LoRaWAN is widely adopted as a low-power networking solution for the Internet of Things, connecting massive numbers of devices in smart cities, agriculture, and industry. However, ensuring reliable communication remains a challenge when confirmed uplinks require acknowledgments under strict duty-cycle and capacity limits. This thesis proposes a probabilistic framework to evaluate reliability and scalability in multi-cell LoRaWAN networks and explores AI-enabled strategies for improving acknowledgment control. The results provide insights into how LoRaWAN can support more dependable IoT services at scale, bridging theoretical analysis with practical design for next-generation large-scale deployments.