AI Agents

An AI Agent stores the instructions and voice configuration used by your calling and browser channels.

​

Create a voice agent

1. Open Configure > AI Agents and choose Create Agent.
2. Choose an Incoming agent for inbound connections and website voice widgets, or a Flow Agent for a visual workflow.
3. Enter a name, first message, and system prompt. Define what the agent should do, what information it can use, and how it should handle missing information.

​

Choose a Voice Provider

In Voice Provider, select an option available to your account. Provider selection controls the voice transport and compatible settings; it is separate from model selection.

​

Understand the provider options

The names and subtitles below match the Voice Provider cards. The direction column describes phone-call capability when the required number, account access, and provider configuration are ready.

​

Select a voice and model

Choose a compatible model, voice, and language from the lists shown for your selected provider. Model availability is separately controlled; selecting a provider does not enable every model.

​

OpenAI voice models

Use the model list returned by the application. OpenAI Realtime includes gpt-realtime-2.1 and gpt-realtime-2.1-mini in the current catalog; availability is controlled by active catalog entries and account access. Realtime models are separate from the language models used with ElevenLabs or Chat Agents. GPT Live 1 (gpt-live-1) is a separately gated OpenAI voice runtime with a GPT-6 Luna delegation backend. Its presence in source does not mean it is enabled for your account. Check the model selector’s unavailable reason. GPT Live SIP permits read-only tools and ending the call; actions such as transfer remain blocked where playback confirmation cannot be verified. Temperature is disabled for OpenAI Realtime voice models and for GPT Live on Realtime AI. Use supported conversation and attentiveness controls instead of assuming every provider accepts the same settings.

​

ElevenLabs speech configuration

The supported speech catalog includes Flash v2.5, v3 Conversational, and v4 Turbo, alongside supported earlier models. v4 Turbo requires explicit administrator enablement and compatible provider access. The effective speech model comes from the saved agent setting or administrator default. Deployments using the earlier settings editor expose a Speech model control with the administrator default and models allowed for the account. The current unified settings editor does not expose that selector. If the control is absent, ask your administrator to verify the effective model rather than looking for an invented menu. Choose the language from the current provider/model-specific list. A language supported by v3 or v4 may be unavailable with Flash. Preview a voice in Configure > Tools > Voice Library, then select it on the agent. See Voice Library. When available, voice settings include stability, similarity boost, and speed. Expressive Speech and response eagerness control different behavior: expression changes delivery; eagerness changes when the agent takes its turn. v4 Turbo uses prompt guidance for audio tags; its v3 Expressive Mode switch does not apply.

​

Understand inbound and outbound calling

Premium AI and Realtime AI use compatible Twilio phone-number configurations for supported inbound connections and outbound workflows. Premium AI via SIP uses ElevenLabs SIP for inbound and outbound calling. Realtime AI via SIP uses OpenAI SIP and supports inbound only. It does not support outbound calling. Do not select it for an outbound campaign or assume its capabilities match Premium AI via SIP.

​

Prepare numbers and SIP trunks

For phone calls, configure a compatible owned number and the relevant routing before testing. Use Deploy > Phone Numbers for available rentals and Deploy > Connections for supported inbound assignments. SIP providers additionally require an enabled SIP Engine feature, plan permission for the engine, a configured trunk, a registered number, and compatible agent assignment. Follow SIP Trunks and its provider-side routing instructions. A saved agent or trunk alone does not prove that a phone route is ready. Browser Talk to Agent tests do not require a phone number.

​

Configure advanced settings

Select existing Knowledge Base items from Configure > Knowledge Base to make them available to the agent. Depending on provider support, the editor also offers conversation style, attentiveness, language detection, maximum duration, ending behavior, transfer, appointment booking, privacy, monitoring, and guardrails. Enabling a setting does not make every channel support it. Realtime monitoring checks the ElevenLabs account used by the agent and may be unavailable. Recording and retention follow the saved provider privacy settings. Integration actions also require the connection and explicit action permissions described in Integrations.

​

Test with Talk to Agent

Save the configuration, then open Configure > AI Agents > Talk to Agent. Review the configured mode, provider, model, readiness messages, credits, and any recording consent before starting. Talk to Agent is a credit-consuming Voice Only browser test; it does not validate a phone route or real external actions. A workflow Test places a phone call and has separate requirements.

​

Save and manage agents

Save the agent after selecting knowledge and supported conversation settings. Switching Voice Provider can change valid models, voices, languages, and capabilities; review the resulting configuration before saving again. Search the agent list, edit an agent, or open Version History to inspect and restore an earlier saved configuration. Check connections, workflows, and widgets that use an agent before deleting it. Creation limits, credits, and available actions remain subject to your plan and permissions.