TL;DR

Google has added a visual layer to its live voice model. Gemini 3.8 Live with Live Avatar generates a speaking video persona in near real time, lip-synced to the model’s speech, so an AI agent can appear as an animated face rather than a voice alone. It is now available in Gemini Enterprise, following the launch of Gemini 3.8 Live itself the week before.

What it does

The avatar takes in audio and video from the user at the same time and answers with synchronised speech, facial expression and turn-taking. Google pitches it at enterprise uses such as customer service and guided product walkthroughs.

Two features stand out. First, the avatar can call tools asynchronously, so it keeps talking while it fetches data in the background. Google’s demonstration shows it checking a guest into a hotel without pausing the conversation. Second, it handles 97 languages and can switch between them mid-conversation, adjusting its lip movements and expressions to match, and Google says this happens without the video quality degrading.

Businesses can choose from a set of ready-made avatars or build their own from a single high-quality reference image, keeping a brand’s look or a character’s identity. Custom avatars are restricted for now to enterprise customers on an allowlist.

The safeguards question

A product that turns one reference image into a talking, responsive face raises obvious misuse concerns. Google says every piece of audio and video Live Avatar produces carries SynthID, its invisible watermark, so the output can be identified as AI-generated. The allowlist gate on custom avatars suggests Google is also controlling who can create a likeness.

That matters in the UK, where AI-generated video of public figures is already a lure for fraud. This week a Carlisle family described how a fake Nigel Farage clip led an 82-year-old into a £35,000 scam. Watermarks help platforms and investigators spot synthetic media, but only when someone checks, and a fake made with a tool that adds no watermark leaves nothing to find. The EU has already made marking AI output a legal duty for the largest model providers; the UK has no equivalent rule.

Looking forward

For UK firms running contact centres, a multilingual avatar that can act on back-office systems mid-call is a real option to test. The questions to settle before deploying are practical ones: whether customers accept talking to a synthetic face, how the business discloses it, and what the allowlist rules demand of anyone building a custom likeness.