What are AI guardrails?
Guardrails are the rules and checks that keep an AI system inside safe and approved limits, such as blocked topics, required formats and actions that need approval.
Key takeaways
- Guardrails can sit in the instructions, in the software around the model, or in both.
- Typical guardrails limit what the AI may say, what data it may see and what it may do.
- Instructions alone are a weak guardrail. Hard limits in software are stronger.
How it works
Some guardrails are instructions in the system prompt, such as "never quote a price". Stronger ones are built into the tool. A filter checks input and output, a permission list limits which tools the agent may use, and an approval step stops risky actions until a person agrees.
Example
A firm puts an AI assistant on its website. The assistant can answer questions from the public help pages. It cannot see customer records, and it must pass any refund request to a human.
Why it matters
Before you let AI talk to clients or act in your accounts, decide what it must never do. Then make those limits real in the tool settings, not only in the prompt.
Common mistake
People rely on a polite instruction to stop a bad outcome. If an error would be costly, remove the ability, or require approval.
Related terms
See all terms in the AI Encyclopedia, the numbers on the AI statistics pages, or build a prompt with the prompt builder.