The year 2025 marked a turning point in how artificial intelligence (AI) is applied to two of the most consequential problems in supply chain management: anticipating disruption and forecasting demand. Three developments converged. First, time-series foundation models matured from univariate research prototypes into covariate-aware, production-ready systems, most notably Amazon's Chronos-2 and Google's TimesFM 2.5, and became available through open-source releases and managed cloud services. Second, large language models (LLMs) and knowledge graphs began to turn unstructured news, regulatory notices and supplier documents into structured, network-aware risk signals. Third, agentic AI, meaning LLM-driven systems that plan, call tools and act, moved from vendor roadmaps into planning platforms, at the same time as a volatile tariff environment pushed firms toward AI-enabled scenario planning and digital twins.
Using an integrative literature review of 261 sources, comprising 243 applied studies screened from a consolidated peer reference list together with technical reports, preprints and industry surveys from Gartner, McKinsey, WTW and CIPS, the paper synthesises these advances. It finds that technical capability advanced faster than organisational readiness: model performance was rarely the binding constraint, whereas data quality, multi-tier visibility, governance and accountability for automated decisions frequently were. The paper proposes an integrated framework that links probabilistic demand forecasting, risk sensing and scenario simulation through a governed decision layer, together with six testable propositions. It then discusses theoretical, managerial and policy implications, and sets out future research directions covering evaluation under regime shifts, explainability, AI as a third-party risk domain and human oversight of autonomous agents.