Gemini 3.8 Live Avatar Gives Enterprise Agents a Face

Your support agent could already talk. Now it has a face to go with the voice. On September 24, 2026, Google made Gemini 3.8 Live with Live Avatar generally available, and Gemini 3.8 Live Avatar puts a lip-synced video persona on a real-time voice agent. It is available in Gemini Enterprise only.

What shipped: Gemini 3.8 Live Avatar in Gemini Enterprise

The details below come from Google's blog post and the Google Cloud announcement.

• Availability: general availability since September 24, 2026, with US and EU endpoints, provisioned throughput, enterprise compliance and data governance.
• Voice: native speech-to-speech (the model works on audio directly instead of chaining transcription and text-to-speech), in 97 languages with automatic language detection.
• Avatars: a library of preset avatars. Custom avatars use system instructions plus a single reference photo and an audio sample, but only through an allowlisting and verification process.
• Tools: tool and API calls run in the background while the conversation continues. The model also takes live camera feeds and screen shares alongside audio.
• Watermarking: every audio and video stream carries an imperceptible SynthID watermark.

The Gemini 3.8 Live Extended Thinking variant is still in private preview, so don't plan around it.

Where a lip-synced agent fits on a service desk

This section is my read, not Google's. Google's own examples are customer service, interactive walkthroughs and hotel check-in.

A face earns its cost where the conversation is long, guided or multilingual:

• Guided troubleshooting: screen share input means the agent can watch the user's screen while it talks them through a fix.
• Onboarding walkthroughs: a persona that walks a new user or employee through a process feels less like a manual.
• Front-door intake in languages you can't staff: 97 languages is a wide net for a small support team.

It earns nothing on password resets or order-status checks, where text or plain voice is faster. Keep angry customers, refunds and anything with legal weight on a path to a human.

Background tool calls matter more than the face. A service desk agent lives on lookups in the ticketing system or the CMDB (configuration management database), and an avatar that keeps talking while the query runs beats one that freezes mid-sentence.

What Google hasn't published about Gemini 3.8 Live Avatar

• Latency: the announcement says near real-time and gives no numbers.
• Language quality: 97 is a count, not a quality score. Expect lip sync and accent quality to vary by language until you test them.
• Pricing: the announcement points to the pricing page and states no rates. Third-party write-ups quote per-minute prices and a later rate change. I couldn't confirm either on Google's pages, so I'm leaving the numbers out. Read the official pricing page and model a full session before you write a business case.

What to do on Monday morning

1. Check eligibility. Confirm you are on Gemini Enterprise and that the US or EU endpoint fits your data residency rules.
2. Try a preset avatar. Open the Gemini Enterprise console, then read the Configure live avatars page in the Live API docs. Take exact field names from there, not from a blog post, this one included.
3. Pick one flow and score it. Use a template like this one, which is my own worksheet and not a Google config format:

pilot: avatar-front-door
scope: <one flow, e.g. new-hire onboarding walkthrough>
languages: [<top 3 customer languages>]
measure:
  latency_to_first_response_ms: <measure yourself, report p50 and p95>
  lip_sync_quality_by_language: <native-speaker rating, 1-5>
  interruption_recovery: <native-speaker rating, 1-5>
  silent_seconds_during_tool_call: <avatar idle time while a lookup runs>
  containment_rate: <sessions resolved without a human>
  cost_per_resolved_session: <from your Cloud Billing data>
guardrails:
  on_screen_ai_label: true
  human_handoff_triggers: [<angry customer>, <refund over X>, <legal request>]

The numbers Google didn't give you are the ones this scorecard collects. Have native speakers rate each language, because that is the gap in the announcement.

4. Start the custom avatar conversation early. Custom avatars need allowlisting and verification, and Google says to go through your Google Cloud sales contact. Get written consent from whoever's likeness and voice you are using.
5. Label it on screen. SynthID is imperceptible by design, so it helps detection but tells the customer nothing. Put a visible AI label on the video.

A face is a new surface to test, disclose and budget for, so run it as a scored pilot before it becomes your default support channel.