# Guardrails

> Guardrails are the controls placed around an AI system to keep its inputs, outputs and actions within acceptable limits. They include input filters, topic restrictions, output validation, blocked actions, spending and rate caps, approval requirements and escalation rules. Effective guardrails are enforced in code and permissions around the model, not only through instructions in the prompt.

## Why it matters for your business

A language model will occasionally do something unexpected, so the real question is what the surrounding system allows it to do when that happens. Guardrails turn business rules into hard limits: an agent cannot refund more than a set amount, cannot email anyone outside approved domains, and cannot give medical or legal advice. They also include checks on output, such as blocking responses that contain account numbers. Buyers should ask which guardrails are enforced in code, which rely on the prompt, and how violations are logged and reviewed.

## Example

A healthcare practice's intake assistant collects symptoms and insurance details before appointments. Guardrails block it from offering diagnoses, detect phrases that suggest an emergency and immediately show instructions to call 911, mask insurance member IDs in logs, and prevent it from writing to the patient record until a staff member reviews the intake.

Related terms: https://www.lumeor.studio/glossary/prompt-injection/, https://www.lumeor.studio/glossary/human-in-the-loop/, https://www.lumeor.studio/glossary/ai-governance/, https://www.lumeor.studio/glossary/system-prompt/, https://www.lumeor.studio/glossary/hallucination/

---

Source: https://www.lumeor.studio/glossary/guardrails/