Collins (2026a) presents the first large-scale empirical analysis of AI agent governance in global financial services, mapping 49,231 autonomous AI agents across 263 institutions in 45 countries. Three findings emerge: a near-total governance void (95.5% halt absence, 79.4% orphan rate), a classification problem (81.1% of human-in-the-loop agents lack a detectable halt mechanism), and a structural vulnerability termed the SPOF paradox (institutions that invested most in governance created single points of failure whose removal collapses their governance posture to zero). Three formal hypotheses are confirmed: governance investment creates topological concentration (H1), that concentration creates single points of failure (H2), and halt scarcity is structural and not correlated with governance effort (H3). This paper argues that all three findings are structurally predicted by substrate governance theory and were not contingent empirical discoveries. They are the necessary outcome of applying organizational governance tools—committees, oversight functions, reporting lines, ethics boards—to a problem class that requires substrate-layer architecture. The Collins findings are not a measurement of how far institutions have to go. They are a measurement of what happens when the wrong governance layer is applied to an agentic system at scale. The paper develops four propositions. First, the governance void is the predictable result of governing the execution layer while leaving the substrate ungoverned. Second, the HITL classification problem is the predictable result of confusing organizational intention with operational architecture. Third, the SPOF paradox is the predictable result of building network topology governance with hierarchy tools. Fourth, H3—halt scarcity as structural and uncorrelated with governance effort—is the definitive empirical proof that organizational governance and substrate governance are different problem classes requiring different architectural primitives.
Narnaiezzsshaa Truong (Tue,) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: