Integrating computer vision into agriculture offers transformative potential for sustainable crop management, particularly in the early detection of plant diseases. Traditional deep learning models, such as Convolutional Neural Networks and Vision Transformers, have shown high accuracy in disease classification but are limited by their dependence on large labeled datasets, lack of generalization, and difficulty adapting to new tasks. This study explores the applicability of CLIP, a multimodal Vision-Language Model, for plant disease detection using zero-shot and fine-tuning strategies. We evaluate CLIP in different experimental setups. Our results highlight the strengths and limitations of CLIP capabilities in the agricultural domain and assess how prompt engineering and domain adaptation can improve performance. The findings contribute to the understanding of how foundation models can be adapted for precision agriculture, potentially reducing the reliance on chemical treatments through early, automated, and scalable disease detection.