Glossary
/
AI Agents

Guardrails

Rules that stop an AI agent from breaking price, credit or data limits.

Definition

Guardrails are rules that keep an AI agent within safe limits. They stop it from sharing data it shouldn't, making promises the business can't keep, or taking actions outside its role. In commerce, they also enforce trade rules like pricing, credit limits and discounts.

Key points

<ul><li>Example: a retailer asks for a 20% discount when the limit is 10%. The agent declines or passes the request to a person, and doesn't approve it.</li><li>Rules that touch money or stock should be enforced by the system, not left to the AI model's judgement.</li><li>Guardrails sit in the agent harness and are tested with evals before the agent goes live.</li></ul>