The rapid evolution of cloud-native architectures has fundamentally transformed how organisations design, deploy, and maintain digital infrastructure. Modern enterprises operate thousands of interdependent microservices across multi-cloud environments, creating an operational complexity that far surpasses the capacity of traditional manual or rule-based provisioning approaches. Human-driven workflows introduce latency, inconsistency, and a persistent risk of misconfiguration factors that directly impair system availability, performance, and cost efficiency. This paper presents the AI-Driven DevOps Automation Framework (ADAF), a novel architecture that integrates machine learning, large language models (LLMs), and deep reinforcement learning (DRL) to achieve fully autonomous infrastructure provisioning. ADAF operates through a closed-loop control cycle continuously ingesting telemetry, predicting workload demand, synthesising Infrastructure-as-Code (IaC) configurations, orchestrating deployments via Kubernetes, and executing self-healing responses to detected anomalies all without requiring human intervention. The framework was evaluated across three cloud environments (AWS, GCP, Azure) using both synthetic benchmarks and a production-grade microservices application. Results demonstrate that ADAF reduces average provisioning time by 92% compared to manual DevOps processes (from 42.6 to 3.4 minutes), decreases infrastructure costs by 43% over a six-month deployment window, improves Mean Time to Detect (MTTD) anomalies from 14.2 minutes to 1.3 minutes, and achieves a workload forecasting MAPE of 4.2% using a Transformer-based time-series model. The DRL provisioning agent converges after approximately 320 training episodes and maintains a 91.4% autonomous deployment success rate. These findings establish ADAF as a significant advancement in AIOps and autonomous systems research, with practical implications for enterprise-scale DevOps, Site Reliability Engineering (SRE), and FinOps practices. Future directions include extension to edge computing environments, federated learning for privacy-preserving cross-organisation AIOps, and formal verification of LLM-generated IaC plans.