Salesforce Decode
Salesforcedecode
Back to questions
AgentforceBeginnerguardrailssafetycompliance

What are guardrails in an AI agent?

Real World Scenario

Legal asks what stops the agent from discussing competitor products or sharing internal pricing. You explain guardrails at a high level.

Expected Answer

• Guardrails are rules and filters that limit what the agent can say and do • Topic boundaries restrict which subjects the agent handles • Instructions tell the agent what to refuse, disclaim, or escalate • Action allowlists prevent unauthorized operations like deleting records • Output checks can block prohibited phrases or sensitive data patterns

Follow-Up Questions & Answers

Click to expand — each follow-up includes a direct, interview-ready answer

No. Effective guardrails combine instructions, topic design, action permissions, and human escalation. Prompt text alone is easy to bypass with clever user input.

Architect Perspective

Junior answer: guardrails = seatbelts for the LLM. Mention defense in depth, not just "we added a disclaimer."