Designing Accountability In: Technical Architecture and Human Oversight Across AAMM Levels

The first two papers in this series established that AI decision-making authority is expanding across organizations faster than governance frameworks can keep pace, and applied the AI Authority Maturity Model (AAMM) to documented enterprise deployments to demonstrate the consistency of this pattern. This paper addresses the underlying technical question: how does system architecture itself enable or undermine human oversight, and what specific engineering practices are required to maintain accountability at each AAMM level? The paper introduces a distinction between governance (organizational policy) and guardrails (technical enforcement), and argues that this distinction is consequential: governance without guardrails is unenforceable, while guardrails without governance are arbitrary. Drawing on the NIST AI Risk Management Framework (2023), the EU AI Act’s Article 14 human oversight requirements, the ‘Moffatt v. Air Canada’ (2024) liability ruling, and documented software engineering patterns in agentic AI systems, the paper maps specific technical mechanisms — domain bounding, human-in-the-loop interrupt architecture, circuit breakers, immutable audit trails, and output guardrails — to the AAMM levels at which they become necessary rather than optional. The central argument is that accountability must be designed into AI systems at the architectural level, not retrofitted after deployment, and that the choice of architecture is itself a governance decision with measurable legal and operational consequences. Keywords: AI system architecture, human-in-the-loop, HITL, guardrails, audit trail, circuit breaker, domain bounding, AAMM, agentic AI, software engineering, AI accountability, NIST AI RMF, EU AI Act