A face-to-face AI video agent, explained

What you see on the Agents page

The Agents page carries a section headed “Ask our AI agent about Run AI, face to face.” It describes itself in one line: “A two-minute video call with an AI agent briefed on this company: the four divisions, how an engagement starts and how to reach the team. Speak to it the way you would on a call.”

• Nothing loads until you press “Start a 2-minute call”. Until then no request goes to the video-avatar provider and no microphone prompt appears.
• Your browser will ask to use your microphone. The call needs it: the agent listens to what you say.
• Its first words are the disclosure: “Hi, I'm the agent for Run AI Intelligence. I'm an AI, and I know this company well. What brings you here today?”
• The call lasts up to two minutes. The limit is set twice: in the provider’s own session cap, and by the page, which removes the call when its timer reaches zero or when you press “End call”.

Under the Start button, before any call, the page says in writing: “This is an AI, not a person: its face and voice are computer-generated. To run the call, your voice goes to our video-avatar provider; the privacy notice explains how.” The privacy notice is one click away, and it names the provider; this post does not, because this site keeps vendor names to the legal pages.

How it works, end to end

Five things happen between your question and the answer you hear:

1. The brief. Before the agent takes any call it is given a written brief: the same facts this website is built from (who Run AI is, the four divisions, how an engagement starts, how to reach the team), rendered into text and cut to what a two-minute call needs, with the rules in the next section. When the site changes, the brief is rendered again and replaced. The agent never knows anything the site does not say.
2. Listening. Your microphone audio goes to the video-avatar provider that runs the call, where speech recognition turns it into words.
3. Answering. The provider’s language model writes a short spoken reply from the brief, under its rules: one to three short sentences, one idea per turn, in your language (English, Hindi or natural Hinglish). If a detail is not in the brief, it is told to say so and to offer the person who has it.
4. Speaking. The reply is voiced and shown through a computer-generated face. It is not a recording of a person.
5. What stays where. The provider keeps a transcript of the call under its own privacy policy; Run AI does not record the audio or copy the transcript anywhere else. Who the provider is, where the data goes and on what basis are set out in the privacy notice.

The face is the least important part of that list. Most of the engineering is the brief and the rules, and both are text a business can read before the agent says a word.

What “done properly” means

A talking face is the easy part. The discipline is in what the visitor is told before the call, and in what the agent is told it may not do. This is the checklist the demo above follows, and the one to take into any vendor meeting:

• Disclosure first. It opens with “I’m an AI”, and the page says so in writing before the call starts. In the Voice page’s words: “Callers are told they’re speaking with AI wherever law or ethics require it — configured per market, logged per call.”
• Consent before anything loads. No request, no microphone prompt and no third-party storage until the visitor presses Start, and the privacy notice is beside the button.
• A hard limit. Two minutes, enforced twice: by the provider and by the page.
• Grounded answers only. The brief is rendered from the site’s own records, and the agent answers from it alone. “Answers only from your approved knowledge — never from imagination.”
• Scoped authority. The brief’s rules, in its own words: never state or hint at a price or a range; it comes in a written proposal for their scope; never promise a result, or pressure: no urgency, scarcity or fear of missing out; never ask for their name, number or email; never name a vendor, tool, model or platform, including the one it runs on; never say or suggest it is a human.
• A route to people. When they ask about cost, timing or how to start, offer one next step from the contact channels, once. It can name the contact page; it cannot take a booking, a payment or your details.
• A trail. For this demo the trail is the provider’s transcript. For an agent deployed in a business the rule is the Agents page’s: “Every action, decision and hand-off recorded — replayable end to end.”

The demo answers questions; it does not act in any system, and that is deliberate. On a public page there is no account for it to act in, and an agent that can act needs what the Agents page describes, not a microphone: “Every agent gets its own least-privilege service account — never a human’s password.” and “Irreversible actions require sign-off — payments, sends, deletions never run unattended.”

How a business gets one of its own

The same build serves three jobs. From the Voice page:

• Digital human presenters: A face for the brand: website avatars, kiosk concierges, livestream hosts — the same brain, now with presence.
• Founder & expert twins: Consent-based digital twins from one capture — your expert, available in 70+ languages, on every surface.
• The website: A live avatar that greets, guides and converts visitors in conversation.

The order of work is one role, one workflow, a metric attached. In the Agents page’s words: “One narrow workflow, one pre-agreed metric, then expansion. The anti-moonshot method — because that’s what survives.” The contained first step is a pilot: “One workflow, weeks to live, a pre-agreed baseline metric — the contained first step that proves the model.”

A voice role goes live in ~4 weeks, and it is rehearsed before anyone outside hears it: “Every release passes a written quality bar before a caller ever hears it.” Bring three things to the first conversation:

• The brief: who the agent is for, what it may say, and where it must stop. Write what it is not allowed to do first; the demo’s own rules are a fair template.
• Disclosure and consent settled, especially for a twin of a real person: “Every twin runs on written, revocable consent with a documented chain — no scraped voices, ever.”
• Data terms: “Client data never trains public models. PII-scoped, sandboxed, access-controlled — auditable end to end.”

A conversation starts with the role you have in mind. Commercials are tailored to scope and shared after a brief conversation.

Before you try the demo

• Use a device with a microphone; a laptop is easiest.
• Ask it what Run AI does, how an engagement starts, or which division fits you. That is what it is briefed on.
• Anything outside the brief is for the team; ask it for the contact page.
• Do not share personal or confidential information on the call. The privacy notice asks the same.

Questions

Is the face a real person?

No. This is an AI, not a person: its face and voice are computer-generated. To run the call, your voice goes to our video-avatar provider; the privacy notice explains how.

Do callers know they’re talking to AI?

Callers are told they’re speaking with AI wherever law or ethics require it — configured per market, logged per call. In regulated sectors the agent never forgets the script it’s required to say.

What happens when the agent isn’t sure?

Below the confidence threshold, the agent stops and asks. Judgment calls escalate to a person with full context. Human validation checkpoints are built in from day one — an agent that does most of the volume perfectly beats one that does all of it badly.

Can a caller always reach a human?

Any caller can reach a human in one utterance — warm handoff, full context attached. AI takes the after-hours, the overflow and the routine. People keep the moments that need people.

Who runs it after launch?

We do. Most voice pilots die after the demo because nobody operates them, so running the agent is our product, not your problem: call QA, tuning, drift watch, monthly reporting.