A support example
A customer-support system may allow an AI to draft a refund reply but not approve a refund. An output guardrail can require an order ID and policy citation; a tool guardrail can restrict which records the system can read; a human gate can stop the final action.
Where guardrails sit
- Input: detect prompt injection, sensitive data, or out-of-scope requests.
- Model request: enforce system rules and allowed context.
- Tool call: validate arguments, permissions, and spending limits.
- Output: check schema, citations, unsafe claims, and prohibited content.
- Workflow: require approval before irreversible actions.
What guardrails do not do
A guardrail is not proof that an answer is true. It can check whether a citation exists or a field is present, but the source itself may be wrong. Combine controls with reliable data, evaluation, logging, and human review for high-impact work.