Conventional fixed-time traffic signal operations at urban intersections are typically based on prescheduled plans that presume stable and recurring traffic patterns, particularly during peak commuting hours. However, recent societal changes—including flexible work schedules, telecommuting, and evolving workweek structures—have introduced greater variability in traffic demand, thereby diminishing the effectiveness of traditional peak-hour-focused control strategies. This study investigated the performance of an AI-based adaptive traffic signal control system that operated independently of predefined time plans. A field demonstration was conducted in Jeju City, South Korea, where the system was deployed in both the cyclic and acyclic operation modes. By leveraging real-time traffic data obtained from AI-enabled video detectors, the system adjusted the signal timings on a per-second basis in response to dynamic traffic conditions. The performance was evaluated against the conventional time-of-day (TOD) control method under diverse traffic scenarios, including typical weekdays, weekends, and local event days. The AI-based system achieved substantial reductions in intersection delays—24.3% on weekdays, 22.2% on weekends, and 17.1% on event days—compared with the TOD baseline. Moreover, it preserved a comparable level of traffic progression (measured by the proportion of non-stop vehicle flows) even during acyclic operations. The greatest efficiency gains were observed during the nighttime and low-traffic periods, underscoring the capacity of the system to minimize unnecessary delays under variable conditions. These results highlighted the potential of AI-based adaptive signal control as a viable alternative to conventional fixed-time operations, offering enhanced responsiveness and operational flexibility in increasingly complex urban traffic environments. Future research will focus on scaling the system to larger networks and developing integrated optimization strategies across multiple intersections.