From Machine Learning to Foundation and Agentic AI: Evolution of Intelligent Decision Systems Across Domains

Abstract

Intelligent decision systems have undergone a multi-stage architectural evolution—from conventional machine learning and structured analytics through convolutional deep learning, attention-based transformers, graph neural networks, multimodal fusion systems, federated and privacy-preserving frameworks, to generative AI and emerging agentic decision architectures. This evolution is not merely technical: it changes how systems acquire representations, explain decisions, operate across institutional boundaries, integrate into professional workflows, and support high-stakes decisions in healthcare, business, industry, smart infrastructure, agriculture, cybersecurity, assistive technologies, and sustainability. This review characterizes ten evolutionary stages from structured ML through agentic decision systems—and maps their expression across seven application domains. Synthesis reveals that while deep learning and transformer architectures have substantially advanced representational capability, the deployment-critical properties of trustworthiness, validated explainability, uncertainty quantification, and governance accountability have not evolved at the same pace. Generative AI and agentic systems represent a qualitative shift toward interactive and workflow-embedded decision support, but introduce hallucination risk, accountability gaps, and governance demands that exceed current frameworks. A structured research agenda addresses evolution-aware benchmarks, trustworthy foundation-model adaptation, human-in-the-loop evaluation, federated multimodal intelligence, and governance-aware reporting standards for agentic decision systems.