When collecting data for decision-making, the informativeness of the data is crucial. In the predict-then-optimize framework, a common method for obtaining personalized decisions, training a prediction model with informative data points can significantly reduce the expected risk of the decision. In this talk, we examine how to sequentially collect informative data for general decision-making problems in the predict-then-optimize framework. To capture the ‘importance’ of each data point regarding the decision, we propose a new concept: the ‘value of one data point.’ This concept estimates the expected risk reduction of the downstream decision-making problem when adding a specific data point to the current training set, before observing the data point’s label. The value of one data point is closely related to the current confidence in the downstream decision. To estimate this value, we introduce a feature-dependent upper bound, which provides useful insights into the importance of each data point. We use this upper bound to develop a personalized incentive policy for effectively collecting survey data from customers to obtain their preferences. We provide theoretical guarantees for our algorithm, which demonstrate that, under certain conditions, our policy requires smaller cumulative incentives than any fixed incentive policies to achieve the same level of decision risk. Our numerical results further validate our personalized incentive policy across various downstream decision-making problems.