Gemini's Live Avatar: A Guide for Support and Sales Teams

Picture a customer support call that isn't audio-only: the AI on the other end has a face, holds expressions, keeps eye contact, and switches languages mid-sentence without losing its lip-sync. That's not a prototype anymore. On September 24, Google introduced Gemini 3.8 Live with Live Avatar, now generally available inside Gemini Enterprise, and it's worth understanding before a vendor pitches it to you as the next upgrade to your contact centre.

What just became generally available

Gemini 3.8 Live with Live Avatar combines Google's real-time voice model with a visual, animated persona that listens, sees and speaks back. According to Google's announcement and corroborating coverage from outlets including ITdaily, the avatar handles natural turn-taking, holds facial expressions while it talks, and performs "asynchronous tool calling" — meaning it can call a backend system to look something up while continuing to talk naturally, rather than going silent while it fetches an answer.

Google says the model can switch across 97 languages mid-conversation while keeping its lip movements in sync with whichever language it's speaking. Businesses get a library of preset avatars to start with; building a fully custom avatar is currently limited to enterprise customers through an allowlist process. Generated audio and video both carry Google's SynthID watermark, which is meant to make AI-generated calls identifiable as such if anyone needs to verify that later. A more advanced sibling, Gemini 3.8 Live Extended Thinking, remains in private preview and isn't generally available yet.

Two details matter more than the demo reel. First, general availability right now is scoped to US and EU endpoints — Google hasn't announced India-specific availability, data residency or pricing as of this writing, so treat it as new and not yet confirmed for Indian deployments. Second, published pricing for the video portion of a Live Avatar call is still being pieced together from third-party reporting rather than Google's own public pricing page, so get a written quote before you assume a specific per-minute cost.

What it's actually good for

The honest use case for a talking, expressive avatar isn't replacing your phone-based customer support queue. It's higher-stakes, lower-volume interactions where a face adds something real: a product demo that needs to show empathy, an onboarding walkthrough for a new employee or customer, a premium-tier support experience, or a sales conversation where a prospect is deciding whether to trust what they're being told. In those moments, an avatar that holds eye contact and reacts in real time is a genuinely different experience from a chatbot window or a disembodied voice.

It's a weaker fit, today, for the volume use case most Indian businesses actually have: thousands of calls a day in Hindi, Hinglish or a regional language, where customers code-switch mid-sentence and the cost per interaction has to stay low. Google's "97 languages" claim is about breadth, not depth in any one of them, and general availability hasn't been confirmed for India yet. That's a different, harder problem — one that's exactly why our sister company IndusLabs builds voice AI and workflow automation designed specifically for Indian languages, accents and code-switching, rather than a general-purpose model retrofitted for them.

Five checks before you pilot it

1. Confirm where the data actually sits

Ask your Google account team, in writing, whether India is covered by any current or planned endpoint, and where call audio, video and transcripts are stored and processed. "Generally available" so far means US and EU; don't assume India is included until someone confirms it.

2. Test the languages you actually need, not the headline number

"97 languages" is a capability claim, not a guarantee of quality in Hindi, Tamil, or Hinglish specifically. Run a real test with your own customers' phrasing, accents and code-switching patterns before you trust it with a live queue.

3. Get the real cost per interaction in writing

Video generation is billed separately from audio, and a few reports put the video component at roughly $1 per million output tokens — but that figure hasn't been confirmed on Google's own pricing page as of this writing. For always-on, high-volume use, that line item can matter more than the per-seat cost you're used to budgeting for software.

4. Decide who reviews what the avatar actually said

A transcript review process isn't optional for anything customer-facing. Sample conversations regularly, the same way you would for a new hire's first weeks, and make sure someone is actually assigned to look, not just empowered to.

5. Start with one low-stakes scenario, not your main support queue

A product walkthrough or an internal onboarding flow is a safe place to learn how the avatar behaves under real conditions. A live customer support or sales queue, where a wrong answer costs you a customer, is not the place to find out.

The bottom line

Gemini's Live Avatar is a real step in making AI voice interactions feel more human, and it's worth knowing about if you run a customer-facing team. But "generally available" currently means US and EU endpoints, unconfirmed India pricing, and a language claim that needs testing against your own customers before you believe it. Treat this the way you'd treat any new vendor capability: pilot it narrow, measure what actually happens, and only then decide whether it earns a bigger role in how your team works. If you're weighing where voice AI fits in your business and want an implementation partner rather than a vendor demo, that's the kind of decision we help work through in AI consulting — and our piece on what Mahindra Finance's voice AI rollout teaches you is a useful look at what a mature, India-specific deployment actually looks like at scale.