Guardrails are controls placed around an AI system's inputs and outputs. Input guardrails filter what goes into the model — preventing prompt injection, restricting topics, validating format. Output guardrails check what the model produces — blocking harmful content, flagging low-confidence responses, enforcing format requirements, rejecting policy violations.
A banking AI that cannot make specific financial recommendations, blocks responses containing other customers' personal data, and flags any response below a confidence threshold for human review — is running within a guardrails framework. Guardrails are not a sign of a limited system. They are a sign of a well-governed one.
