Plain English
Guardrails
Rules and limits that keep an AI tool within safe, agreed boundaries.
Limits built around an AI system to stop it doing or saying things it shouldn’t, such as blocked topics, required approvals, spending caps or checks on what goes out. Some come from the provider; others you set up yourself.
Why it matters
Instructions can be ignored or talked around; hard limits, like a permission the tool simply doesn’t have, can’t. Use both, and keep a person on anything important.
At work
A booking assistant can offer times from your calendar but can’t take payments or promise discounts, and it hands upset callers to a person.
Think of it like the bumpers at a bowling alley: they don’t throw the ball for you, they keep it out of the gutter.
Related terms
Where it appears