What a Voice Agent Knows

Where a phone agent finds its answers — the brand's own knowledge — what else it can do while a caller is on the line, and the one thing that decides whether it can run a workflow or call an API tool mid-call.

Where the Answers Come From

A voice agent answers from its brand's knowledge. Everything on Brand → (your brand) → Knowledge — knowledge bases, documentation sets, uploaded files, question-and-answer entries and pages of your website we read for you — plus the brand's product catalogue is searched in one pass on every lookup. It is exactly the same material the brand's chat AI reads.

There is still no separate "voice knowledge base" to maintain, but the reason has changed: the phone agent and the chat AI share the same brand, not the same bot. A brand does not need a chat AI at all for its phone agent to answer from that knowledge.

Practical consequence: anything you add on the brand's Knowledge screen improves chat and phone at the same time. There is nothing to copy across and nothing to re-link.

What the Agent Can Do on a Call

Look something up

Searches the brand's whole body of knowledge and its products in one pass. This is the default for almost every factual question — the agent is told to try it whenever a question could plausibly be about your company.

End the call

When the conversation is finished the agent can hang up by itself. Every agent has this, on every brand.

Run a workflow or call an API tool

Fetching live data from your own systems — an order status, stock, a booking — or running one of your scripted chat workflows. Offered only when the brand also has a Vex AI; see the next section.

Without a Vex AI on the Brand

Knowledge works. Workflows and API tools do not. Hanging up is the only ability every voice agent has unconditionally; the knowledge lookup is offered whenever the agent's brand has anything indexed at all, and a brand with nothing indexed yet gives its agent no lookup to make. Your scripted workflows and your own API tools are offered on a call only when the agent's brand also has a Vex chat AI. Most people never meet this limit, because an agent created on a brand that already has a Vex is linked to that bot automatically.

Calls handled by an agent whose brand has no Vex are saved completely normally — transcript, inbox conversation, recording, minutes billed. The one thing they do not produce is AI answer-quality reporting, because those figures are counted per chat bot and there is no bot to count them against.

This used to fail silently and badly. An agent with no chat bot behind it had every knowledge lookup refused, only the caller's half of the conversation was saved, and privacy masking was skipped — with nothing on screen saying so. All three are fixed.

API Tools Are Your Own Endpoints

"API tools" are not built-in helpers that read the caller's number or the time of day. They are endpoints you define yourself, which the agent calls mid-conversation to fetch something that changes.

They live in the lower half of Brand → (your brand) → Knowledge, under Other sources → API tools, and the brand's chat AI uses the same ones. In the tool editor the box you type it into is simply labelled Description; in the API tools list on the Knowledge page that same text appears under a column headed "When it is used" — and that column name is the hint about what to write, because the AI reads the description to decide whether to call the tool. Describe the situation, not the endpoint. The allowance is per brand: one on the free plan, twenty on a paid plan.

When the Agent Uses a Tool

The agent decides, turn by turn, whether to answer directly or call a tool. In practice:

  • Generic factual question ("what are your shipping zones?") → knowledge lookup.
  • Personal or account question ("where is my order?") → an API tool that takes the order number.
  • Something you scripted ("I'd like to book a slot") → the matching workflow.
  • Small talk, or something already in the agent's prompt ("are you a robot?") → a direct answer, no tool.

The agent says things like "let me check that for you" when it needs a beat. That pause is deliberate — it tells the caller that real work is happening.

If callers complain the agent is slow on basic questions, your knowledge probably has too many short sources competing for the same question. Consolidate into one canonical page per topic and the hesitation goes away.

Answering in the Caller's Language

The brand's material is searched in every language it is written in, up to three at a time — so an English knowledge base and a Hungarian one both keep their full search accuracy on the same call.

When the caller speaks a different language from the one your material is written in, the search wording is translated before the lookup so it still matches. The reference for that translation is the chat AI's content language; for a brand with no chat AI it is the brand's own language, set at Brand → Brand settings → General.

When Nothing Matches the Question

There is no control for this on the phone. What the agent does when the knowledge holds no answer comes from the agent's own stored value first, and from the brand's chat AI (Brand → Vex → Personality, When the AI Doesn't Know the Answer) only when that value is unset. Nothing in the interface can write the agent's own value, so in practice a brand with no chat AI always falls back to the phone default: the agent hedges and answers from general knowledge. And because a call cannot be handed to a person, the "Offer to connect with a human" choice becomes a plain, polite refusal on the phone.

Privacy Masking on Calls

Masking is a brand setting: Brand → Brand settings → Advanced → Privacy masking. It hides email addresses, phone numbers and card numbers.

It behaves differently on the phone than in chat, and you should know which. In chat it hides those details before they are sent to the AI provider. On a call the caller's speech has already reached the AI in real time, so what masking protects is what gets stored — the saved transcript and the call's messages in your inbox. It is best-effort, never a guarantee: if the masking service is slow or unavailable the conversation is stored as it was, and nothing records that it happened. It is not encryption. Calls recorded before you switched the toggle on stay as they were.

Knowledge

Everything a brand's AI can read, and what makes a good source.

Call flow and transcripts

What happens from greeting to hang-up, and where the transcript ends up.