Guardrails

Guardrails are the control layers around an AI system: checks on inputs, checks on outputs, limited permissions for tools and data access, hard limits for amounts and actions. The decisive part sits in the word "outside": an instruction in the `system prompt` is a request to the model, and clever inputs can override it. A guardrail is a rule in the surrounding software that holds even when the model tries to break it.

The boundaries to the neighbouring terms: the system prompt describes the desired behaviour, human-in-the-loop places a person in front of individual decisions, and guardrails enforce limits automatically on every single interaction. Together, the three form the control side of an AI agent: describe, constrain, escalate when in doubt. Typical building blocks are format checks, allowlists for permitted tools, amount limits, blocklists for topics and an escalation rule that hands unclear cases to a person.

In practice, guardrails decide whether a pilot is fit for production. In a pilot their absence goes unnoticed, because selected users test in good faith. In production they are the answer to the question of what happens in the worst case - and that is precisely the question every approval body asks. The proven direction: start narrow and open up deliberately once model evaluation shows a limit is no longer needed. The opposite direction - start wide and close down after the first incident - costs trust that is hard to win back.