This study was carried out to validate a software-defined sensing (SDS) architecture that combines AI-vision-enabled image sensors with a deep learning model to replace traditional hardware sensors, such as infrared and ultrasonic sensors, for complex human–computer intent recognition. The architecture was implemented in Picbot, an AI-driven, naked-eye augmented reality interactive platform for smart retail that integrates visual AI with high-performance displays. The Picbot model identifies 1293 2D facial key points, significantly exceeding the 106 points provided by Google MediaPipe, an open-source framework widely used for real-time perception tasks such as facial mesh and skeletal tracking. In operational reliability testing, the architecture achieved a 94.56% success rate in recognizing and triggering composite user intent in controlled indoor environments. Deployment in dynamic outdoor environments achieved a 50.41% social share rate, demonstrating the SDS architecture’s practicality for brand engagement and social dissemination. These results highlight SDS architectures as flexible, high-precision AIoT sensing solutions, marking a transition from hardware-constrained sensing to intelligent, algorithm-defined perception. Despite these results, the transition to algorithm-defined sensing has challenges of system stability in high-traffic environments with fluctuating ambient light. Such challenges can be addressed by employing a hybrid AI tracking strategy and real-time edge processing in complex interactive settings.