Before you start
Confirm these are ready:- STT/LLM/TTS provider keys in Settings.
- At least one outbound/inbound carrier number if the bot will be tested by phone.
- CRM variable names and sample customer data.
- Expected dispositions and post-call fields.
- Knowledge bases, if the bot must answer from uploaded documents.
Bot wizard steps
The CXB Console bot builder has six steps:- Identity
- Instructions
- Policy
- Voice Pipeline
- Behavior
- Tools & Integrations
1. Identity
Set:| Field | Guidance |
|---|---|
| Bot Name | Use a clear ops name: client, use case, language, and environment if needed. |
| Opening Message | The first thing the customer hears. Keep it short, natural, and variable-aware. |
| Re-engagement | Enable when silent customers should be prompted before the call is marked RNR. |
{{CUSTOMERNAME}}, make sure the CSV, API variables, or CRM pre-fetch really provides that field.
2. Instructions
If policy engine is disabled, the System Prompt drives the bot. Include:- who the bot is
- call objective
- language style
- what variables mean
- when to use tools
- when to end call
- how to handle objections
- what not to say
- very long legal scripts that sound robotic
- unsupported promises
- vague goals like “convince customer”
- field names that do not exist in CRM/CSV
3. Policy
Use Policy when the use case matches an available template, such as India collections. When policy is enabled:- CXB API generates the runtime system prompt.
- Manual system prompt becomes fallback only.
- CRM field mapping becomes critical.
| Field | Example |
|---|---|
| Company Name | DMI Financial Services |
| Agent Name | Kavya |
| Customer Name Source Field | CUSTOMERNAME |
| Total Due Field | due_amount |
| DPD Field | CURRENT_DPD |
4. Voice Pipeline
Configure STT, LLM, and TTS.| Layer | Common choices | Ops guidance |
|---|---|---|
| STT | Streaming STT, turn-detecting STT, multilingual STT | Use the turn-detecting STT engine only where its built-in turn detection behavior is desired and configured. |
| LLM | Multiple language model providers and the managed LLM platform | Use the model approved for the client and cost profile. |
| TTS | Multiple TTS engines (including Indian-language TTS) | Use client-approved voice, language, and speed. |
5. Behavior
Review:- max call duration
- voicemail handling
- dead-air / re-engagement behavior
- transfer behavior
- ending behavior
6. Tools & Integrations
Review:- End Call tool
- transfer targets (only
websocket_reverse_transferandlivekit_sip_refer— see Tools & integrations) - custom HTTP tools
- CRM pre-fetch
- post-call push
- SIP headers
- Knowledge Base RAG
- Agent Desk handoff (a separate path, not a transfer method)
- Callback detection toggle
Callback detection toggle
The bot-level callback toggle (callback_prompt_injection) makes the bot detect “call me back later” requests and emit callback fields in post-call analysis.
This is a double-gate. The callback analysis prompt is injected only when both are true:
- the bot has
callback_prompt_injectionenabled, and - the campaign running the call has
config.callback_detection.enabled = true.
Save and test
After saving:- Open the bot detail page.
- Use Test Call.
- Confirm the bot speaks the right opening message.
- Confirm variables render correctly.
- Complete one normal conversation.
- Inspect call detail: transcript, recording, events, analysis, QC, CRM push.
- Fix bot config before creating campaigns.