What are AI guardrails?
Guardrails are enforced constraints on an AI system’s behaviour. The useful distinction is between guardrails that are asked for — instructions in a prompt — and guardrails that are enforced outside the model, where the model cannot argue with them.
Prompt-level guidance is worth having, but it is advisory. It degrades under unusual inputs, long contexts and adversarial phrasing, and it cannot be audited: there is no record proving the instruction was followed.
Enforced constraints — what is callable, which tenant, which credential, which actions need approval — hold regardless of what the model was persuaded to attempt.
Every boundary that matters is enforced outside the model: publication controls what is callable, the token controls the tenant, the vault holds the credential, and approvals gate execution.
Related terms
- Prompt injection — An attack where instructions hidden in data an AI system reads are treated as commands from its operator.
- Human in the loop — A design where a person approves or rejects a consequential action before it executes.
- Agentic AI — An umbrella term for AI systems that plan across multiple steps and act on the world, rather than answering one prompt at a time.
See it working rather than described.
The product demo runs one agent across three systems and stops at the approval gate.