03 / Tools & Environment
Guardrails
Rules and checks that keep an AI system inside acceptable limits: what it will discuss, what it can do, and what gets blocked or handed to a person.
How you’ll hear it
“What stops it giving financial advice?”
“Guardrails. That topic is blocked and passed to a person.”
Avoid
Relying on instructions alone. Enforce the limits in the system, not just in the prompt.
The longer version
Guardrails combine instructions, filters on what goes in and comes out, permission limits and routes for escalation. The strong ones are enforced outside the model as well as inside it, because a model can be talked out of an instruction but not out of a permission it never had.
In depth
Guardrails are the rules and controls that keep an AI system within acceptable limits: what it will discuss, what it is allowed to do, what it must never reveal, and when it hands over to a person.
They work in layers. Instructions tell the model how to behave. Filters check what goes in and what comes out. Permissions limit which tools and data it can reach. Escalation routes pass difficult cases to people. Each layer catches what the others miss.
The most common mistake is relying on the first layer alone. A model can be persuaded to ignore an instruction, but it cannot use a permission it was never given. Anything with real consequences, such as sending money, changing records or contacting customers, should be controlled outside the model.
Guardrails should be tested, not assumed. Red teaming shows where they fail before your users do.
What to do
- List what the system must never do, then check each item is enforced by permissions, not just instructions.
- Design the hand-over to a person as carefully as the automated path.
- Have someone try to break the guardrails before launch, and again after every major change.