The brand holds the knowledge
A voice agent has no knowledge base of its own, and it does not borrow one from a chatbot either. It answers from its brand — everything on Brand → Knowledge and Brand → Products: knowledge bases, documentation, uploaded files, website pages Yaplet has read for you, question-and-answer entries and your product catalog.
That is exactly the same material the brand's chat AI reads. There is no linking step, no picker, and no separate voice copy to keep up to date.
What the agent can do on a call
A voice agent has a small set of abilities:
- Look something up in the brand's knowledge. Offered whenever the brand actually has something indexed. The agent uses it for anything that could be specific to your company: products, pricing, services, policies, opening hours, contact details.
- Talk about your products. It can name your categories and the products in one of them, with their live prices, in the same most popular order your chat AI uses, and it can filter or sort when the caller asks. This is offered whenever the brand has at least one active product, and it does not need a Yaplet chat AI. Nobody hears an id or a link.
- End the call. Always available.
- Run one of your scripted workflows, or call one of your API tools. Offered whenever the brand has any that can run on a phone call. The brand does not need a Yaplet chat AI for this. See the conditions below.
What it knows about time, and what it cannot do
A voice agent knows the time, but a few things are still out of its reach:
- The agent knows the current date and time. It reads them once, when the call starts, in the time zone of the brand's team hours (without one, Budapest time for a Hungarian agent, otherwise UTC).
- It has no separate opening-hours setting. If you want it to know your opening hours, they have to be either written into its prompt or published as an article it can find. It can then compare them with the current time. It also knows the brand's team hours, but those are only when your human team is usually online, not your opening hours, and it brings them up only when a caller asks when they can reach a person.
- It cannot hand the caller to a person. There is no live transfer. When it cannot help, it says so plainly and gives the company's contact details from your knowledge, or names the company's website if it finds none. It takes a message or offers a callback only if you write that into When a caller asks for a person on the agent's page; when it then agrees something your team has to do, your team gets an alert and the conversation waits in the inbox.
What happens when a caller asks something
| Caller asks | What the agent does |
|---|---|
| "What are your shipping zones?" | Looks it up in the brand's knowledge and answers from what it finds. |
| Something your knowledge does not cover | Looks it up, finds nothing, and says it does not have that information, then gives the company's contact details from your knowledge. It never answers from general knowledge instead. |
| "Are you open right now?" | Looks it up and compares your hours with the current time. It answers well only if your opening hours are in your knowledge or in the agent's prompt — there is no separate opening-hours setting. |
| "Where is my order?" | Needs one of your API tools. If the brand has none, the agent can only say it cannot check. |
| "Hi, how are you?" | Answers directly, with no lookup. |
| Something covered in the agent's prompt | Answers directly from the prompt, with no lookup. |
Workflows and API tools do not need a chat AI
Workflows and API tools are offered on a call whether or not the brand has a Yaplet chat AI. Anything the brand has — its knowledge, its API tools and the workflows that can run on a phone call — is offered to its voice agent. Knowledge lookup and ending the call always work.
One consequence is worth knowing for a brand set up deliberately as phone-only: its calls are billed and transcribed exactly as normal, but they do not appear in the AI answer-quality reports, because those figures are kept per chat AI.
Two conditions apply to workflows specifically. A workflow only runs on a call if its trigger is User says, and if none of its steps rule out phone calls — a form, a link, buttons, a reply-time step, a request-agent step, collect-data, ask-a-question or a stop step all do. The Runs on column on Brand → Workflows tells you which of your scripts qualify. And a caller never counts as identified, because a phone number can be faked, so an API tool or workflow switched to Only for identified visitors never runs on a call.
What makes good voice knowledge
The same articles serve chat and phone, but a few habits make a real difference when the words are spoken:
- One canonical answer per topic. Three overlapping articles on the same question make the agent hesitate while it picks. Merge them.
- Spell out abbreviations. The agent reads what you wrote. "U.S. Postal Service" sounds like a company; "USPS" sounds like four letters.
- Lead with a one-sentence answer. On a call the agent keeps its replies short unless the caller asks for more, so it will use your opening line and drop the rest. Make that first line the answer.
- Go easy on lists. Heavily structured bullet points come out as a recital. Smooth prose sounds like a person.
If callers say the agent is vague on basic questions, the usual cause is a fragmented knowledge base — lots of short overlapping articles competing for the same query. Consolidating into one clear article per topic fixes it more reliably than rewriting the prompt.
When the agent's language differs from your content's
An agent set to Hungarian can answer from an English knowledge base without you doing anything. Every call starts in the agent's configured language, whatever the caller happens to speak, and the spoken reply comes back in that language. It changes language only when the caller clearly asks to continue in another one; a short phrase, a name, an unclear sound or even a whole sentence that sounds like another language does not make it switch.
What is stored, and what is masked
Both halves of the call are saved — the caller's side and the agent's replies — as a conversation in your inbox. If the brand has Privacy masking switched on at Brand settings → Advanced, email addresses, phone numbers and card numbers are stripped from that stored transcript and from the call's inbox messages.
Be clear about what that does and does not mean on the phone. It hides those details in what Yaplet keeps. It cannot hide them from the AI during the call, because the caller's speech reaches the AI as they say it. It is best-effort rather than a guarantee — if the masking service is slow or down, the conversation goes through unmasked and nothing logs that it happened — and it does not change conversations recorded before you switched it on.
Updating knowledge once, for every channel
Because it is one brand's knowledge, publishing an article makes it available everywhere at once: the chat AI uses it on the next message, the voice agent on the next call, the inbox on the next suggested reply, and AI social posting the next time it writes. You never maintain a separate copy for voice.
One signal to watch for: on Brand → AI agent → Phone, an agent's row in the list warns "This brand has no knowledge yet — the agent can't answer questions." If you see that, the agent is answering the phone with nothing behind it: it still greets callers and takes the call, but it answers questions only from its prompt, and for anything else it says it does not have that information.
What to read next
To add or check the material itself, see Add knowledge to your brand. To handle calls outside your working hours, see Set business hours for a phone number.