PolicyGuard Enhances LLM Agent Adherence to Company Policies

Ensuring Large Language Model (LLM) agents adhere to company policies is crucial for organizational workflows, especially when handling user requests and making tool calls. Existing safeguarding methods often focus on blocking non-compliant actions at an argument level, which can be insufficient for complex, multi-turn conversations. This research introduces PolicyGuard, a novel sub-agent verifier designed to address this broader challenge. PolicyGuard operates by sharing the LLM agent's full dialogue context, allowing it to reason over the policy in light of the entire conversation. It provides conversation-specific, actionable feedback to guide the agent's next turn, rather than just blocking an action. This approach acknowledges that policy adherence often depends on explicit user confirmation, prerequisite information, and the overall dialogue flow. Evaluations across multiple LLM vendors (GPT-5.4, Claude Sonnet 4.6, Gemini 2.5 Pro) on an airline domain benchmark showed significant improvements. PolicyGuard increased PASS4 scores by up to +12.0 percentage points, demonstrating higher policy-violation recall while blocking actions less frequently than traditional argument-level guards. This indicates a more nuanced and effective approach to ensuring LLM agent compliance.