The Signal
A face on your AI agent raises the standard for the handoff
Gemini Live Avatar adds visual presence to conversation. The business test is whether customers finish the task and know when a person must take over.

The Signal · Customer systems · September 29, 2026
A convincing digital face can make an interaction feel more complete than the underlying work actually is. For a business adopting a visual agent, the first question should be what the customer can successfully finish.
Google announced Gemini 3.8 Live with Live Avatar on September 24. The release pairs live dialogue with generated video and supports tool calls in the background while conversation continues. Google states that it is available in Gemini Enterprise. Custom avatar creation is separately restricted to enterprise allowlisting.
That is a meaningful interface change. It is not evidence that adding a face improves bookings, customer satisfaction, or resolution in your business. The announcement does not establish those outcomes.
Presence and completion need separate scores
A demo can show a natural conversation while hiding the awkward part: the booking service was unavailable, the customer’s account could not be matched, or the requested change needed a person.
Evaluate conversational quality and operational completion separately. The agent should be understandable, responsive, and clear about its role. It should also report the true state of the task. Those requirements can conflict if the system prioritizes keeping the conversation smooth over admitting that it is stuck.
A friendly “you’re all set” is harmful when the appointment was never recorded. A less polished explanation that the booking is still unconfirmed is more useful.
Test one customer journey
Consider a hypothetical equipment-rental company exploring a visual assistant on its website. The proposed job is helping a visitor identify suitable equipment and request a pickup time.
Start with the point of friction. If customers struggle to understand dimensions or accessories, a visual explanation might help. If they already know what they need and simply want availability, a short form or voice conversation might be faster. The avatar should earn its place in the journey.
Give the pilot a limited task: explain a small approved catalog, retrieve availability, and prepare a request. Keep final confirmation tied to a successful response from the booking system. If the system only records an inquiry, the assistant must call it an inquiry.
Make the difficult cases ordinary
Before expanding the pilot, test interruptions, corrections, background noise, slow connections, and unavailable tools. Use people who were not involved in building the script. They are more likely to ask questions the design team did not anticipate.
For the rental example, try a customer who changes dates halfway through, asks for unavailable equipment, or uses a product nickname. Test what happens when two items have similar names. Check whether the assistant repeats the final selection accurately and whether the destination record matches it.
Include a customer who wants a person immediately. The handoff should preserve the relevant context without forcing them to restart. If no person is available, the interface should say what will happen next and avoid promising a response time the business cannot meet.
Keep an alternative path
A visual agent should not become the only way to complete the task. Some customers cannot or do not want to use audio, video, or a conversational interface. Provide a readable path and a clear route to human help.
Treat likeness and brand presentation as deliberate choices. An abstract or clearly artificial character may communicate the role better than a realistic persona. Whatever the appearance, explain that the customer is interacting with AI and do not imply that a particular employee is present.
The design goal is clarity. A customer should understand who is responding, what has actually been done, and what remains uncertain.
Measure the work, then the experience
For a pilot, define completion using the system of record: a confirmed booking, an accurately submitted request, or a successfully transferred conversation. Then examine abandonment, repeated explanations, incorrect confirmations, and review effort.
Do not call every human handoff a failure. A timely escalation may be the correct outcome. Conversely, an interaction that stays entirely automated can still fail if it records the wrong date or leaves the customer confused.
The practical takeaway is to test the job before investing in the face. Live Avatar creates a new presentation option. Use it where visual presence removes a real obstacle, and retain a simpler interface where it does not.
Related: Gemini’s connected apps and the action behind the integration.
Primary source: Google’s Live Avatar announcement, checked September 29. The rental scenario and evaluation procedure are hypothetical recommendations. No conversion lift or Xovion deployment is claimed.
The Signal
Read next.
Coding · Tools
Codex can review the change. You still need proof of the behavior.
The new code-review experience makes investigation easier. Use it to assemble evidence, while keeping tests, permissions, and release decisions explicit.
Trust · Risk
AI evaluators are moving inside the lab. Ask what they can report.
Anthropic’s Accenture partnership points toward deeper external scrutiny. Access, independence, and public evidence still need separate examination.