---
title: "What are AI guardrails?"
excerpt: "Checks around a model that constrain what goes in and what comes out. The reliability layer between a clever demo and something you'd let touch a customer."
definition: "Guardrails are validation and policy checks placed around an LLM — on inputs, outputs, and tool calls — that constrain behavior, block unsafe actions, and enforce format or content rules the model cannot be trusted to follow on its own."
date: "2026-07-09"
lastModified: "2026-07-09"
author: "Teo Deleanu"
authorAvatar: "/team/teo.jpg"
tags: ["Guardrails", "AI Engineering", "Production AI"]
keywords:
  - "what are ai guardrails"
  - "llm guardrails explained"
  - "how to constrain llm output"
---

Guardrails are the checks that sit around a model — validating what goes in, what comes out, and which actions it is allowed to take. They exist because a model will, eventually and confidently, produce something wrong: an invalid JSON payload, a leaked secret, a tool call that shouldn't fire. Guardrails are the deterministic layer that catches it.

## Why it matters in production

The gap between an impressive demo and a system you would put in front of a customer is almost entirely guardrails. Input filters stop prompt injection and malformed requests; output validators enforce schema and policy; action gates require confirmation before a [tool call](/glossary/tool-calling) does something irreversible. Without them, one bad generation becomes one bad incident.

## The common mistake

Trusting the model to guard itself. "I told it in the prompt not to do X" is not a guardrail — it is a suggestion the model can and will ignore under the right input. Real guardrails are code that runs regardless of what the model decided, and they fail closed. The other mistake is guarding only the output while leaving inputs and tool calls wide open.

## How we actually use it

We validate output against a schema and reject-or-retry on failure, gate destructive actions behind explicit confirmation, and keep the checks deterministic so they behave the same on every run. Guardrails also produce the signals that feed [observability](/blog/silent-failures-observability) — a guardrail that trips is a failure you can see instead of a [silent one](/blog/why-ai-agents-fail-in-production). They pair naturally with [evals](/glossary/llm-evals): evals measure quality in aggregate, guardrails enforce policy per request.