How to Create Your Chat and Voice Assistant

Voice AI scales work that used to require real people — but only if it actually replicates how a person handles a call. This guide covers the assistant settings that make the difference between an assistant that sounds capable and one that doesn’t.

Before you start

Have your business information filled in and a knowledge base created if your assistant needs reference material. The AI drafts your assistant’s starting prompt by researching your business website and the details you’ve entered — you fine-tune it from there.

Handling background noise

Real phone calls have background noise, echo and other audio issues. The orchestration model handles most of this automatically, but two settings help in noisier environments:

• Responsive Speed (Queuing): set to 0.9 to make the assistant more deliberate, reducing false triggers from background noise.
• Sensitivity to Interruption (Noise Sensitivity): set to 0.8 to help the assistant tell real speech apart from background interference.

For very noisy environments — construction sites, busy streets — experiment with even higher settings.

Handling voicemail

Your assistant can detect voicemail and either hang up or leave a message automatically. The system runs voicemail detection continuously during a call:

• Enable Voicemail Detection in the assistant’s call settings, under “Enable Voicemail & Message.”
• Leave the message field blank to have the assistant hang up the moment voicemail is detected — its job is considered done.
• Enter a message to have the assistant wait its turn and leave that message, which can include variables and custom variables for context.
• If the detection timeout is reached without a voicemail being detected, the system stops trying and continues the call as normal.

Custom variables

In the “Make AI Call” action node, five custom variables let you dynamically change the prompt based on context or conditions. Practical uses:

• Feed historical interactions into a variable — pull context from a prior post-call webhook, transcript, form-fill or email trigger, and reference it in the prompt.
• Summarize and reuse context between calls — run the transcript through an LLM call after a post-call webhook to get a markdown-formatted summary, store it in a custom field, then feed that field into the custom variable on the next call. The first call has no prior context; every call after is contextually aware.
• Shift task order dynamically — structure your prompt’s task list with placeholders like {{custom.one}} so an outbound campaign can swap in different first steps per call.

Post-call tags

After every AI call, tags are automatically added to the contact to signify outcome and disposition — see AI call disposition tags explained for the full list. Use these instead of, or alongside, a post-call webhook to build campaigns based on how a call ended.

Pre- and post-call webhooks

A pre-call webhook fires to your specified URL when an outbound or inbound call starts, carrying to, from and contactId — useful for call tracking or start notifications.

A post-call webhook fires after the call ends, carrying the full transcript, call duration, sentiment analysis, disconnection reason, recording URL and more. See how to set up a post-call webhook for transcripts and workflows for the setup steps.

Next step

Once your assistant is created, write a prompt that actually sounds like a person: AI prompting guide — best practices. To deploy it, see how to deploy an inbound and outbound AI call assistant or how to create a voice widget for your website.