Breast cancer treatment decisions heavily rely on biomarker assessments, traditionally obtained through resource-intensive chemical processes and genomic assays. This introduces challenges, including high costs, long turnaround times, and inter-observer variability. Additionally, these methods may be unavailable in some countries. This work explores deep learning-based analysis of gigapixel whole slide images (WSIs) to predict biomarker expression, guide and assist the treatment process, focusing on hematoxylin and eosin (H&E)-stained slides as a cost-effective alternative. We employed multiple instance learning (MIL) frameworks, progressed from attention-based MIL to transformer-based MIL and vision transformers, leveraging inter-patch relationships in WSIs to handle slide-level labels effectively. Models were trained and validated on large multi-institutional cohorts, demonstrating robust performance with a high area under the ROC curve (AUC) for certain biomarker predictions (0.93–0.96 across independent cohorts). For recurrence risk, integration of the Prov-GigaPath foundation model achieved strong OncotypeDX score discrimination (external AUC of 0.82) with excellent generalization. Clinical utility was rigorously evaluated through sensitivity-specificity trade-offs, negative predictive value (NPV) for low-risk identification (21–25% of patients with NPV 0.96–0.97), and survival analysis, yielding significant hazard ratios (2.0–4.1) for recurrence-free and breast cancer-specific survival. Following these evaluations, the results highlight the potential of deep learning on WSIs to spare unnecessary testing for specific patient groups, enhancing accessibility, reducing costs, and refining personalized treatment strategies—such as guiding chemotherapy decisions and forecasting recurrence risk—ultimately paving the way for the clinical integration of AI-assisted pathology.