Research

Problem definition: Forecasts for key variables in a variety of operational decisions may be generated by human experts or by artificial intelligence (AI) technologies. A decision maker can benefit from the distinct advantages that each source may offer by providing AI assistance to the experts, allowing them to augment the information contained in the AI forecast by incorporating their own knowledge about the uncertain target variable. When multiple experts are available, accuracy can be further improved by utilizing the wisdom of crowds, forming a consensus by averaging each of their AI-assisted forecasts. However, the potential accuracy of a crowd of AI-assisted demand forecasters may be limited by two structural characteristics. First, because the AI assistance is valuable to each expert at an individual level, the opinion of the AI can end up being overrepresented in the crowd’s consensus. Second, the experts may fail to appropriately utilize the AI assistance when forming their forecasts by underemphasizing the information it provides.

Methodology/results: Using a stylized Bayesian model of information aggregation and a two-step forecast elicitation procedure, we develop a prescriptive method that can recover the most accurate consensus forecast given all information collectively observed by the AI and every expert in the crowd. This procedure works by pivoting the crowd's average AI-assisted forecast either toward or away from the crowd’s average initial forecast, elicited before the forecasters have observed the AI advice. We test the predictive performance of a data-driven pivoting method in three laboratory experiments. Across 3,188 participants and three AI accuracy treatments, we find that the pivoted forecasts match, and in many cases outperform, the AI-assisted crowd, the AI advice itself, and the unassisted crowd of forecasters.

Managerial implications: Firms can use the proposed forecasting method to improve human-AI collaborative forecasting accuracy by identifying and adjusting for the net bias in the crowd's AI-assisted forecast.

Data-driven solutions to the feature-based newsvendor problem are widely used when the underlying demand distribution is unknown. In such settings, interpretable policies that organize products into operationally meaningful categories are particularly appealing. We study a class of interpretable, tree-based prescriptive policies that can be learned via honest empirical risk minimization. We develop a regret decomposition that separates approximation, estimation, and learning effects, and use it to characterize how the optimal complexity of an interpretable policy should scale with available data. Our results show that the optimal degree of segmentation grows sublinearly with sample size, reflecting a fundamental tradeoff between customization and statistical reliability. These findings provide ex ante guidance on how much segmentation a manager should allow when designing a data-driven ordering policy. We further study the ex post reliability of leaf-level prescriptions after the policy has been deployed. Conditioning on the learned tree, we develop observable, leaf-specific measures of risk that distinguish between statistical instability arising from finite data and structural heterogeneity within a category. These diagnostics enable the identification of regions of the feature space where prescriptions are less reliable.

Our results highlight how interpretability enables transparent, partition-based decision rules and facilitates explicit control and assessment of the risks inherent in data-driven operational decisions.