Background: Artificial intelligence (AI) and machine learning (ML) are used to support prehospital triage and transport decisions, but their comparative performance versus conventional scores and guidelines remains heterogeneous. Objective: To synthesize original studies evaluating AI/ML models that use data available to EMS at dispatch or on-scene to predict critical outcomes or guide transport modality, and to contextualize findings against recent reviews of AI in prehospital care. Methods: Following PRISMA guidance, we included seven original studies that developed or validated AI/ML models in prehospital settings and nine review articles for background and discussion. We extracted setting, population, inputs, models, comparators, outcomes, and discrimination. Results: Across 7 studies (N ranging from 2,604 to 219,323; mixed retrospective and prospective cohorts), AI/ML models consistently matched or outperformed conventional tools. Deep learning trained on national ED data predicted need for critical care with AUROC 0.867 and outperformed ESI, KTAS, NEWS, and MEWS. Random forest improved one-day and 30-day mortality prediction versus NEWS; adding blood glucose further improved discrimination. An ensemble model for suspected COVID-19 predicted 30-day death or organ support in 7,549 EMS patients. Gradient-boosted triage using EMS vitals and injury patterns improved sensitivity for severe trauma (ISS≥16) versus field triage rules. Large regional cohorts showed ML enhanced NEWS2/DEPT with fewer false positives. Text-based models modestly predicted subsequent events after non-conveyance. Conclusions: AI/ML can augment prehospital risk stratification and triage, particularly when integrating standard vitals with select additional signals (blood glucose) or structured injury features. Prospective external validation, calibration reporting, and workflow-aware evaluation are needed before routine deployment.