A Chain Is Only as Strong as Its Weakest Link: A Scoping Review of System Integration Audits in AI

Abstract

As artificial intelligence (AI) systems are integrated into diverse interfaces and applications, model-centric audits are insufficient to address risks arising from their deployment. In safety-critical domains such as aerospace, system integration has long been a focal point for technical system audits. Yet, its role in AI auditing remains underexplored. Scanning 4,259 documents, we present a scoping review of AI audits that treat integration as a core tenet of evaluation (n = 58). We find that this emerging form of AI auditing extends beyond individual components to evaluate interactions within an AI system, between an AI system and its deployment environments, and across multiple interconnected AI systems. Specifically, these audits aim to identify integration-related risks, coordinate highly distributed evaluation tasks, and support end-to-end assurance. Unlike other AI audits, they also assess whether the intermediaries connecting elements operate in alignment (compatibility), whether sufficient information is available to assess the system as a whole (completeness), and whether responsibilities are clearly assigned (oversight). Despite this progress, system integration audits in AI remain fragmented, with few explicit measures targeting risks that emerge at system interfaces or their interdependencies. Like other AI audits, meeting traditional audit expectations, such as independence, remains limited by upstream access constraints. We call on the AI auditing community to prioritize system integration as a core strategy for capturing failures beyond the reach of component-level evaluation.