This preprint presents a trust architecture for LLM-assisted workflows in multi-tenant enterprise systems. Modern enterprises increasingly use large language models and agentic workflows to automate operational tasks, but these systems introduce new challenges related to privacy, access control, multi-tenant governance, hallucination mitigation, and enforcement of organizational policies. This paper proposes a structured architecture combining tenant-aware RAG routing, schema-validated tool interactions, multi-level guardrails, RBAC-based authorization, and human-in-the-loop (HITL) verification paths to improve reliability and safety. The architecture enables secure and consistent LLM-assisted task execution across diverse enterprise applications while preserving tenant boundaries and ensuring regulatory compliance. The work includes design principles, workflow diagrams, guardrail pipelines, and practical considerations drawn from large-scale enterprise platform engineering.