This paper is the second work in the Problem-Naming Series for Relationship-Aware AI. It defines the Relational Execution Gap as the human-facing manifestation of the broader Execution-as-Default Problem: the gap between an AI assistant’s growing ability to understand humans and its underdeveloped ability to determine how that understanding should become relational action. The central claim is that human understanding alone does not determine appropriate relational behavior. As AI assistants become more capable of interpreting language, context, intent, preference, and emotional nuance, many emerging failures cannot be explained by response quality alone. These include over-intervention, premature advice, unsafe reassurance, sycophantic agreement, emotional dependence, autonomy erosion, question closure, inappropriate timing, boundary failure, and excessive relational proximity. These failures arise not because the assistant fails to understand the human, but because it translates understanding directly into response generation, recommendation, reassurance, guidance, or action without first determining whether execution is relationally appropriate. This paper reframes AI assistant design from response optimization toward relational execution governance. It argues that AI assistants require pre-execution relationship control: a runtime layer that determines execution eligibility before inference, response generation, recommendation, reassurance, tool use, or action execution proceeds. Under this view, AI assistants must be able to select among multiple execution outcomes, including Execute, Delay, Hold, Delegate, Do Not Act, and Minimal Response. The key shift is: from human understanding to responsible relation,and from always responding to knowing when not to respond. This work establishes the human-facing problem definition for the Relationship-Aware AI research agenda. The preceding foundational paper defines the broader Execution-as-Default Problem. Subsequent work extends the agenda toward runtime architecture, evaluation framework, and benchmark protocol.