MBAs Considered Harmful?

Executive briefing

The finding. Organisations systematically discard their capacity to evaluate the claims their futures depend on, and they do it while every indicator they watch improves. The discarding is not stupidity and not malice. It is the rational output of a governance architecture — metric-based, comply-or-explain, claimant-evaluating — that prices deviation and makes domain judgment inadmissible as evidence. The capacity being discarded appears on no instrument, so it is free to cut and slow to miss: an organisation that has lost the ability to evaluate claims has also lost the ability to detect the loss.

The mechanism, in four parts. Expert objections are rationally discounted because the expert’s true objection and a self-interested incumbent’s objection arrive as the same sentence, while the evaluator’s own interests go unpriced. Claims that cannot be checked are flattened into single metrics that can mean anything. The metric becomes a comply-or-explain regime: compliance files itself, deviation requires a case argued to evaluators who cannot evaluate it. And there is no appeal — each layer above the room reads fewer of a claim’s properties, down to a single bit at the top. The condition this produces is not blindness. The experts can still read the situation with usable accuracy; they are simply not in the escalation path.

The canonical case. Intel declined to supply the iPhone on a cost forecast that was wrong on both axes, and sold its ARM division the year before — decisions that were individually defensible, collectively ruinous, and vindicated by five years of record results while the consequences incubated in a segment the instruments did not measure. The customers who could evaluate Intel’s claims directly — Amazon, Apple, Microsoft, Google — evaluated them and left, on a timeline visible from 2015. Intel’s data-centre revenue peaked in 2020 and fell off a cliff in 2022. The one channel that never files an objection is the one that ends the argument.

The current instance. The 2024–25 AI layoff wave ran the same machinery at economy scale: cuts filed under a technology story (by Gartner’s own survey, only one in five was genuinely about AI), capability discarded on metrics that improved, reversals now arriving — a third of US hiring managers have rehired for AI-eliminated roles — under new job titles that let organisations make the same decision twice without recording that the first was wrong.

Held honestly. Domain expertise is not a guarantee: Intel’s 10nm disaster was authored by experts with full evaluation capacity, and the telephone industry’s engineers defended a doomed category with flawless measurements. Evaluation capacity is not correctness. It is the ability to tell an evaluated claim from an unevaluable one — and to have the argument at all.

The implication. The credential is not the problem; the convention it normalised is: that the portable layer suffices and the room needs no one who can check the claims. An organisation that discards the domain model does not make its dashboards wrong. It makes them unfalsifiable — and in a regime where every deviation needs a case, the safest word available is no, which requires no domain expertise at all.