Upload a PDF — and people can call it and ask out loud

Here's the idea in one line: you upload your PDF — a schedule, a manual, your notes — get a phone number, and anyone can call and ask, out loud, whatever is in that file. The agent answers by voice. Not text in a chat — an actual voice on the line.

And here's the fun part — a year ago this wasn't this easy. "A voice assistant for my document" used to mean: speech recognition, then a model, then voice synthesis, then telephony — weeks of assembly. On July 1, xAI launched the Grok Voice Agent Builder: no code, right in the browser, in a couple of minutes. You get a free phone number, 25+ languages, near-instant replies even over background noise, $0.05 a minute. The whole build is: upload a file and write a couple of lines about who this agent is.

Why this one

This is the first project that lives in a real phone call, not on a screen. And it's personal right away: a number for wedding guests (when's the gathering, where to park — from your PDF), a spoken study buddy (you call, it quizzes you), an "apartment manual" for guests (how to turn on the boiler, where the trash goes). You turn a boring sheet of paper into someone you can call.

What you'll learn

• Voice as the interface. Input and output are speech, not text. For the first time you build something people talk to, not type at.
• Knowledge = your file. The agent answers not "from its head" but strictly from the uploaded PDF. That's RAG in its simplest form — and here that limit is the feature: fewer inventions, more truth.
• The agent's system prompt. Who it is, what it talks about, in what tone, and — crucially — what it does when the answer isn't in the file. One paragraph of instruction decides whether the agent is useful or starts making things up.

A ready starter prompt

Don't tell the agent to "answer questions about the document" — it'll answer ones that aren't in the file too, and invent things. Set the role, the limits, and the behavior when it doesn't know:

The strong prompt sets the role, clear limits ("only from the PDF") and a fallback for "I don't know" — so the agent won't invent what isn't in the file. That paragraph is exactly what separates a useful helper from a rambler.

What you end up with

You give your guests one number. Someone calls at eleven at night: "when's the gathering and where do I park?" — and the agent calmly answers by voice, from your PDF, in the caller's language. You never picked up, never got interrupted — and every question is already answered. From a plain file, you built someone people call.

Start with one file, get it to a live call — and you'll have not "another chatbot" but a voice people actually call.

Source: xAI — Grok Voice Agent Builder