Interceptors
Interceptors are rule-based overrides for moments where you do not want to leave behavior entirely up to the model.
Interceptors
Interceptors are rule-based overrides for moments where you do not want to leave behavior entirely up to the model.
What Interceptors Are Good For
Use interceptors when you need a predictable response to a predictable trigger.
Common examples:
- escalate when a customer asks for a human
- block unsafe or unsupported requests
- tag conversations that mention a key topic
- force a standard reply for a known question
How They Work
An interceptor matches text using keywords or regex-style patterns, then runs one or more actions.
Supported action types include:
- respond
- escalate
- tag
- block
Action Types In Plain Language
| Action | Use it when |
|---|---|
| Respond | The agent should send a known reply every time this trigger appears |
| Escalate | A human should step in instead of letting AI continue alone |
| Tag | You want reporting, filtering, or follow-up later |
| Block | The request should not proceed at all |
Priority matters when multiple interceptors could match the same message.
If two interceptors could both fire, the higher-priority rule should be the one you trust most. For example, a billing-policy escalation rule should usually beat a generic friendly-response rule.
When To Use An Interceptor Instead Of Editing Sources
Use an interceptor when:
- the behavior should be mandatory, not optional
- a compliance or support policy must always win
- you need escalation or tagging, not just a better answer
Use sources when:
- the problem is missing knowledge
- the agent needs more context, not a hard override
Practical Workflow
- Find the repeated failure in Conversations or Feedbacks
- Decide whether you need a response, escalation, tag, or block
- Create the interceptor with the narrowest trigger that still catches the issue
- Test it in Playground
- Watch live conversations to confirm it fires as expected
Best Practices
- start narrow and expand only if you miss valid cases
- avoid overlapping interceptors unless priority is intentional
- keep rule names descriptive so other teammates know why they exist
- review old interceptors periodically so temporary fixes do not become permanent clutter
Related Docs
Conversations
Review chats, reply as human, and inspect voice sessions.
Playground
Test answers, retrieval, prompts, and overrides before production.
Agent Settings
Tune AI, retrieval, escalation, and voice defaults for one agent.