Guardrail types
Choose the right rule for exact patterns, personal data, topics, or business limits.
Each guardrail type handles a different kind of risk.
| Type | Use it for |
|---|---|
| Pattern | An exact word, code, or text format |
| PII | Structured personal data such as cards, emails, IDs, and phone numbers |
| Custom rule | A business rule described in plain language |
| Topic | Keeping the conversation within allowed subjects |
| Full PII | Context-based personal data such as names and locations |
Pattern
Matches specific text or a known format. Use it when the value can be identified with an exact pattern.
PII
Finds common structured personal data and either blocks or masks it. This is the recommended baseline for every assistant.
Custom rule
Uses AI to check a rule written in plain language, such as “Do not provide legal advice.” On outgoing messages, it can rewrite a reply instead of only blocking it.
Topic
Allows the subjects you list and blocks clearly unrelated requests. Use it for assistants with a narrow role.
Full PII
Adds context-based detection for details such as names, locations, identifying dates, and health-related identifiers. Use it when stricter privacy is required.
Combine rules carefully
You can use several guardrails in each direction. Start with the smallest set that covers the real risks, then test it with normal and edge-case messages. Too many or overly broad rules can block valid customer requests.