The Tacit Dimension of Agentic AI: Relational Competence, Delegated Trust, and the Irreducibly Human Core of High-Stakes Interaction

Agentic AI systems that act autonomously on behalf of human users are being deployed in medical, legal, and financial domains with governance frameworks designed primarily around task competence: accuracy, efficiency, and cost reduction. This paper argues that task competence is a structurally insufficient governance standard for agentic AI in high-stakes human domains and that two additional competence dimensions—recovery competence and relational competence—are mandatory prerequisites for responsible deployment. We introduce delegated trust failure: the governance gap that emerges when an AI agent acts on a human's behalf, fails, and leaves the human to carry the recovery burden without a pre-built remediation mechanism. We introduce the convenience asymmetry: the systematic divergence between deployer-measured efficiency and user-experienced convenience. And we introduce the tacit knowledge problem: the finding that the most critical professional capability in high-stakes human interaction—what practitioners describe as sensing the vibe and adjusting to the person before asking the first question—is not explicitly definable, not trainable from data, and not certifiable through any current assessment mechanism. Drawing on Polanyi's concept of tacit knowledge and clinical communication research, we propose a three-tier governance framework: task competence for low-stakes automation, recovery competence for consequential transactions, and relational competence for emotionally sensitive high-stakes domains. We argue that in domains where tacit relational knowledge is the critical professional capability, agentic AI should be restricted to administrative functions, with the relational core of the interaction protected as irreducibly human.