What are AI guardrails?
Guardrails are the limits around an AI system: what it may say, which tools it may call, what a person must approve, and what data it must not touch.
Also called: guardrail
In plain English
A guardrail can be a rule in the prompt ('never invent a fee'), a product setting (draft only), or a workflow stop (a manager clicks send). One sentence in a prompt is the weakest kind. A tool that cannot send mail is a stronger kind. Test the fence, do not admire it. Try to make the assistant quote a price you did not provide. Try to make it send. If it can, the guardrail is a sentence, not a lock. Move the lock into the tool: remove the action. Write down the three failures you actually tested, and retest them when you change models.
How an owner would use it
Write three guardrails for any agent that touches customers: no invented numbers, no send without approval, no access to files outside the job folder. Review them when the offer or the staff changes.
Common mistakes
A paragraph of 'be careful' in a prompt is the weakest guardrail, and owners stop there. A guardrail you can trust removes the tool: the agent cannot send mail because send is not offered. Do not stack twenty soft rules and call it safety. Pick the three actions that would hurt, and make those impossible or human-approved. Review the rules when the offer changes, or the assistant will keep enforcing last season's policy.
A worked example
A bookkeeper
No tax guesses
This is an illustration, not a customer result. The assistant may list unmatched deposits from the file. A guardrail blocks it from stating a tax amount or emailing a client. Those two actions are not even offered as tools.
Related terms
Prompts and guides
Questions owners ask
What are AI guardrails?
Guardrails are the rules and product limits that keep an assistant inside the work you intended, including what it can say and which tools it can use.
How do I set guardrails?
Start with draft-only, a ban on invented prices, and a folder limit. Turn off tools that send, pay, or delete until a person has watched the drafts for a while.
Are guardrails the same as a content filter?
A vendor's content filter blocks some harmful outputs for the platform. Your guardrails are the business limits: no invented fees, no send, this folder only. You need yours even if the vendor has theirs.
Product guides are on AI apps for business.