Voice agents

A voice agent is the AI persona that picks up calls (inbound) or makes them (outbound). Each agent has a voice, a personality, a system prompt, and an optional knowledge base.

Anatomy of an agent

Writing a good system prompt

Short, concrete, written like a manager briefing a new hire:

• One sentence on who they are ("You're Zara, the appointment booker for…")
• 2-3 lines on what they should do ("Greet the caller. Ask their name. Find a time that works…")
• A few hard rules ("Never quote prices over the phone. If asked about pricing, say…")
• An end condition ("Once a time is booked, confirm and end the call.")

Avoid long preambles, fictional backstory, or "act like…" framings — they make the agent rambly. Treat the prompt as the job description, not a script.

Voice & personality

The 5 personality presets translate to different prompt prefixes:

• Professional — leads with the answer, one sentence of context, no filler.
• Friendly — first names, light warmth markers.
• Casual — contractions, occasional small-talk.
• Formal — drops contractions, uses Sir/Ma'am.
• Empathetic — validates feelings before redirecting.

If none of these fit, leave it on Professional and shape tone in your system prompt instead.

The voice library lives at /voices — every voice has a sample you can play.

Knowledge base

Upload documents (PDF, plain text) on the agent's Configure page → Knowledge Base. The agent's system prompt is augmented at call time with the relevant passages. Good for:

• Product specs / FAQs the agent should know.
• Pricing sheets the agent can quote from.
• Scripts / talking points.

Bad for:

• Real-time data (uploaded docs are static).
• Long-form documents — paste-or-summarize the relevant section instead.

Connecting to a phone number

A voice agent doesn't accept calls by itself — it needs a phone number wired to it. Two ways:

1. Inbound — Phone Numbers page → Connect Agent on the number row. All inbound calls to that number ring the AI.
2. Outbound — used by Campaigns and the developer API.

A single agent can handle inbound on one number and outbound for a campaign at the same time.

Recognizing returning callers

On an inbound call, the agent checks whether the caller's number is already known — a saved contact or someone you reached in a past campaign. If it matches, the agent greets them by name and has their email, notes, and history on hand instead of asking who they are. Put {{NAME}} in the greeting and a recognized caller hears their own name in the opening line. Unknown callers are handled normally. See Contacts → Caller recognition.

AI-only vs human-only mode

A phone number can be in one of two modes (see Phone numbers):

• AI mode — all calls go straight to the connected agent.
• Human mode — calls ring on-duty human agents in a workspace, with the AI as a fallback. Useful for hours when you want real people.

Switch via the AI ↔ Human toggle on the phone-numbers page.

Transferring to a human mid-call

If a caller asks for a real person ("speak to a human"), the agent triggers a transfer. Calls on a workspace number ring that workspace's on-duty team using its routing strategy; other calls ring on-duty members of your organization. The first to answer takes the call.

• Nobody is on duty on a campaign (outbound) call: the agent tells the person no one is available and carries on the conversation.
• Nobody answers: the caller is moved to the workspace fallback (voicemail, callback request, or AI).

To be reachable, open Live Agent and go On duty.

Appointment booking

Under Appointment Booking on the agent's page you control what the agent must collect before it books:

• Require email / Require phone — both off by default; name, date and time are always enough unless you turn these on.
• Timezone — the zone the agent uses for dates and times. Defaults to your organization's timezone.

The agent checks for conflicts before confirming and offers other times if a slot is taken. See Appointments.

Cloning a voice

Dashboard → Voice Cloning lets you create a custom voice from a recording.

1. Click Clone a Voice, name it, and add an optional description.
2. Provide an audio sample: record in the browser or upload an MP3, WAV or M4A file. Speak naturally for 1–5 minutes (at least 30 seconds); read an article or describe your day. Use a quiet room.
3. Submit. When the profile is ready, pick it as the agent's voice.

Only clone voices you have permission to use.

Limits per plan

Deleting an agent

Click the trash icon on the agent card. If the agent is still connected to a phone number or used by a campaign, you'll get a clear message telling you exactly which campaign is blocking deletion — reassign or delete that campaign first, then try again.