Governance-Aware Adaptive Agentic AI Orchestration for Enterprise SaaS: A Design Science and Multi-Case Evaluation

Abstract

Enterprise software-as-a-service (SaaS) workflows increasingly expose large language model agents to heterogeneous tasks, cross-application dependencies, external tools, governance constraints, and human approval requirements. Fixed orchestration patterns can be effective for stable workflows but do not explicitly adapt their coordination structure to runtime task conditions. This study developed and evaluated a Governance-Aware Adaptive Agentic Orchestration Framework (GAAOF) using design science research, a controlled comparative experiment, and an embedded multi-case design. The benchmark contained 180 enterprise workflow tasks spanning Customer/Revenue Operations, IT/Service Operations, and Finance/Administrative Operations. Each task was executed under five orchestration configurations and repeated five times, yielding 4,500 primary executions. GAAOF achieved a 91.6% task-success rate, 94.0% workflow completion, and 96.0% correct tool/action execution. Relative to the strongest fixed baseline, task success increased by 5.8 percentage points while latency, execution cost, and tool calls decreased by 12.3%, 22.0%, and 17.5%, respectively. Governance-aware routing reduced the policy-violation rate to 1.3%, and mixed-effects analysis showed approximately 74% lower odds of policy violation. Significant pattern-by-complexity, pattern-by-risk, and pattern-by-dependency interactions demonstrated that orchestration effectiveness was context contingent. Ablation, subgroup, cross-case, and sensitivity analyses further showed that governance-aware selection, runtime adaptation, and human escalation contributed complementary benefits. The findings support adaptive orchestration as a practical design principle for enterprise agentic systems requiring joint optimization of effectiveness, efficiency, reliability, and governance.