Artificial Intelligence for High-Stakes Decision Support: Architectures, Applications, and Deployment Challenges

Abstract

Artificial intelligence (AI) is increasingly embedded in consequential decision-making processes across healthcare, assistive technologies, smart infrastructure, agriculture, business analytics, cybersecurity, and sustainability. Unlike general-purpose AI deployments, high-stakes decision support demands not only predictive accuracy but also explainability, robustness, privacy, scalability, human oversight, and governance readiness. This structured critical review synthesizes to map the current landscape of AI for high-stakes decision support using a four-axis taxonomy: application domain, data modality, architecture family, and deployment concern. The review identifies six application domains, healthcare and biomedical decision support, human-centered and assistive AI, smart infrastructure and cyber-physical systems, agriculture and sustainability, business and enterprise decision support, and cybersecurity and distributed intelligence, and eight architecture families ranging from conventional machine learning and convolutional neural networks to vision transformers, graph neural networks, Bayesian models, generative AI, and federated learning systems. The synthesis reveals that while significant architectural advances have been made, deployment-critical properties such as uncertainty quantification, privacy-preserving inference, real-time feasibility on edge devices, and governance-aligned reporting remain inconsistently addressed. Future research must prioritize cross-domain benchmarking, trustworthy and auditable AI pipelines, human-in-the-loop frameworks, and evidence maturity standards appropriate for high-stakes contexts. This review provides an evidence-grounded taxonomy and actionable research agenda for researchers and practitioners to build the next generation of responsible AI decision-support systems.