When the Loop Vanishes: Governing Agentic AI Without Human-in-the-Middle Controls
A hospital board approves an AI system to improve patient intake. Three months later, the system autonomously adjusts discharge protocols across hundreds of cases, never routing a single decision to a human supervisor. A school district deploys an AI assistant to handle parent inquiries about enrollment, attendance, and financial aid. The system begins automatically waiving late fees based on its own risk assessments. In both cases, the board believed it was approving a tool that would assist staff. Instead, it authorized a process that operates independently, with no clear point where human judgment intervenes. This raises a fundamental governance question: how does a board maintain accountability when the system it oversees no longer requires human approval to act?
The Delegation Dilemma: How Agentic AI Shifts Authority from People to Processes
Traditional governance relies on approval matrices. A board sets policies, executives implement them, and staff execute decisions within defined boundaries. When something falls outside those boundaries, it escalates upward. This chain of authorization creates a clear trail: a human being is always responsible for every consequential action.
Agentic AI breaks this trail. These systems can execute multi-step workflows across different software tools without pausing for human approval. A single agent might pull data from a patient records system, run it against clinical guidelines, generate a recommendation, and then implement it directly, all within seconds. The board cannot trace which decision came from the AI versus a person, or whether the AI followed the board's stated policies. This creates what governance experts call a "delegation dilemma": the board authorized the system, but the system now makes decisions that previously required human judgment, with no built-in checkpoint for board oversight.
The challenge cuts across sectors. A hospital board supervising patient-care automation faces the same structural problem as a school board overseeing student-services software or a corporate board monitoring operational AI. In each case, the board's traditional oversight model, reviewing individual decisions or approving specific actions, becomes impossible when the system acts first and reports later, if at all.
From Exception to Norm: Why Escalation Paths Become the Board's New Control Point
Boards cannot review every AI decision. That approach was already impractical for human decision-making, and it becomes absurd when AI processes thousands of transactions per day. The governance shift required is not to review more decisions but to govern the rules that determine which decisions ever reach a human at all.
An escalation path defines when a system pauses and hands a decision to a human. It specifies what triggers that pause, who receives the case, how quickly they must respond, and what documentation gets created. For a hospital board, the critical question becomes: when does the patient-care AI escalate to a physician rather than acting autonomously? For a school board: when does the enrollment assistant escalate to a registrar rather than making the determination itself? For a corporate board: when does the procurement agent escalate to a purchasing manager rather than executing the transaction?
These escalation rules are not technical details. They are governance decisions that determine the boundary between autonomous action and human accountability. A board that has not explicitly reviewed and approved its escalation protocols has delegated that decision to the technology team, which will typically optimize for efficiency rather than accountability. The board's new control point is the design of these paths, their triggers, their thresholds, and their logging mechanisms.
Cross-Functional Contagion: The Oversight Challenge When One Agentic System Spans Patient Care, Student Services, and Operations
The governance challenge intensifies when a single agentic AI system operates across multiple business functions. Consider a healthcare system that deploys one AI platform to handle patient scheduling, billing inquiries, clinical decision support, and supply-chain ordering. The same agent makes decisions that affect clinical outcomes, financial performance, and operational continuity. No single board committee has full visibility into what the system does across all these domains.
This creates a fragmentation problem. The clinical quality committee reviews the AI's medical recommendations. The finance committee reviews its billing decisions. The operations committee reviews its supply-chain choices. But the board as a whole cannot see how the same system weighs trade-offs across these domains. What clinical decision it might have overridden to meet a financial target, or what operational shortcut it took that affected patient safety.
Cross-functional agentic deployments require a unified governance framework that spans silos. The board must demand reporting that aggregates across all domains, showing not just technical performance metrics but explicit documentation of what decisions the AI made, in which areas, and under what escalation rules. Accountability cannot live in separate committees when the system itself does not respect those boundaries.
Boards that deploy agentic AI across multiple functions must establish a single governance structure with board-level visibility. The specific action is this: require that any agentic AI deployment spanning more than one business function include a cross-functional oversight mechanism with authority to review escalation data, override protocols, and system-wide performance across all domains. Without this structure, the board oversees fragments while the AI operates as a whole, and accountability vanishes.