Modern supply chains depend on having the right products in the right place at the right time. When companies keep too little stock, customers face delays and production can stop. When they keep too much, they tie up money and risk waste. Improving inventory decisions therefore matters not only for business costs, but also for the reliability of high-tech manufacturing, spare-parts services, and other sectors that society depends on. This dissertation shows how artificial intelligence (AI) can help make these decisions better. Deep reinforcement learning (DRL), a form of AI that learns decision rules through repeated interaction with data or simulations, is attractive for inventory management because stock decisions must be made step by step under uncertainty. It can learn flexible decision rules that adapt to changing situations by taking into account current stock levels, uncertainty about future demand, and the long-term consequences of ordering too much or too little. However, using this type of AI in practice is not easy. It often requires a great deal of expert knowledge, computing power, and careful adjustment before it works well. Inventory problems make this even harder: demand and delivery times are unpredictable, which can make learning unstable; when conditions change, models often need to be retrained; and in larger systems, the number of possible decisions can quickly become too large to handle efficiently. As a result, standard DRL methods, originally developed for games and robotics, are often not well suited to inventory problems as they are. The central conclusion of this thesis is therefore simple: generic AI is not enough, but AI tailored to the structure of inventory problems can outperform widely used decision rules. To demonstrate this, the thesis develops and tests three tailored approaches. First, it introduces a new algorithm called Deep Controlled Learning (DCL) for single-item inventory problems. This method is designed specifically for settings in which uncertainty makes inventory decisions especially difficult; for example, when unmet demand leads to lost sales, when products can perish before they are sold, or when delivery times are unpredictable. Across these settings, DCL consistently achieves lower costs than both standard DRL methods and leading decision rules. Second, the thesis shows that AI can be made more practical by reducing the need for retraining. In real organizations, demand patterns and other parameters often change, and retraining a model every time is costly. To address this, the dissertation develops the Train, then Estimate and Decide framework. With this approach, one policy is trained across many possible situations and can then be reused for new cases by updating the estimated system conditions. This makes AI-based inventory control more scalable in practice. Third, the thesis extends these ideas to large multi-item systems governed by service agreements, where companies must meet contractual availability targets. Here, the challenge is that the number of possible joint decisions grows explosively. The dissertation shows how to simplify this by letting the AI choose among smart precomputed strategies instead of building every decision from scratch. This leads to further cost savings while maintaining service levels. Overall, this research helps bridge the gap between AI and real operational use. It shows that, when designed around the realities of inventory systems, AI can become a practical and scalable tool for improving supply-chain decisions.