Neural Network Procure-to-Pay and Supply Chain Intelligence -Augmented Cloud ERP Architecture for Predictive

Abstract

Cloud ERP programs increasingly expose operational events to external analytics and AI services, but the practical question is how a predictive model should be placed inside a governed transaction process. This paper proposes a neural network-augmented cloud ERP architecture for predicting late-delivery risk in procure-to-pay and supply-chain execution. A reproducible simulation study was conducted on 60,000 synthetic ERP-like purchase and shipment records using only pre-outcome features available at or before an operational decision point. A multilayer perceptron was compared with logistic regression and random forest baselines using a chronological 80/20 split. The neural model achieved an ROC-AUC of 0.818 (95% bootstrap interval 0.810–0.826) and 0.700 accuracy; random forest achieved 0.816 ROC-AUC and the highest F1-score of 0.759. The small performance gap is itself operationally relevant: model complexity should be justified by measurable benefit rather than assumed. The architecture maps ERP events, cloud integration, feature services, model inference, workflow controls, and human approval into a single deployment pattern. The study contributes a leakage-aware evaluation protocol and a practical blueprint for integrating predictive intelligence into SAP S/4HANA-style cloud ERP landscapes without allowing the model to become the system of record.

Keywords: cloud ERP, neural networks, SAP S/4HANA, SAP BTP, predictive supply chain, procure-to-pay, late-delivery risk, machine learning, Industry 4.0.