Accurate demand forecasting is a foundational capability for modern supply-chain competitiveness: it reduces inventory holding and shortage costs, improves service levels, and enables leaner, more sustainable operations. This paper investigates an AI-based predictive analytics framework that integrates advanced time-series decomposition, representation learning, and hybrid machine-learning models to improve demand forecast accuracy and translate forecasts into inventory decisions that maximise fill-rate while minimising total cost. We first survey state-of-the-art forecasting architectures — including LSTM/GRU variants, Transformer-style sequence models, and ensemble tree methods — and discuss methods to incorporate exogenous signals (promotions, price, calendar, weather, macro indicators) and hierarchical cross-product dependencies. Next, we propose a modular solution combining (1) multi-resolution time-series decomposition to separate trend, seasonal and high-frequency components; (2) a hybrid deep learning forecaster that fuses sequence encoders with attention mechanisms; and (3) inventory decision logic that converts probabilistic forecasts to (s, S) and newsvendor-style policies using scenario-based optimisation. We evaluate the approach on multiple retail and manufacturing datasets, demonstrating consistent improvements in point and probabilistic forecast accuracy (reduction in MAE and CRPS relative to standard baselines) and measurable inventory benefits (lower days-of-supply and fewer stockouts) under realistic lead-time and promotion scenarios. We close by outlining operational considerations for deployment — data pipelines, model governance, explainability, and integration with ERP/APS — and identify research directions including causal demand drivers, continual learning under concept drift, and joint forecasting-inventory optimisation..