Welding is a crucial process in manufacturing, but traditional inspection methods are slow, prone to errors, and labor-intensive. Automated detection using artificial intelligence offers a sustainable way to enhance efficiency and cut waste. In this study, we created an AI-driven framework to identify resistance spot welding defects by combining deep learning with experimental validation. ResNet18, ResNet50, UNet, and ResUNet were used to classify and segment weld images. Tensile testing pinpointed defective joints with stresses below 25 MPa, whereas resistivity tests showed that defective welds had a significantly higher electrical resistance. ResNet50 achieved the highest classification accuracy at 95%, and ResUNet provided the best segmentation with a mean Dice coefficient of 87%. These results show that combining AI with mechanical and electrical validation offers a reliable, efficient, and sustainable method for detecting welding defects. The proposed framework supports smart manufacturing and helps advance circular economy practices.