What are AI guardrails?
AI guardrails are the checks and limits around an AI system, such as input filters, output validation, permission scopes and approval steps, that keep its behavior inside defined bounds.
In more detail
A guardrail can be a prompt instruction, but the ones that hold live in code. Examples: a schema check that rejects output missing a required field, a rule that blocks any payment over a threshold, a list of tools the agent may call, a filter that strips card numbers before text reaches a model. Layer them. Input checks reduce bad data in, output checks catch bad data out, and action limits cap the damage when both fail. Guardrails add latency and sometimes block legitimate work, so measure false blocks as well as misses. Log every trigger. The log shows where the system is drifting before a customer does.
Example
A refund agent for an online retailer can issue refunds up to a set amount, only on orders under 30 days old and only to the original payment method. Anything else goes to a person with the case summarized.