Crop diseases significantly impact global agricultural productivity, causing losses exceeding $220 billion annually. Traditional disease detection methods rely on human intervention and expert judgment, which are often time-consuming and inefficient. To address these limitations, this study proposes a deep learning-based AI-driven approach for automatic detection of crop diseases. The performance of MobileNetV2, InceptionV3, LeafNet, and ResNet54 is compared with a custom-designed CNN model, AGDFNet (Attention-Guided Multi-Scale Disease Feature Network). AGDFNet integrates Adaptive Lesion Module, Adaptive Focal Loss, and Feature Aggregation techniques to enhance feature extraction, focus on infected regions, and detect rare diseases. The models are trained on 30,500 images across 61 distinct crop disease classes. Experimental results demonstrate that AGDFNet achieves 90% accuracy, outperforming the other pre-trained models. Furthermore, a chatbot interface is incorporated, enabling farmers to upload crop leaf images and receive real-time diagnostic feedback and treatment recommendations. This AI-driven solution provides an effective and scalable approach to minimize agricultural losses caused by crop diseases.