On October 1, Tavus introduced Griffin, a video AI that watches, listens, and talks back in real time on a live video call. In a study Tavus ran with 54 people, each person had a one-minute call with Griffin. Afterward, 48 percent said they thought their partner was a real person. Tavus's previous system fooled 2.4 percent. Griffin also topped Nvidia's benchmark for real-time video conversations, scoring 3.83 out of 5 against a human reference of 3.92.
Griffin is not publicly available yet—Tavus is limiting access while it builds disclosure and safety features—but the test shows what is coming.
Your team could use this for the first call, the demo, or the intake form. Imagine a video call that qualifies a lead, walks someone through your service, answers the top ten questions, and hands the person to you only when they are ready to decide. Or a demo that runs at 2 a.m. in a time zone you do not staff. The person on the other end sees a face, hears a voice, and gets interrupted when they interrupt. You set the script, the tone, and the boundaries. The AI runs the call. You review the transcript and follow up with the ones that matter.
If half the people cannot tell, some will feel tricked when they find out. Test it, but also test the disclosure.
Decide now which questions a person answers and when you tell them. Write a list of the questions where your customer deserves a human: pricing exceptions, complaints, anything involving their kids or health, contract changes. If the AI cannot answer it, route the call to a person immediately. Also decide when you disclose. Some teams say it at the start of the call. Others say it when the caller asks or when the conversation hands off. The wrong answer is: never. If someone has to guess whether you are real, you have already lost their trust. Build the rule before you take the first call.