Three-Tier Explainability: A Governance Framework for Audience-Differentiated Explanation Obligations in Regulated AI Systems

Explainability in AI governance is not a single obligation. Organisations subject to regulated AI deployment obligations encounter explainability requirements from three structurally distinct audiences: regulators conducting examinations and demanding tamper-evident evidence of decision logic; operators supervising autonomous systems in real time and requiring rationale sufficient to intervene or override; and affected individuals whose rights to understand decisions that materially affect them are established in data protection, consumer protection, and sector-specific law. These three obligations are not a spectrum. They are independent. Satisfying one does not satisfy the others. Treating them as interchangeable, or designing an explainability architecture to satisfy the most prominent audience and assuming the others are thereby met, is a governance design failure with specific regulatory consequences. This paper introduces the Three-Tier Explainability framework as a governance architecture for regulated enterprises deploying autonomous AI systems. The framework distinguishes Tier 1 — Regulatory-Grade Audit Trail, an immutable, tamper-evident, pre-execution record constituting evidence that survives regulatory examination; Tier 2 — Operator-Facing Rationale, a real-time structured account of system reasoning sufficient to support human oversight and intervention; and Tier 3 — Individual-Facing Plain-Language Explanation, a rights-fulfilling account of decision basis accessible to a non-specialist affected person. The paper argues that the independence of these three tiers is not merely a design preference but a governance requirement: each tier has distinct legal grounds, evidentiary standards, failure modes, and remediation obligations. The paper further introduces a Conflation Failure Typology, identifying the four governance failure patterns that arise when organisations collapse two or more tiers into a single architecture. It demonstrates that existing governance frameworks — ISO/IEC 42001:2023, the EU AI Act, GDPR, APRA CPS 230, the NIST AI RMF, and the IMDA Model AI Governance Framework for Agentic AI — establish explainability obligations but do not require the three-tier distinction to be implemented as independent governance obligations. This paper argues that the gap is structural and that closing it is a precondition for regulated enterprises to satisfy their explainability obligations across all audiences simultaneously.