AI-driven optimization in cloud computing: a systematic review of cost, resource management, and security

Cloud computing environments face persistent structural challenges in cost control, dynamic resource allocation, and security risk management, which traditional infrastructure approaches fail to address adequately. This systematic literature review aimed to synthesize empirical evidence on the application of artificial intelligence (AI) and machine learning (ML) models for cost optimisation, resource management, and security enhancement in cloud computing environments. Following the PRISMA 2020 guidelines and the Kitchenham–Charters methodology, a structured search was conducted across IEEE Xplore, Web of Science, ScienceDirect, and the ACM Digital Library, covering the period 2020–2025. From an initial pool of 216 records, 18 primary studies were selected after applying the PICOC framework, predefined inclusion and exclusion criteria, and a dual-reviewer quality assessment process yielding substantial inter-rater agreement (Cohen's κ = 0.86). The synthesized evidence demonstrates that predictive provisioning systems and intelligent load-balancing mechanisms reduce operational costs by up to 85%, metaheuristic algorithms such as the Whale Optimization Algorithm and Particle Swarm Optimization improve energy efficiency by 30%–40% and increase resource utilization by up to 80%, and deep learning–based intrusion detection systems achieve accuracy levels exceeding 92%. These findings confirm that AI constitutes a structural mechanism for strengthening economic efficiency, operational resilience, and the sustainability of cloud infrastructures. However, heterogeneity in simulation environments, limited validation in production-scale deployments, and insufficient coverage of virtual machine migration dynamics represent critical gaps requiring standardized benchmarking frameworks and empirical validation in hybrid and multicloud architectures. A quantitative synthesis ( Table 1 ) reveals that metaheuristic algorithms achieve 30%–40% cost and energy efficiency improvements, while ensemble deep learning approaches attain >97% security threat detection rates.