Maritime ports operate under increasing pressure to strengthen hazardous cargo inspection while maintaining operational throughput and supply chain continuity. Although AI-assisted screening systems improve risk prioritization, inspection resources such as scanning lanes, personnel, and processing capacity remain constrained. This study examines hazardous cargo inspection as a capacity-limited operational decision problem using a discrete-event simulation framework grounded in queueing theory. Containers arrive stochastically and receive probabilistic AI-generated risk scores. Selected containers enter a multi-server M/G/c inspection system with parallel inspection lanes and variable service times. Four inspection strategies are evaluated: random inspection, rule-based inspection, AI accuracy-maximizing inspection, and congestion-aware balanced inspection. System performance is measured using hazardous cargo detection rate, false positive rate, dwell time, queue length, throughput stability, and operational cost across varying utilization regimes. Results show that AI-assisted screening improves detection performance under moderate utilization conditions. However, aggressive accuracy-maximizing policies create nonlinear congestion escalation as utilization approaches system capacity, resulting in unstable queues and increased dwell times. In contrast, congestion-aware balanced policies achieve superior system-level performance by maintaining operational stability while preserving strong detection capability. The findings demonstrate that improved AI-based risk scoring alone does not guarantee operational efficiency. The value of AI in maritime inspection systems depends on how risk information is integrated with adaptive, capacity-aware inspection policies. This study provides a practical engineering framework for evaluating AI-enabled cargo screening strategies in modern port operations and contributes to research on resilient and congestion-aware maritime logistics systems.