Runtime assurance for enterprise agentic AI systems: A policy-gated control model with quantitative autonomy-risk scoring

Enterprise adoption of agentic artificial intelligence (AI) is moving from passive text generation toward autonomous planning, tool use and cross-system workflow execution. This transition creates a control gap: conventional model governance evaluates outputs or development processes, while agentic systems create risk through sequential actions, delegated authority and changing operating context. This paper develops a runtime assurance architecture (RAA) for enterprise agentic AI and formalizes a quantitative Autonomy-Risk Exposure (ARE) score for deciding when an agent may execute, must be sandboxed, requires human approval or must be blocked. A design-science method was used to synthesize requirements from AI risk-management standards, generative AI security guidance and agentic AI threat literature. The model was then evaluated through a reproducible scenario simulation of 2,000 enterprise agent episodes across knowledge assistance, data retrieval, internal workflow and external transaction tasks. Results show that the full RAA configuration reduced mean ARE from 35.7 to 24.4 points (31.5% reduction), decreased invalid or policy-conflicting actions from 9.8% to 4.6%, eliminated unsupervised pass-through of high-risk invalid actions in the simulated environment, and improved mean audit evidence coverage from 0.61 to 0.89. The control benefit was achieved with a mean latency overhead of 95 ms and human approval for 12.6% of episodes. The paper contributes a practical reference architecture, a risk-scoring equation, a policy decision algorithm and implementation guidance for organizations deploying agentic AI in regulated or high-consequence workflows.