Customer automation is gaining eyes as well as a voice. Google Cloud has made Gemini 3.8 Live with Live Avatar generally available for enterprise use, combining real-time conversation, video or screen understanding and background tool calls. For growing businesses, that opens useful service patterns beyond the traditional chatbot. It also introduces more sensitive data and a greater risk that customers mistake a fluent interface for a trustworthy decision-maker.
What happened
Google Cloud announced general availability on 25 September. The service provides native speech-to-speech interaction, conversational video avatars, tool and API calling, automatic language detection across 97 languages, and the ability to process live camera feeds or screen shares alongside audio.
Google says the service is available through US and EU endpoints with enterprise compliance and data-governance features. Custom avatars remain allowlisted, while a curated library of pre-built avatars is available. Generated audio and video streams include SynthID watermarking intended to make synthetic media identifiable.
The company demonstrated an insurance claims intake scenario in which a customer talks and shows damage on camera while an agent team checks policy information, applies rules and prepares a packet for an adjuster. Google also cites customer and partner applications in vehicle shopping, voice assistance and customer service. These examples are vendor-provided demonstrations and statements; they should be treated as patterns to evaluate, not guaranteed outcomes for every deployment.
Why this matters for growing businesses
A live multimodal agent can reduce the gap between a customer describing a problem and a team understanding it. A technician can see an error light through the customer’s camera. An onboarding assistant can highlight the correct control on a shared screen. A sales agent can discuss requirements while navigating relevant options. This may shorten calls and reduce unnecessary site visits.
The same capability processes richer personal and business information. A camera can capture faces, addresses, documents, surroundings or bystanders. A screen share can expose passwords, notifications and unrelated customer records. Video avatars can also encourage users to attribute human judgment or authority to software. Governance must therefore be designed around collection, purpose and action, not added after the experience is polished.
Three practical use cases
1. Remote technical triage
A customer shows equipment, packaging, an appliance or a software screen while the agent follows a controlled diagnostic tree. The system can collect model numbers, identify visible indicators and create a structured service case. It should stop before safety-critical instructions or uncertain diagnoses and transfer to a technician with the captured context.
2. Guided onboarding and support
With consent, an agent can view a shared application screen, explain the next step and highlight where to click. This can help users complete forms or configure a product without giving the agent unrestricted control. Sensitive fields should be masked, and the agent should never ask the user to expose a password or authentication code.
3. Evidence intake
For insurance, maintenance, returns or property management, the system can guide a customer through capturing required views and documents. It can check completeness and prepare a case file, while a person or deterministic rules make consequential decisions. The agent should not infer facts that the image cannot reliably establish.
Guardrails to define before launch
- Minimise collection. Request camera or screen access only when needed, explain the purpose and end the stream promptly. Avoid retaining raw media when structured observations are sufficient.
- Separate assistance from decisions. Let AI guide, collect and summarise. Keep eligibility, liability, pricing exceptions, safety judgments and adverse decisions behind validated rules or human approval.
- Make identity clear. Tell users they are interacting with AI. Use approved avatars, preserve watermarking and prohibit designs that impersonate employees or real people without explicit authority.
- Protect the visual channel. Detect and mask sensitive information where possible. Provide a one-click pause. Warn users to close unrelated windows and keep authentication codes out of view.
- Constrain tool calls. Each action should use least-privilege credentials and server-side validation. The model proposes; the application enforces permissions, limits and business rules.
- Design for uncertainty. Low confidence, contradictory information, repeated failure or customer distress should trigger a handoff. The human should receive a concise summary and evidence links, not an unfiltered transcript dump.
- Set retention and access rules. Document where media, transcripts and derived data are stored, who may access them, and when they are deleted. Review cross-border processing and processor terms with privacy counsel.
Pilot and measurement approach
Start with prerecorded or employee test scenarios before live customer use. Include difficult lighting, background noise, weak connections, accents, interruptions and misleading visual inputs. Red-team the agent with content displayed on screen that tries to override instructions. Verify that tool permissions and backend validation prevent unsafe actions even when the model is confused.
Measure completion rate, time saved, transfer quality, correction rate, privacy incidents and customer satisfaction. Compare remote resolution against site visits or ordinary support calls. Do not optimise only for fewer human interactions; an early handoff can be the correct outcome.
How VDS Consultoria can help
VDS Consultoria can assess where multimodal service creates genuine value, map the data flows, connect approved actions to Odoo or other back-office systems, and establish privacy, security and human-approval controls. We can structure a limited pilot with measurable outcomes and a clear stop condition.
The interface may feel human, but the operating model must remain explicit: collect only what is needed, constrain what the agent can do, and keep accountable people in charge of consequential decisions.