time of qualified mentors can commit the same hour twice. A productivity target that spends every available hour on output can leave replacement learning unfunded.

This is an institutional choice to investigate, not evidence that every AI deployment destroys expertise. The Institute's bounded research identifies conditions under which protected learning and present verification can coexist. It also identifies ways the constraint can loosen. [1]

One budget, two obligations

Start with a defined task and an independent standard for performing it. A qualified reviewer must demonstrate the relevant capability; a job title, a course completion or a polished AI-assisted answer cannot substitute for that demonstration. Valid tools can be part of the assessment, provided checking the answer does not depend on trusting the same unverified answer.

The accounting is ordinary: next period's qualified capacity equals retained capacity plus newly formed qualifications and qualified inward arrivals. Keep those additions separate. Hiring someone already qualified can solve one firm's shortage while transferring it to another. It does not, by itself, replenish the profession.

The difficult choice appears when existing qualified people also mentor successors. Reserving their time for learning reduces the minutes available for current review. Reserving too little may fail to replace losses. Both commitments therefore belong in the staffing decision. Calling training important after allocating all qualified time to production does not make the commitments compatible.

That is the extension beyond warning about deskilling or overwork. Today's adequate staffing can depend on qualifications accumulated under earlier working arrangements. An organization needs to ask whether its new arrangements reproduce the capabilities it still requires, as well as whether its current staff can handle the queue.

An invented comparison

Consider a deliberately synthetic organization with 100 qualified reviewer-equivalents. Each has 300 productive minutes per workday. It receives 8,000 cases a day, each requiring three minutes under an unchanged review protocol. Assume two reviewer-equivalents are lost each month. These are invented inputs, not measured professional rates. [1]

With no mentor time reserved, aggregate time could cover 10,000 cases a day. Today's 8,000 fit, but the next month starts with 98 qualified equivalents if nobody qualifies or enters.

Now assume an established training pipeline can replace those two losses when mentors reserve a tenth of their productive time. Current review-time capacity falls to 9,000 cases a day, still above demand, and the opening qualified stock is maintained next month.

Reserve a quarter of qualified productive time for mentoring instead, under the same assumed linear training relationship, and formation rises to five equivalents. But current review-time capacity falls to 7,500 cases. Future capacity grows while today's requirement is missed.

The comparison establishes a possibility: some reservations can cover both obligations, while others cannot. It supplies no universal training percentage. It assumes trainees already in progress, sufficient candidates, comparable qualifications and a training response that may fail in practice. Novices do not become experts on a monthly timetable merely because the accounting uses months. Even sufficient aggregate minutes do not guarantee timely, accurate review when arrivals bunch, specialties differ or rework grows.

Assistance can change the constraint

The strongest objection is that the model charges training against a scarce resource that better technology might release. AI can help people learn. It can reduce the work requiring review. A reliable external check can make an old human skill unnecessary.

That objection matters. If successors can demonstrate the necessary competence while using less mentor time, the trade-off improves. If independently required review work falls faster than qualified capacity, coverage can improve even while fewer people retain the old skill. Preserving obsolete work would be a poor definition of institutional resilience.

The retained studies show why learning design deserves attention, while leaving the professional replacement question open. Shen and Tamkin randomized AI access for developers learning the Trio programming library and assessed immediate unaided understanding. AI access impaired learning on that immediate evaluation. Their qualitative interaction patterns do not establish the causal effect of a particular mentoring policy, and the short study does not measure the formation or attrition of professional verifiers. [2]

Bastani and colleagues randomized high-school mathematics classes to a standard AI interface, a guarded tutor or no AI. Guardrails mitigated the unassisted-exam harm found with the standard interface; the guarded tutor did not establish a positive unassisted-exam effect. This is evidence about learning design in that setting, not an estimated supply of future engineers or reviewers. The two studies cannot fill the model's professional formation rates. [3]

Make the commitment testable

A useful institutional commitment would specify the capability being maintained, its continuing purpose and the resources reserved to develop it. Count people as qualified only after an appropriate independent assessment, including delayed assessment where retention matters. Record losses, internal formation and already qualified arrivals separately. Track whether the work still requires the original capability.

There are costs outside the illustration: trainees' time, training delays, specialist supervision and uncertainty about who will qualify. Money for courses alone does not settle those costs. Nor does a mentor-time reservation establish that learning occurred.

The decisive missing evidence is an attributable record following those quantities together. Preserved independent competence and adequate review under assisted learning would challenge the depletion account. Formation persistently below losses, while the capability remains necessary, would support it. A dependable replacement for the capability would change the question altogether.

This is ordinary human-capital accounting applied to a particular dependency, not a new economic law. Its practical force is that a plan can appear efficient by spending a resource it does not replenish. Where verification remains necessary, training successors belongs inside the cost of sustaining oversight. Whether an organization can afford both obligations must be demonstrated in its allocation and outcomes, not assumed from today's completed queue.