Google Gives Gemini 3.8 Live Agents Talking Avatars

Google has added talking, seeing avatars to its Gemini 3.8 Live voice models, formally launching the feature as Gemini 3.8 Live with Live Avatar for enterprise customers.

The move signals where Google thinks voice agents are heading: not just listening and speaking, but appearing on screen as visual personas that can handle sales support and customer service. In its announcement, Google described the result as "an experience that listens, sees and speaks with a dynamic visual persona."

The launch builds on work Google finished just last week, when it released the Gemini 3.8 Live and Live Extended Thinking models and called them its most advanced live dialogue models to date. At that release, Google framed the models as "building blocks for reliable, production ready voice agents" for "developers and enterprises." The earlier update also made speaking back to Gemini more natural, letting users carry out complex tasks with voice commands.

Live Avatar extends that formula with near real-time visual presence layered on top of the live dialogue models.

What the avatars can do

Videos in Google's blog post show both realistic and cartoon-like avatars. According to Google, they are capable of precise lip-syncing, natural expressions and "fluid turn-taking." The company says the avatars enable natural, multimodal conversations by communicating through facial expressions while simultaneously looking, listening and speaking.

Under the hood, the avatars draw on Gemini's advanced reasoning. Google says Live Avatar can trigger tool calls and fetch data in the background while continuing an active dialogue — meaning a customer-facing avatar could, for example, look something up mid-conversation without the exchange going silent.

Customization, but only behind an allowlist

Google will offer a library of what it calls "diverse, preset avatars." Organizations that want something bespoke can customize their own avatars from high-quality reference images, a process Google says preserves "reference likeness, brand styling or character identity."

Those custom avatars are available only through enterprise allowlisting, the company added — a restriction that matters given the obvious potential for misuse when a photorealistic AI face can be generated from reference images.

Google says Live Avatars ship with "strict safeguards designed to respect identity and keep AI-generated content transparent." Every avatar is watermarked using SynthID, Google's system for marking AI-generated content.

Enterprise-only — for now

Live Avatar is designed strictly for Google's Gemini Enterprise customers. Consumers will not get the feature directly. But the company's stated use cases — sales support and customer service — mean the avatars are likely to reach ordinary users anyway, appearing on computer or phone screens when they contact a company.

That prospect cuts both ways. On one hand, an avatar that maintains eye contact, syncs its lips and takes turns naturally could make automated support feel less like navigating a phone tree. On the other, the reception is far from guaranteed. Based on what Google has shown so far, and on the public's documented distaste for AI-generated content, the avatars may not be universally popular.

The stakes for Google are straightforward. Voice is becoming a primary interface for AI assistants, and every major lab is racing to make spoken interaction feel instantaneous and natural. Adding a visible, expressive face to a voice agent is Google's bid to make enterprise bots feel less like software and more like a presence — before rivals define the category.

For enterprises, the appeal is equally concrete: a single avatar that can see a user's screen context, speak, react with facial expressions and quietly call backend tools in the background could consolidate what today requires separate chat windows, phone queues and human handoffs.

The open question is whether end users will accept a synthetic face at all. Google has addressed the trust problem with SynthID watermarks and allowlist controls, but transparency measures do not by themselves make an avatar likable. If customer-service avatars land badly, Google and its enterprise customers will learn quickly — and publicly.