What are AI guardrails?
AI guardrails are controls that keep AI systems safe, accurate and within policy, such as filtering harmful inputs, checking outputs, limiting topics, enforcing permissions and requiring human approval for risky actions. They are essential before letting AI agents act in business systems.
Types
| Guardrail | Example |
|---|---|
| Input | Block prompt injection |
| Topic | Stay on product support |
| Output | Check for PII or policy violations |
| Grounding | Answer only from sources |
| Action | Approval for refunds over $100 |
| Access | Respect user permissions |
Best practices
Layer guardrails, test them with evals and red-teaming, and log every action.
