Human-on-The-Rail: A Sovereignty Model for Agentic AI Governance — Why HITL Is a Category Error and What Must Replace It

Human-in-the-loop (HITL) is the dominant oversight model in agentic AI governance. The EU AI Act Article 14, NIST AI RMF, ISO 42001, and SR 11-7 all invoke variations of it as the primary mechanism for maintaining human authority over AI systems. This paper argues that HITL is a category error when applied to agentic systems—not an insufficient implementation of the right model, but the wrong model applied to the wrong problem class. HITL assumes four properties that agentic systems structurally violate: that the human can reconstruct the agent's reasoning at the point of review, that the human can override the agent before consequences materialize, that the human's judgment is the authoritative governance layer, and that the human's review is causally meaningful rather than ceremonial. In multi-step agentic workflows with autonomous tool use, parallelized actions, high-tempo decision chains, and stateful reasoning across long horizons, none of these assumptions hold. The human becomes a rubber stamp—not because humans are inadequate but because HITL was designed for tools, not actors. The paper proposes Human-on-the-Rail (HotR) as the correct replacement model. HotR does not attempt to insert the human into the agent's execution loop. It constrains the agent within a governance corridor the human has defined and can verify. The human's role is: defining the authority boundary, defining the permissible action space, defining the interpretive constraints, receiving audit-ready artifacts, reconstructing the agent's reasoning after the fact, and intervening at governance boundaries rather than at every step. HotR tries to constrain the machine so it cannot leave the human's governance corridor. HITL tries to slow the machine down to human speed. These are different models with different architectural requirements. The paper further argues that dissolving the human/machine distinction—as proposed in recent legal scholarship—is a productive move only if it is replaced with a sovereignty model: the distinction between sovereign and non-sovereign decision authority, interpretive and non-interpretive labor, reversible and irreversible actions, and auditable and non-auditable reasoning chains. Without a sovereignty model, dissolving the human/machine distinction produces humans treated as interchangeable components, agents treated as quasi-persons, no clear locus of accountability, no mechanism for detecting interpretive drift, and no architecture for enforcing role integrity.