In plain English
Guardrails sit around the model and check what goes in and what comes out. They can block harmful requests, hide personal information, keep a chatbot on its job and stop an AI agent from taking risky actions without approval.
Example
A bank's support chatbot has guardrails that stop it from giving investment advice and strip account numbers out of its replies before customers see them.
Not the everyday meaning
Rails along a road that keep cars from going over the edge.
Why it matters
Guardrails are what let companies put AI in front of customers and connect it to real tools. They reduce the risk of harmful, off-brand or leaked responses, and they help teams meet legal and safety rules. As AI agents take more actions on their own, guardrails decide what needs a human's sign-off.
Common confusion
Guardrails vs. fine-tuning. Fine-tuning changes the model itself by training it on new examples. Guardrails leave the model alone and check what goes in and comes out. Most teams use both.