Workflow-Centric AI Governance: A Sociotechnical Architecture for Accountable Human–AI Decision Systems

The integration of AI into high-stakes domains has revealed a critical sociotechnical gap: while model performance improves, institutional capacity to responsibly operationalize AI outputs remains fragile. Without robust procedural safeguards, human oversight often devolves into performative ritualism rather than substantive review, especially in high-volume decision environments.

To address this, we propose ALTRION, a workflow-centric governance architecture that introduces structured friction into AI-assisted decision pipelines. ALTRION formalizes four governance gates—from baseline constraints to human arbitration—that are designed to enforce cognitive engagement and preserve meaningful human agency. We develop a probabilistic simulation, grounded in literature on alert fatigue and compliance decay, to examine how ALTRION mitigates error rates under conditions of context drift. Our probabilistic model suggests the existence of a regime in which institutional non-compliance stays below a critical threshold, beyond which oversight collapses. Finally, we outline a governance-oriented representation of workflow interactions to render oversight practices auditable and revisable. We argue that explicit workflow governance is essential to prevent AI-enabled procedures from becoming mechanisms of mere moral crumple zones.