AI guardrails sit around a model or an agent and apply rules at run time. Input guardrails screen prompts, for example by detecting personal data or requests outside an approved topic. Output guardrails check responses for format errors, policy violations, or claims that conflict with the provided sources. For AI agents, guardrails also limit which tools an agent may call and when a person must approve an action. Teams usually combine several kinds: fixed rules, classifiers, and a second model that reviews the first.
A bank chatbot, for example, might refuse to give investment advice, hide account numbers in replies, and pass refund requests above a set amount to a human. Guardrails govern behavior at the edges of a system and are one part of AI safety practice. Whether the data the system relies on fits the task is a separate question that guardrails do not answer.