We design an automated decision-making system for inventory management using a Transformer-based neural network. Leveraging contextual information and historical data, the system makes two key decisions: (1) order timing and (2) order quantity. To accommodate random vendor lead times, we develop an imitation-learning framework, in which the Transformer imitates ex post optimal decisions computed from historical data to directly output inventory actions. The model adopts the GPT-2 architecture—an off-the-shelf large language model—for efficient fine-tuning under the inventory context. Our framework, InventoryGPT, addresses challenges faced by large e-commerce platforms that manage millions of stock-keeping units while serving customers with stochastic and nonstationary demand and lead-time patterns. By incorporating rich contextual data, the model learns to improve service levels and reduce costs. Empirical results using real-world data from a leading e-commerce platform show that InventoryGPT outperforms traditional and state-of-the-art benchmarks. Moreover, by carefully designing the Transformer’s input and output structures, the proposed InventoryGPT model achieves good interpretability. Our study highlights the potential of Transformer-based neural networks for large-scale decision making in service operations. History: This paper has been accepted for the Service Science Special Issue on the Impact of AI on Service Design and Delivery. Funding: This research was supported by the National Key Research and Development Program Grant 2024YFB3311500 and the Research, Academic and Industry Sectors One-Plus Scheme Grants RAISe+ and RAI-24-1-096A. Supplemental Material: The online appendix is available at https://doi.org/10.1287/serv.2024.0236 .
Liu et al. (2026) studied this question.