A Layered Control Architecture for Trustworthy AI Agents on Salesforce Agentforce: Integrating Hybrid Reasoning, Platform Enforcement, and Human-in-the-Loop Approval

Abstract

Enterprise CRM platforms now host AI agents that act on customer records, not just answer questions. Salesforce Agentforce improves agent reliability with hybrid reasoning, which separates deterministic Agent Script logic from large language model (LLM) reasoning inside each agent. However, current agent designs still face fundamental limitations: business rules written as natural-language instructions that the LLM may skip, rules enforced only inside one agent while console users, integrations, and other agents change the same records, no systematic route for decisions that exceed delegated authority, and audit evidence scattered across prompts, scripts, and actions. This paper addresses these challenges and asks: can a layered architecture that places each decision in the layer able to guarantee it prevent these failures while preserving the conversational flexibility of LLM agents? The proposed architecture organizes an agent-enabled CRM process into five layers: channel, agent, action contract, platform enforcement, and oversight and monitoring. A control placement model assigns each decision element to one of four zones (platform enforcement, scripted control, LLM reasoning, and human authority) using six criteria applied in a fixed order. Five enforcement patterns bind the zones: double-locked actions, policy as data, draft and approve, explain from state, and warm handoff. A governance lifecycle defines zone-specific tests and releases agent scripts, actions, platform rules, and policy records as one versioned unit. A refund case study and a design-level analysis of eight failure scenarios, drawn from security guidance and platform documentation, evaluate the architecture. The proposed design addresses all eight scenarios by construction, whereas a script-controlled agent fully prevents two and partly prevents four, and an instruction-only agent prevents none. The results show that the two gaps left open by agent-level control, cross-channel enforcement and delegated authority, close when controls are placed below and above the agent. Index Terms—Salesforce Agentforce, AI agents, hybrid reasoning, Agent Script, large language models, enterprise CRM, human-in-the-loop, excessive agency, AI governance, software architecture