Enterprise digital twins (EDTs) are increasingly adopted across sectors such as logistics, supply chain, and manufacturing to simulate, analyze and optimize complex operational systems. They support forward simulation, or “what-if” analysis, which allows decision-makers to evaluate the impact of different operational policies. However, many enterprises demand the inverse inference (i.e., “if-what”): determining suitable changes in the operational policies required to meet specific performance targets. While this inverse inference capability is critical for strategic decision-making, particularly in uncertain multi-objective environments, it remains largely unsupported in current state-of-the-practice EDT frameworks. We address this gap by proposing a general framework to enable “if-what” reasoning in stochastic EDTs using multi-objective Bayesian optimization (MOBO). Our approach treats an EDT as a black-box objective function with noisy multi-dimensional outputs and uses probabilistic surrogate models to efficiently search the input space. We demonstrate the efficacy of this approach through a real-world case study involving a high-fidelity actor-based digital twin of a logistics sorting terminal. The optimization seeks to infer optimal workforce allocation policies that maximize throughput and minimize processing time. We evaluate the performance of two MOBO variants (scalarization and Pareto optimization), analyze the trade-offs between competing objectives, and highlight the sample efficiency and robustness of the method. We further compare our approach with established optimization methods, including simulated annealing, genetic algorithms and gradient-based optimization using neural surrogates, to benchmark performance across diverse strategies. The proposed framework extends the capabilities of EDT technology by adding automated decision support.