Admin Settings → Guardrails lets you attach policy checks to AI interactions so that requests and responses are evaluated against your organisation's rules — before harmful, off-policy, or non-compliant content reaches a user or an external system.
What a guardrail does
A guardrail is a named policy rule that can be applied to chats, templates or workflows. When enforcement is active:
- Content is evaluated against the rule as it flows through the platform.
- Content that violates the rule is blocked, and the user sees a clear message such as "Your request was blocked by a security guardrail."
- Allowed content proceeds normally.
Typical policy families organisations encode:
| Family | Example rule |
|---|---|
| Content boundaries | No answers about topics the company excludes from AI use. |
| Data leakage | Refuse inputs/outputs containing secrets like credentials or customer data patterns. |
| Regulated domains | Block clinical, legal or investment advice while allowing informational help. |
| Brand and tone | Enforce required disclaimers on certain assistant outputs. |
Managing guardrails
| Action | Notes |
|---|---|
| Create / Edit | Define the rule and where it applies (chats, templates, workflows). |
| Delete | Remove a rule; enforcement stops immediately. |
| Dry-run this guardrail | Test the rule against sample content without enforcing, to see what it would block. Always dry-run before applying a rule to production users. |
Rollout approach
- Start broad but dry — define the policy, dry-run against realistic samples.
- Pilot group — enforce on a small user group; gather false-positive reports.
- Tune — refine wording; false positives teach users to route around controls.
- Enforce workspace-wide — and communicate the policy in your acceptable-use documentation.
- Review periodically — policies drift; so should guardrails.
Note:
Guardrails are one layer of governance, alongside Permissions (what users may do) and capability enablement (what exists). Use all three together.