Agentic artificial intelligence is moving enterprise engineering from passive decision support to autonomous task execution, tool use, workflow coordination and continuous optimization. This transition creates a design problem that is neither purely technical nor purely managerial: enterprises need AI agents that improve productivity while remaining auditable, bounded, reversible and accountable. Existing AI governance standards provide essential principles and process requirements, yet many organizations still lack an implementable system architecture that translates these requirements into operational controls at the level of agent planning, tool invocation, data access, human escalation and post-deployment monitoring. This paper develops a governance-aware reference architecture for agentic AI in enterprise engineering systems and formalizes it through a quantitative risk-control model. Using a design-science methodology, the study synthesizes AI risk management standards, algorithmic auditing literature, human-centered AI design and recent work on AI-form organizations, distributed responsibility and AI stakeholder recognition. The resulting artifact, termed the Governance-Aware Agentic AI Control Architecture, integrates six layers: strategic intent and risk appetite, agent orchestration, data-model-tool infrastructure, governance control plane, observability and evidence, and human escalation with adaptive assurance. The paper introduces three formal constructs - Productivity-Adjusted Residual Risk, Governance Debt and Human Override Threshold - to guide deployment decisions. An illustrative scenario evaluation across customer service, finance operations, human-resource screening and supply planning demonstrates how the model can convert abstract governance principles into measurable engineering checks. The contribution is a practical and theoretically grounded architecture for organizations seeking to deploy agentic AI responsibly without reducing governance to after-the-fact compliance documentation. The paper concludes that trustworthy enterprise AI requires control systems designed into the architecture itself, not merely ethical statements attached to autonomous workflows.
Kwan Hong Tan (Wed,) studied this question.