From Pilot to Production: A Decision-Centred Framework for Governed AI in Manufacturing Operations

Artificial intelligence (AI) is now widely piloted in manufacturing, yet many pilots do not become dependable production systems. This paper argues that the main obstacle is not algorithmiccapability but the absence of a decision-centred design discipline that links data, models, peopleand governance. Drawing on the author's book AI-Driven Production Management (2026), it proposes a framework in which every AI application starts from a named operational decision, itsowner and a measured baseline, and is bounded by explicit authority limits, human override andmonitoring. The framework is applied across four core domains: quality control, productionscheduling, supply-chain management and predictive maintenance. It is extended to digital twins,agentic AI and large language models. A reproducible synthetic case study on component-readiness prediction for a hypothetical UK hydraulic-equipment manufacturer illustrates the evaluation discipline. A gradient boosting classifier reached a ROC-AUC of 0.669 against 0.587 for atransparent rule, while the rule kept higher recall (0.768 against 0.652). The paper also sets outstage gates with kill criteria, an NPV sensitivity showing that a 40% shortfall in benefit realisation can turn a positive case negative, and the UK and EU regulatory position as of October2026. Implications for practice and research are discussed.