Decision Authority, Silence, and Governance: A Framework for Safe Use of Large Language Models in Financial Contexts

Abstract The rapid adoption of large language models (LLMs) in financial environments has shifted the primary risk profile of artificial intelligence systems. While existing discussions emphasize accuracy, bias, and data quality, a more fundamental failure mode remains underexplored: the uncontrolled escalation of decision authority. This paper argues that the core risk of financial AI systems is not incorrect outputs, but confident outputs delivered in situations where no response should be given. LLMs are optimized for fluency and plausibility rather than for assessing whether responding is appropriate in high-stakes, uncertain, or emotionally charged financial contexts. As a result, even compliant and technically correct responses may function as implicit guidance, transferring perceived decision responsibility from human to machine. To address this gap, the paper introduces a governance-first architectural framework that operates independently of any specific model. The proposed Decision Governance Layer evaluates user intent, contextual and cognitive risk, ethical and regulatory constraints, and decision authority before authorizing system output. Crucially, the framework formalizes silence, refusal, and constrained speech as valid and auditable system actions. The framework is designed to complement, not replace, existing compliance and model risk management structures. By separating response generation from permission to respond, it provides a missing layer of pre-decisional control applicable across internal and client-facing financial AI systems. This work contributes a conceptual foundation for authority-aware AI governance in finance, reframing safety not as improved prediction, but as disciplined restraint.