Guardrails

Guardrails are runtime checks placed around an AI system to constrain inputs, outputs, and tool calls within safety, compliance, and business policy.

What is Guardrails?

Guardrails sit before and after the model. Pre-checks include prompt injection detection, PII redaction, topic allow-lists, and rate limits. Post-checks include schema validation, profanity and toxicity classifiers, citation verification, fact-grounding checks, and tool-call allow-lists. Guardrails can be implemented as classifiers, regex rules, validation libraries, or LLM-based judges. They are most effective when combined.

How does Guardrails apply to enterprise AI?

Enterprise AI without guardrails is uninsurable. The EU AI Act, GDPR, and sectoral rules effectively require pre-deployment risk controls, including filters on personal data leaving the organisation.