Efficient resource allocation in heterogeneous distributed computer systems is a growing challenge due to the complexity of diverse task requirements and the dynamic nature of system nodes. Traditional scheduling approaches—such as centralized, decentralized, market-based, and optimization-based methods—often fail to adapt effectively to fluctuating workloads and varying node capabilities. To overcome these shortcomings, this research introduces an AI-driven optimization framework that utilizes neural networks and deep learning to dynamically assign tasks to nodes, continuously adapting to real-time changes in system conditions like computational efficiency, security, fault tolerance, and data transfer latency. The proposed framework was tested through extensive experiments comparing two neural network architectures: feedforward neural networks (FFNN) and convolutional neural networks (CCNN). These experiments utilized datasets simulating distributed systems ranging from 100 to 10000 nodes. The FFNN outperformed the CCNN, achieving validation accuracies between 90 and 99.8%, demonstrating its superior scalability and ability to model intricate relationships among node attributes. In contrast, the CCNN showed inconsistent performance, particularly with larger, non-spatial datasets. This approach offers significant benefits over traditional methods, including real-time adaptability, flexible task-node mapping, and enhanced resource utilization, making it a robust solution for heterogeneous environments. Given the Feedforward Neural Network’s (FFNN) demonstrated high accuracy and scalability, it is well-positioned for practical deployment in distributed systems; however, further research is warranted to explore the development of hybrid architectures that integrate the FFNN’s proficiency in modeling complex relationships with the Convolutional Neural Network’s (CCNN) ability to extract hierarchical features from node attributes.