Governing agentic AI autonomy: an accountability-contingent frontier of organizational decision outcomes

Abstract Organizations are delegating consequential decisions to agentic AI systems faster than theory can explain when delegation is safe. The literature contains two traditions that give opposite answers. The delegation tradition, rooted in agency theory, treats autonomy as a resource whose value rises with capability and therefore predicts monotone gains. The accountability tradition, rooted in theories of responsibility and oversight, treats autonomy as a risk to be bounded and therefore treats governance as a brake. The two traditions talk past each other because each leaves a different assumption implicit: the first omits the cost of errors and oversight from the value of autonomy, the second omits the efficiency that autonomy buys. This paper makes the missing object explicit. It develops an expected-outcome model in which efficiency gains, error costs, and oversight load jointly determine the value of autonomy, and in which the accountability architecture is a design variable that shifts both the height and the location of the optimal autonomy level. Four propositions follow, including that accountability shifts the governance frontier up and to the right while task complexity moves the peak to the left. The framework is confronted with two bodies of evidence. A coded synthesis of sixteen recent studies shows that the reported return to autonomy depends on the evaluation metric: accuracy-scored studies report monotone gains, decision-outcome and oversight-scored studies do not (Fisher exact p = 0.025). A real-data panel of 1,383 incidents from the AI Incident Database, 2019 to 2026, shows an inverted-U in adoption intensity with an implied peak near 73 percent adoption, governance instruments associated with slower exposure growth in the point estimate, and a regime break that trend-based forecasts miss. A stylized computational experiment, labeled as illustrative, reproduces the frontier and shows hybrid accountability raising the optimal autonomy level from 3.7 to 4.3 on a five-level scale. The contribution is a value-theoretic reconciliation of the delegation and accountability traditions: an outcome metric, a validation standard, and a governance frontier that make the question of how much autonomy is safe a tractable empirical one.