AI news
Google announces Gemini 3.8 Live for voice agents
For a business considering a voice assistant, the useful question is whether a conversation ends with the right action. Our response to this announcement is a practical pilot brief, rather than a recommendation to replace an existing customer workflow.
Announcement: Sources checked:
We summarise the official announcement and distinguish it from our analysis and recommendations. Availability details may change after our source check.
What Google announced
Google announced Gemini 3.8 Live and Live Extended Thinking on September 15, updating its announcement on September 17. Google describes voice models with visual context and background tool use, positioning Extended Thinking for more complex tasks. These are vendor claims; Arczen has not independently tested the models.
References: Google: introducing Gemini 3.8 Live and Extended Thinking
Arczen’s take: test the action behind the conversation
We would evaluate a voice assistant against a specific operational job, such as preparing a service callback. A natural exchange is useful, but our acceptance decision would depend on whether the correct request reaches the correct person with the right details.
For that pilot, separate three states: the assistant heard the request, a connected system accepted it, and the business confirmed the outcome. Write the customer-facing wording for each state before connecting live systems. This gives the reviewer something concrete to compare against what the assistant actually says.
An illustrative pilot: a customer changes the request
Imagine a home-services customer requesting a callback for Tuesday, then interrupting to change it to Wednesday. This is an illustrative test scenario, not an observed model result. We would check the final saved request, the spoken confirmation, and the staff notification together.
| Situation to test | Evidence we would require |
|---|---|
| Customer corrects a date | The final record and confirmation use the corrected date. |
| Request connection fails | No false success claim; the customer receives a usable next step. |
| Customer asks for a person | A clear transfer or callback route with conversation context. |
Arczen’s pilot questions before choosing a model
Our proposed checklist applies to any candidate. Test with representative speakers and agreed scenarios, then retain the outcomes rather than relying on an impressive demonstration.
- Can the assistant recover when someone interrupts, changes a detail, or switches between Arabic and English?
- Which actions require confirmation, and which remain with your staff?
- What happens when the customer stops speaking, the connection drops, or a tool returns an unclear result?
- What information is retained, who can access it, and what recording notice does your approved process require?
- How will you measure correct completion, waiting time, unresolved requests, and the pilot’s actual operating cost?
Make the next decision from your own evidence
We would begin with a limited, supervised trial and compare it with the existing route for the same customer task. Keep successful examples and failures in the review. Expand only when the business can explain what worked, what still needs a person, and how problems will be handled.
Sources & further reading
Service guidance can change; check the linked documentation when implementing.