Identity, instructions and AI model

Give the bot a precise role

Describe what the bot represents, what it can do and what it must never do. Prefer short, observable rules ordered by priority.

A good configuration specifies:

• how the bot introduces itself;
• the information it must collect;
• when it should suggest an action or transfer;
• which topics it must refuse or pass to a person;
• how it confirms sensitive information before using it.

Greeting and additional information

The greeting is spoken or sent at the beginning of the conversation. It should sound natural and must not promise an action that depends on an unverified integration.

Additional information provides stable context: opening hours, covered services, emergency rules or specialist vocabulary.

Language and time zone

The bot language controls its conversations, summaries and generated suggestions. The time zone is used to understand phrases such as "tomorrow", "this afternoon" or an appointment date.

When a bot is created, the user's time zone is suggested by default. Change it when the bot works for another country or region.

Choose an AI model

Available models depend on platform settings. A more advanced model may improve reasoning while changing conversation cost. Always test real scenarios before rolling out a model change.

Protection and inactivity

The configuration can restrict anonymous, repeated or abusive calls. The inactivity timeout first warns the caller, then closes the conversation if no speech is detected. Choose a timeout suited to the usual length of your callers' answers.

Quiet voices and short answers

BeAI automatically applies bounded gain to quiet voices before sending them to the voice model. This helps detect short confirmations such as “yes, that is me”. Silence and very low-level noise remain unchanged; processing adds no waiting period and does not modify the original recording. It applies to inbound and outbound calls, for bots configured through either the interface or API. It does not replace a clear microphone signal or guarantee model or external tool response times.

Gemini responsiveness and tool announcements

Gemini defaults to 300 ms of silence to detect the end of speech. Existing and new bots without a custom value inherit this setting. Explicit custom values take priority, including when advanced settings are enabled. Increase this delay in Gemini voice settings if the bot cuts off hesitations; lower it cautiously for faster exchanges. This is the detection delay, not a guarantee of total model or external-service response time.

The equivalent API is PATCH /api/v1/bots/{id}/voice, for example {"gemini_live_advanced":{"enabled":true,"silence_duration_ms":300}}. Without an override, the default applies. The same bot editing permissions apply in the interface and API.

The shared prompt asks the bot to speak a brief, relevant announcement before the first business lookup or action for the same request, such as “I’m checking availability” or “I’m saving your booking”. Further lookups, detail reads and authorized actions for the same request follow without repeating that announcement or the same recap. This includes a planning lookup followed by saving a note. Invalid action arguments are rejected before execution; the bot receives the error to correct them or explain the difficulty. The bot speaks again with the results, an explanation of a failure, or an essential clarification. System date and time lookup (get_time) and number formatting (pronounce_phone_number) stay silent and do not count as a business announcement. If a calendar lookup follows, its announcement comes after the time lookup has returned. Required confirmations still apply and success is only announced after the result. A goodbye suffices before hanging up; post-call internal reports remain silent. Business tool execution no longer blocks the voice engine from receiving audio and interruptions.

While a tool executes, BeAI automatically plays a quiet waiting sound during silence. The phone engine generates it without an extra tool or model decision, and stops it when AI speech resumes. It does not overlap speech and is not triggered for hangup or internal reporting. The sound is included in the call recording, without being added to the transcript or sent as model input.

In a campaign, a short identity confirmation should let the conversation continue normally. The bot can use references, balances and due dates already supplied in the campaign context without an unnecessary lookup. To explain billed services, it reads the exact invoice lines using authorized tools; an invoice header reference is insufficient. The same workflow applies to calls created through the interface and API.