Agentic artificial intelligence (AI) has moved from pilot experimentation to enterprise-scale deployment. The financial case is increasingly well documented, yet the same body of evidence shows that financial upside and governance readiness are diverging: most organizations report deploying agents faster than they can govern them. This paper synthesizes converging analysis on two interlocking questions. First, what structural conditions enable agentic AI to scale reliably, and how should an organization sequence its investments across data foundations, operational maturity, and culture? Second, and more consequentially, how should an organization decide which decisions may execute at machine speed and which must retain human-scale authority? The paper develops an integrated answer: value capture and risk containment are not competing objectives but the same engineering problem, solved by classifying decisions along axes of reversibility, stakes, and ambiguity, and by embedding that classification into a three-layer governance architecture with explicit escalation paths and stop triggers. The analysis further identifies where this boundary is currently eroding in practice — in hiring, healthcare claims adjudication, algorithmic management, predictive risk scoring, and essential-services pricing — and concludes that the placement of governance authority, including the recurring question of whether the finance function should lead it, is itself a design decision with material consequences for both value realization and enterprise risk. The paper closes with a strategic roadmap oriented toward the C-suite and digital transformation leadership, structured around governed data foundations, operational maturity, and organizational culture as the three pillars through which quadrupled ROI and double-digit EBITDA gains become achievable rather than aspirational. Keywords: agentic AI; enterprise governance; human-in-the-loop; decision architecture; AI maturity; return on investment; organizational risk; moral crumple zones; algorithmic accountability
E. Katz (2026) studied this question.