Beyond Predictive Accuracy: A Multidimensional Protocol for Business-Aware AI Recommendation Evaluation

Predictive accuracy does not necessarily capture the economic or governance quality of AI-generated business recommendations. This exploratory study specifies a business-aware evaluation protocol covering incremental utility, downside risk, recommendation stability, and hard feasibility constraints, with a provisional composite index (BARQI). Previously executed repeated-split simulations in inventory, staffing, and manufacturing settings compare prediction-error selection, direct-utility selection, and risk-adjusted approaches. Across scenarios, downstream utility selection often improves on RMSE-based selection, while BARQI does not consistently outperform direct utility. Separately, eight constructed CMS-inspired pricing-transparency rule fixtures were classified as expected; this is a software unit-test result, not real regulatory compliance accuracy. A prospective validation protocol using the public UCI Online Retail II transaction dataset is specified, but raw data retrieval and empirical estimation have not been completed. The evidence supports transparent multidimensional reporting and falsifiable evaluation, not a claim that the proposed index is universally superior.