Voice agents are created at Brand → (your brand) → Voice. Click Create agent and the New Voice Agent dialog opens, where you can start blank, start from a prebuilt template, or use Generate with AI.
A voice agent is always tied to:
- One brand — the brand is where its knowledge, its privacy masking and its company details all come from.
- Zero or more phone numbers (the agent answers calls coming to these).
- One voice + language combination (the spoken side of the conversation).
You can have several agents on one brand — for example a customer support agent on the main toll-free number and a separate after-hours agent on another. Each brand keeps its own set; agents are never shared between brands. For what a minute costs and what limits apply, see Pricing & limits.
The agent list flags two things worth acting on. An agent with nothing routed to it shows "No phone number yet". An agent whose brand has nothing indexed shows "This brand has no indexed knowledge — calls will be answered from general knowledge only." Each row also carries a status of Active or Paused.
Defines who the agent is and how it opens a conversation.
Internal name shown across the dashboard — in the agent list and in inbox call threads. Pick something your team will recognise (for example "EN Receptionist · main toll-free").
The system prompt that defines how the agent behaves during calls. Tell it who it represents, the personality and tone to use, what it should and shouldn't talk about, and any non-obvious rules ("never quote prices for enterprise plans, end the call and let the caller follow up by email").
Read-only. Every call opens with a fixed sentence in the agent's language — in English, "This call is answered by an AI assistant." — before your greeting. It is the disclosure the EU AI Act (Article 50) requires, so it can't be edited, removed or skipped, and callers can't interrupt it. It changes with the Language setting on the Voice tab.
What the agent says right after the AI notice. Leave it empty to use the default greeting for the selected language. Keep it short and don't repeat that the agent is an AI — the notice already said so. If you have recording switched on, this is where your recording consent line goes.
Allow callers to interrupt the greetingboolean
When on, speaking during your greeting stops it and the agent starts listening. Off by default — the greeting plays in full first. Either way the AI notice always plays to the end before your greeting begins.
Write the prompt like you're onboarding a new hire, not like you're writing a chatbot prompt. The agent does best when given a clear role, a few hard rules, and explicit examples of when to stick strictly to what the brand's knowledge says.
The agent's operating language. It controls the default greeting and the language the agent replies in. See
Voice & language for the list of 17 supported languages — and for the two other "language" settings that are easy to confuse with this one.
One of 30 Gemini voices, filterable by name, descriptor and gender, and grouped into families. Press play to hear a sample. The voice itself is multilingual — the same voice handles every supported language, so picking a voice never locks you into a language.
Saving this tab triggers greeting pre-generation — the AI notice and your greeting are converted to audio in the chosen voice and language and cached, so the first seconds of a real call are instant. Changing the voice, language or greeting regenerates that audio automatically. The play button next to a voice previews the notice and the greeting together, exactly as a caller hears them.
Assigned phone numbersmulti-select
The numbers that route to this agent. A number can only be attached to one agent at a time; assigning it here detaches it from its previous owner. You can do the same thing from Settings → Organization settings → Phone numbers — both write the same assignment.
Saves a stereo MP3 of every call, with the caller on one channel and the agent on the other. Recordings are stored privately and capped at 100 per organisation — the oldest are removed first when it is full. Switching it on shows a compliance warning, because recording rules differ by country and use case.
There is no "which Vex bot should this agent use" field any more. A voice agent is a sibling of the brand's chat AI, not a child of it: both answer from the same brand knowledge. That is also why privacy masking and the company details come from the brand rather than from a bot. One thing does still depend on the chat AI: API tools and workflows are only offered to a voice agent when its brand also has a Vex AI. A brand without one still answers calls from its own knowledge, but the agent has no tools it can call.
Masking is switched on per brand, at Brand → (your brand) → Brand settings → Advanced → Privacy masking. It hides email addresses, phone numbers and card numbers.
One toggle, two behaviours — and on the phone it means the opposite of what it means in chat. In chat it strips personal details out of a message before it is sent to the AI provider. On a call it cannot do that: the caller's speech has already reached the AI in real time. What it redacts on a call is what we store — the saved transcript and the call's messages in the inbox. It is best-effort, never a guarantee: if the masking service is slow or unavailable, the text is stored unmasked and nothing records that it happened. It is not encryption, and calls recorded before you switched it on stay as they were.
Every inbound call this agent has handled, defaulting to the last 30 days and filterable by date. Each row shows the caller, date, duration, cost and status — Ended, In progress, Failed or Session lost — and lets you play the recording (if there is one) or jump straight to the conversation in the inbox.
The Generate with AI tile writes the name, the prompt and the greeting for you from a short description of what the agent should do. It pulls this brand's company name, website and tone automatically, so you don't repeat them, and it plays the generated greeting back before you commit. If the brand has no Vex AI yet, the panel warns "This brand has no AI agent yet" and the Generate button stays disabled — start from a template instead.
Click any agent in the list to open the same four tabs in edit mode. The audio cache regenerates whenever you change a field that affects spoken output (voice, language or greeting). You can also Pause an agent from here: while it is paused, incoming calls are not routed to it, and Resume turns it back on.
Editing the prompt does not affect a call already in progress — the agent keeps the prompt it started with for the rest of that call. The new prompt takes effect on the next call.
Deleting an agent no longer refuses while phone numbers are assigned to it. It unassigns them and resets their routing. Delete from the brand's Voice list and the confirmation also tells you how many numbers are affected; delete from inside the agent editor and it only asks you to confirm. Past call transcripts and recordings stay in the inbox. Deleting the brand deletes its voice agents too.
Unassigned is not the same as released. Those numbers stay on your organisation and keep costing their monthly rental until the organisation owner releases them at
Settings → Organization settings → Phone numbers. See
Phone numbers.