Guardrails

The word originally means the protective barrier along the edge of a road. In AI, it means a safeguard that checks incoming requests for dangerous content and filters outgoing answers for personal data or harmful material. It is often a separate checking program that runs alongside the model.

For agents, guardrails also set the range of what they may do, such as "files are read-only" or "a person must approve payments." If they are too loose, accidents happen. If they are too tight, even normal requests get refused and the AI is less useful. Getting this balance right is the hard part.

It is like the bumpers on both sides of a children's bowling lane. No matter where the ball rolls, it cannot fall into the gutter.

Guardrails are the first thing a company looks at when it puts AI into customer support or internal work. Stories about accidents caused by a wrong AI answer usually say these safeguards were broken through or missing.