<p>Agentic AI systems execute multi-step action sequences on behalf of users by invoking tools, accessing data, delegating to sub-agents, and making downstream decisions. Users typically authorize a bounded task, yet the systems that carry it out may operate across action spaces that were neither disclosed nor reasonably comprehensible at the point of authorization. This conceptual article introduces consent collapse to describe the structural failure that occurs when authorization for a task is treated as authorization for dynamically determined action pathways that exceed the user’s reasonable understanding. It explains why organizations may nevertheless adopt agentic systems for speed, scale, continuity, and reduced coordination costs, while incurring new legal and ethical exposure. The analysis situates moral authorization alongside the EU General Data Protection Regulation’s distinct requirements for valid consent, lawful processing, transparency, purpose limitation, and automated decision-making. It then proposes a Bounded Authorization Framework that preserves agentic utility inside declared scope while requiring re-authorization for material scope extensions and explicit gates for irreversible actions. The framework is a normative design proposal rather than a legal safe harbour or an empirically validated intervention.</p>