Expert routing and self-preferencing on AI platforms

This paper develops an industrial-organization theory of AI routing in which a dual-role platform allocates queries between its in-house model and an outside expert. The expert chooses an access price and initial quality, while early expert traffic raises later quality through learning-by-serving. The closed-form equilibrium has a difficulty cutoff in each period. A larger self-preferencing benefit raises the cutoff and reduces expert demand and quality investment; stronger data feedback raises expert investment but also increases the sensitivity of quality and traffic to self-preferencing. Relative to the conditional dynamic first best, the exact period-1 traffic shortfall is the sum of an access-markup wedge, a bias wedge, and a positive data-feedback wedge. Thus neutrality and marginal-cost access remove only the first two distortions. A practical implication is that governance should combine ownership-neutral routing criteria and cost-based access with logged, independently testable routing decisions and controlled external-model trials that preserve learning opportunities. The framework identifies what such audits must measure and clarifies the implementation limits created by stochastic performance, latent costs, and diminishing returns to data.