Variational Autoencoders (VAEs) represent a powerful Generative AI tool that learns the underlying structure of input data and generates new data. This paper applies VAEs to model complex, real-world demand distributions in the context of inventory management. Existing research often relies on traditional methods, such as linear regression or Gaussian mixture models, which frequently struggle to capture the non-linear and multimodal nature of actual demand patterns. VAEs utilize deep neural networks to learn intricate patterns within data and map them to a regularized latent space, enabling the generation of novel demand scenarios. Consequently, using VAEs could result in finding the optimal order quantity more accurately. We hypothesize that VAEs empower businesses to make data-driven decisions to optimize inventory levels and minimize costs, providing a robust and adaptable framework for inventory management compared to traditional approaches. To test our hypothesis, we simulate ten sets of demand distributions with complex patterns and shapes and use VAEs to model these demand patterns. Based on the distributions identified by the VAEs, we calculate the optimal order quantity and expected profits. We then compare these results to scenarios where demand is a assumed to follow traditional distributions, including Normal, Poisson, Erlang, and Uniform distributions. Our findings reveal that, in most cases, the VAE-based approach outperforms the approaches that use traditional distributional assumptions. This research demonstrates the potential of VAEs to contribute to inventory management by providing a more accurate and flexible approach to modeling demand variability, ultimately leading to significant cost savings and improved operational efficiency.