Short answer
Input and output guardrails are the two enforcement surfaces around an AI model: input guardrails screen what reaches the model, and output guardrails screen what leaves it before a customer sees it.Input guardrails inspect the incoming message and its context before the model reasons about anything. They catch prompt injection attempts, requests outside the agent's scope, and personally identifiable information that should never enter the context window. Output guardrails inspect the drafted response and any proposed actions before they ship: hallucination checks, policy screens, tone review, and blocks on unauthorized commitments such as out-of-policy refunds.
The framing to reject is one-sided protection. Some teams filter inputs and assume a clean prompt produces a safe answer; others moderate outputs and let anything in. Both halves fail alone: an innocent question can still draw a policy-violating answer, and a crafted injection can slip past input filters and reveal itself only in what the model does. The two surfaces catch different failure classes; neither substitutes for the other.
Input guardrails vs output guardrails at a glance
| Dimension | Input guardrails | Output guardrails |
|---|---|---|
| What they inspect | The incoming message and context, before the model sees it | The drafted response and proposed actions, before the customer sees them |
| What they block | Prompt injection, off-scope requests, PII entering the context | Hallucinations, policy violations, off-tone replies, unauthorized commitments |
| Example | Hidden "ignore your instructions" text in a message is flagged and neutralized | A drafted reply promising an out-of-policy refund is held and escalated |
Aide, the AI platform for customer service, enforces both surfaces through its Agent Governance Engine: every conversation is screened on the way in, every answer and action is checked on the way out, and anything that fails a check is blocked or routed to a human.