Accurate and interpretable multi-class AD classification from brain MRI is of high value for diagnosis and clinical decision support in early stages. This research presents a hybrid deep model that combines the UNet and Vision Transformer models for utilizing spatially localized features as well as whole contextual information. While the UNet is able to extract detailed spatial features by virtue of convolutional encoding, self-attention is utilized by the ViT for extracting long-range relations. Together, they enable an extensive representation for stable classification. The model is analyzed on the three-class bench named Non-Demented, Very Mild Demented, Mild Demented, and place against a baseline of merely utilizing the UNet. The results indicate that there was better performance of the model of UNet and ViT, 90.01% accuracy, 90.20% precision, 90.17% recall, 90.21% F1-score, better than the baseline of merely using the UNet, 68.96% accuracy. To allow for interpretability as well as trustworthiness, Grad-CAM is utilized for visualizing most salient brain areas responsible for contributing to decision-making, facilitating support for model interpretability. Such interpretable heatmaps justify model integrity and also provide support for clinical relevance. The results indicate the potential of utilizing the fusion of the UNet and the ViT for enhancing diagnosis accuracy as long as explainability is also achieved, showing its promise for use in trustworthy AI for medical imaging.