When Selective Quantum Machine Learning Fails: A Cost-Aware Quantum-Kernel Bandit Audit for Human–AI Allocation

Computation is a scarce input to decisions, so a more capable analytical tool need not be worth using. Building on economics Nobel laureates Herbert Simon's bounded rationality and George Stigler's economics of information, we ask when a Human–AI system should decline optional quantum analysis. Quantum Optimization and Policy Selection (Q-OPS) studies this question in synthetic task-mode allocation, separating a classical allocation, a rule for permitting further analysis, and a learner that chooses whether to request it. On 320 held-out instances, 221 satisfy the rule; the original online learners request quantum analysis on every eligible case and earn negative net sampling rewards at the declared resource price, while a hindsight oracle requests it on 55.7% and earns 4.245. Three exploratory diagnostics sharpen this failure: a classical score-based sampler exceeds both quantum samplers on every eligible case; removing the deeper option raises reward to 0.257 without producing selectivity; and predictors trained on both action outcomes yield selective, positive point estimates whose seed-block intervals include zero. Within this benchmark, the quantum kernel provides no robust advantage over its classical counterpart. The contribution is a reproducible distinction between a statistical sampling signal, the ability to select useful analysis, and benefit to the underlying decision; the present reward measures the first two, and the sampled states never alter an allocation. This distinction gives researchers and decision-makers a concrete way to question resource use before claiming social value. Small exact simulations motivate a sequenced research program in measured computation costs, downstream decision outcomes, and participatory organizational validation.