Architectural Trust Controls in Enterprise Generative AI Platforms: A Systematic Review of Mechanisms, Frameworks, and Governance Gaps

Abstract

The rapid adoption of generative AI in enterprise creates a trust gap between governance infrastructure and technical capability. Transparency approaches focused on model interpretability are insufficient for autonomous AI agents making important multi-stakeholder decisions with limited human oversight․ This systematic article describes architectural solutions for controls of trust in enterprises. GenAI platforms including enforcement primitives, hallucination risk quantification, policy-as-code, multi-agent governance and observability infrastructure. Following PRISMA guidelines, a systematic literature review was conducted across Google Scholar and arXiv. A total of 30 relevant papers were identified from an initial 600․ The inclusion criteria for the papers included requirements for architectural enforcement‚ empirical evaluation‚ and implementation in an enterprise setting․ The results support five emergent pillars of control: telemetry-first observability, zero-trust service meshes, cryptographically audited action gating, structured knowledge substrates, and graduated multi-agent enforcement. For two model governance use cases, hallucination can be reduced by 30-45% through RAG and behavioral guardrails, and adversarial detection can deter 72.9% of attacks with a 2.9% false positive rate․ However, existing single-layer governance systems are vulnerable to over 90% of adversarial attacks. We present Cognitive Gatekeeping, a framework that operationalizes enforced incapacity over intent in multi-layered controlled architectural constructs, with implications for standardized benchmarking, cross-platform interoperability, and enterprise-scale governance of AI.