This paper investigates whether axioms in economic theory can guide AI agents toward more consistent decision-making. We instruct a model to generate choices under varying budget constraints and apply revealed preference theory to automatically evaluate their internal consistency. Using these axiom-derived signals instead of human labels, we use the agent’s own output to fine-tune the model. We show that the resulting agent exhibits substantially higher choice consistency, with improvements that generalize well beyond the original training environment. To validate this approach, we also apply the procedure to a simulated supermarket environment calibrated with real scanner-data prices and household grocery budgets and find that fine-tuning yields highly consistent monthly category allocations. Our findings demonstrate that economic axioms, designed to provide normative benchmarks for human choices, can also serve as a powerful feedback mechanism to improve AI decision quality.