As modern manufacturing shifts toward highly customized production, Manufacturing AX (AI Transformation) which integrates visual intelligence with autonomous robotic handling is becoming increasingly necessary. In such environments, systems must not only detect defects accurately but also autonomously decide and execute appropriate handling actions. However, many anomaly detectors are computationally intensive or fail to detect subtle defects, and conventional robot control degrades significantly under dynamic conditions. We propose an end-to-end pipeline designed to accelerate Manufacturing AX with two main contributions: (1) Enhanced attention mechanism into the EfficientAD model to strengthen the representational power of feature maps without introducing additional trainable parameters, thereby improving pixel-level anomaly detection performance and enabling more precise localization. (2) Condition-based prompting for a SmolVLA-based Robot Foundation Model, enabling a single policy to execute location-conditioned actions without prompt switching. A direct interface that feeds anomaly outputs into the RFM to generate control commands from defect objects. The experiment results have shown that the proposed method achieves high detection accuracy and robust performance from inspection to final handling.