Governance
AI guardrails for your chats, agents and automations.
One switch in Admin, and Crewdle checks messages and replies against your policies. Watch for violations first, or block them. You see counts and reasons, never the message itself.
Updated October 2026
What AI guardrails do
AI guardrails check the messages that go into your AI and the replies that come back. They work at the organization level, so one switch in Admin covers everyone.
You choose which parts of Crewdle are checked, and whether a violation is only recorded or actually blocked.
Guardrails lower risk. They do not catch everything, and they do not replace your own review of important work.
AI guardrails are powered by Fireraven (FireGuard).
Two ways to run AI guardrails: managed or your own Fireraven account
Managed: Crewdle-defined policies, free for paying organizations. Or use your own Fireraven account (any organization).
-
Managed
Free for paying organizations
Crewdle defines the policies for you. There is nothing to set up in another tool.
-
Your own Fireraven account
Any organization
Connect your own account with Fireraven, the provider behind AI guardrails, and use the policies you set up in Fireraven’s dashboard.
What gets checked
Each of these can be switched on or off on its own.
| Where | What is checked | Good to know |
|---|---|---|
| Chat | Your messages and the AI’s replies | |
| Agents | Messages to your agents and their replies | |
| AI automations | The inputs and outputs of each run | Every run called through the API is checked when your organization has guardrails on. |
What counts as a violation, and prompt injection checks
A violation is anything that crosses the line your policy sets. You pick a criticality level, for example “High and critical”.
You can also add a security guardrail. It looks for prompt-injection and jailbreak attempts: attempts to trick the AI into ignoring its rules.
Monitor first, or block
-
Monitor only
Violations are recorded and nothing is blocked. A useful way to see what your policies would catch before you enforce them.
-
Block
The message or reply is withheld, and the person is told it was blocked by your organization’s AI guardrails.
AI guardrails activity: counts and reasons, never message content
The activity view in Admin shows how many checks ran, how many items were flagged and how many were blocked, for chat, agents and AI automations. It also lists the top reasons. It never shows message content.
Where checks are processed
Fireraven processes checks in Canada, mostly in Quebec, with some GPU servers in Toronto. If your organization is in the United States or Europe, the text being checked is still processed in Canada. Fireraven keeps the checked text for monitoring. It does not train on it.
This is the same when you use your own Fireraven account: checks are still processed in Canada.
AI guardrails FAQ
What are AI guardrails?
They check the messages that go into your AI and the replies that come back, against policies. You choose whether violations are only recorded or blocked. They lower risk but do not catch everything.
How do I turn them on?
In Admin, under AI guardrails. It is a switch for the whole organization. You then choose what to check (chat, agents, AI automations) and whether to monitor only or to block.
What does it cost?
Managed: Crewdle-defined policies, free for paying organizations. Or use your own Fireraven account (any organization).
Will it block my team’s messages?
Only if you choose block mode. In monitor only, violations are recorded and nothing is blocked. In block mode, the message or reply is withheld and the person is told it was blocked by your organization’s AI guardrails.
Can an admin read the messages that were flagged?
No. The activity view in Admin shows counts per surface and the top reasons, never the message content. The text being checked is processed by Fireraven, which keeps it for monitoring and does not train on it.
Where are the checks processed, and what does Fireraven keep?
In Canada, by Fireraven: mostly in Quebec, with some GPU servers in Toronto. For organizations in the United States or Europe, the text being checked is still processed in Canada. This also holds when you use your own Fireraven account. Fireraven keeps the checked text for monitoring. It does not train on it.
Do guardrails stop prompt injection?
An optional security guardrail looks for prompt-injection and jailbreak attempts. No filter catches everything, so keep approvals on for sensitive actions and review important work.
Are AI automations checked too?
Yes. When your organization has guardrails on, the inputs and outputs of AI automations are checked, including every run called through the API.
Want AI guardrails for your organization?
Talk to a person about which mode fits, or start free and look around Admin.
