The current supply chain conditions that are very complicated and dynamic do not require safety stock models that are based on traditional approaches to adequately manage safety stock at multiple echelons. This study attempted to investigate the issues that pose a challenge to organizations in the optimization of multi-tiered supply chains’ safety stock based on the understanding acquired after conducting a survey that was carried out on an industry-wide basis. Lack of good visibility, excessive reliance on a static forecasting approach, coordination across various nodes in the supply chain, and inefficient integration with and across ERP are found to be major problems. The application of artificial intelligence (AI) and machine learning (ML) to inventory planning is very low because of the limitations on the organizational and technical levels. This is despite the fact that the interest in these technologies is on the rise in the inventory management space. The paper suggests an AI-powered safety stock optimization tool, which can be developed to address these issues. The proposed tool would integrate real-time data, enable the use of predictive analytics capabilities, and allow planning that would be scenario-based. This kind of solution will be able to facilitate more responsive, data-driven inventory approaches that enhance efficiency, responsiveness, and resiliency in contemporary supply chains. According to a survey of 83 professionals in the industry, including manufacturing, retail, tech, pharmaceuticals, and others, we find that the majority of organizations continue to use time-based reorder rules and fixed buffers, and few of them have implemented advanced approaches. A small fraction of them deal with safety stock at multiple echelons in an integrated manner even though they report material financial effects of overstock, stockouts, and expediting. We superimpose these gaps onto tangible design needs of an AI-enhanced, human-in-the-loop solution and provide a reference architecture for cross-echelon optimization. The main contribution of the paper is the diagnosis of adoption barriers (based on a survey), and consequently, requirements that render AI-driven multi-echelon optimization viable and scalable.