Gemini Live Avatar hits GA: agents with a face

Your support agent can now look at a customer and talk back. Google moved Gemini 3.8 Live with Live Avatar to general availability in Gemini Enterprise on September 24 — speech-to-speech dialogue paired with streaming video of an animated person whose mouth and expressions track the words in real time. For retail and services operators, Gemini Live Avatar turns a question you could previously defer into one you have to answer: should the thing answering your customers have a face?

What actually happened

Google's announcement and the Google Cloud post lay out what shipped:

• Low-latency video generation alongside speech, targeting web, mobile and interactive kiosks.
• 97 languages, with lip-sync and expressions adapting across a language switch mid-conversation without the video degrading.
• Asynchronous tool execution. The avatar acknowledges a request and keeps talking while the API call or lookup finishes in the background, instead of going silent.
• SynthID watermarking woven into both the audio and the video output.
• A curated library of preset avatars, with custom avatar creation gated behind enterprise allowlisting and verification.
• US and EU endpoints, with provisioned throughput available.

Google names live deployments: Cox Automotive building a shopping assistant for Autotrader, Equal AI handling over a million calls a day across nine Indian languages, plus Salesforce Agentforce and Specs. Access runs through the Gemini Enterprise Agent Platform and the Live API.

Why video agents matter for your business

Async tool execution is the actual upgrade. Every voice agent we have shipped has the same failure: it stops talking while a tool runs, and three seconds of silence reads as a dropped call. A model that can say "let me pull that up" and mean it removes the most common reason people hang up. That capability is worth more to a small operation than the face is.

The face is a commitment, not a feature toggle. A voice bot is judged on whether it resolved the problem. A photoreal talking person is judged on whether it was honest. Before you ship one, decide three things and write them down: the avatar discloses it is AI in its first sentence, the handoff to a human is one clear request away, and the identity it wears is not a real employee unless that employee signed off in writing. The allowlist Google put on custom avatars exists because the failure mode here is impersonation, and your own controls should be at least as strict.

Kiosks are the near-term win, not the website. A visual agent earns its keep where the customer is already standing still and cannot type — a service counter, a parts desk, a lobby after hours. On a product page, a video head competes with the buy button. In a showroom at 7pm, it is the only thing open.

Budget for streaming video, not for tokens. Continuous video generation per concurrent session is a different cost curve than text or even voice. Price one session-minute at your real concurrency before you scope the rollout, and check whether your EU customers need the EU endpoint before, not after, the contract.

Sources: Google, Google Cloud blog.