Guardrails keep an AI system within acceptable bounds. They include instructions that define its role and limits, checks on inputs for malicious content, checks on outputs for prohibited content or format errors, restrictions on which tools and data it can reach, and approval requirements for consequential actions.
No single guardrail is sufficient. Effective designs layer several controls and assume any one of them can fail. In sensitive areas, such as finance, health, or personal advice, guardrails also include escalation paths that hand the conversation to qualified people.