Real-time voice AI can now keep talking while tools work in the background, raising the value of good integrations, clear handoffs and precise controls.
Introduction
Voice AI is moving from scripted menus toward conversations that can trigger useful work. Google announced Gemini 3.8 Live and an Extended Thinking version for developers building real-time voice experiences. The models can continue a spoken exchange while calling tools in the background, use visual context, handle structured identifiers and work across many languages. That opens practical possibilities for appointment booking, order status, field support and customer intake. It also raises the stakes. A voice agent that can act inside a CRM or ERP is no longer only a conversational interface. It becomes an operational user whose identity, permissions, data sources and failure modes need to be designed as carefully as those of an employee.
What happened
Google says the new Gemini Live models are available through its Live API and Google AI Studio. The company highlights asynchronous function calling, so a system can invoke an API without freezing the conversation, plus visual grounding and support for more than 97 languages. It also introduced Gemini 3.5 Transcribe for speech-to-text across more than 85 languages. Google publishes benchmark and pricing claims on the product page, but those figures come from the vendor and should be tested against each business’s accents, terminology, call quality and workflows. The important change is architectural: voice, reasoning and tool execution can now happen in one continuous interaction rather than through a brittle chain of separate services.
Why it matters for growing businesses
For an SME, a capable voice layer can extend service hours and remove routine calls from overloaded teams. A caller might confirm an appointment, ask whether an item is in stock or report a maintenance issue without waiting in a queue. Multilingual support is particularly useful for firms serving Belgium, Europe or Latin America. Yet spoken interfaces create special risks. Names, account numbers and addresses are easy to mishear. Background noise can alter intent. A fluent answer can sound authoritative even when the underlying record is stale. If the agent can issue refunds, change delivery details or expose account information, one recognition error becomes a business incident. The safe opportunity is controlled execution, not unlimited conversational autonomy.
Practical actions this week
Choose one narrow call type with a clear system of record, such as appointment confirmation. Map every step, including identity checks, fields read, fields changed, confirmation language and handoff conditions. Give the voice agent a dedicated service identity with access only to those records and actions. Require callers to confirm critical identifiers twice, and send a written summary after any change. Keep payments, refunds, contract changes and sensitive disclosures behind human approval. Test common accents, mixed languages, background noise, interruptions and unavailable integrations. Record structured tool calls separately from audio transcripts so investigators can see what the agent actually did. Publish an easy path to a person; trapping callers in automation is not innovation, merely a cheaper maze.
A four-week pilot
Week one should measure the current process: call volume, waiting time, handling time, repeat contacts, errors and abandoned calls. In week two, deploy internally with staff acting as customers and deliberately testing edge cases. Week three can open the agent to a small customer segment during staffed hours, with every transaction reviewed. In week four, compare outcomes by language and intent, not only as a single average. Track correct resolution, incorrect record selection, handoff quality, staff rework and customer complaints. Set stop conditions before launch, including any privacy exposure or repeated wrong action. The pilot succeeds only if customers complete the task accurately and staff workload falls without hiding unresolved work behind a high containment percentage.
How VDS Consultoria can help
VDS Consultoria can identify voice workflows that connect cleanly to CRM, Odoo and scheduling systems, then design the integration around minimum access and human approval. We map identity checks, data scopes, multilingual test cases, audit logs and recovery procedures before production. We also build dashboards that show completed outcomes, corrections and escalations rather than vanity metrics such as total conversations. For businesses already using Odoo, the ERP can remain the authoritative record while the voice layer performs only narrowly approved actions. The goal is a useful service channel that sounds natural without pretending uncertainty has vanished. Good automation should shorten the route to resolution, not make accountability harder to find.
Management checklist
Confirm who owns the service outcome and who owns the technical integration. List the exact actions the agent may perform, the data it may disclose and the events that force a human handoff. Verify that callers are told when they are interacting with AI where required. Check retention rules for audio, transcripts and tool logs. Confirm that the system can revoke access immediately and continue service manually during an outage. Review a random sample of successful and failed calls every week. Re-test after model, prompt or connector changes. Finally, compare total cost with the baseline, including implementation, telephony, model usage, review time and support. Voice AI earns expansion through evidence, not because it can say “certainly” with impressive confidence.
Source
Build real-time voice applications with Gemini 3.8 Live and 3.5 Transcribe, Google, 2026-09-15.