Google has released two new AI models built specifically for voice conversations: Gemini 3.8 Live, made for handling large volumes of calls cheaply, and Gemini 3.8 Live Extended Thinking, made for harder tasks that need real reasoning while still talking naturally.
The core idea is simple. Most voice bots today go quiet while they think, which feels awkward and breaks the conversation. Google's new models keep talking while they work in the background, using phrases like "let me check that" before pulling up an account, running a search, or completing a multi-step task, then narrating progress as it happens. That small change matters a lot for anyone who has ever been on hold with a robot that goes dead silent for ten seconds.
Google is also making a pointed claim about quality. It says its Extended Thinking model now sits in first place on the Artificial Analysis Speech to Speech Quality Index, a ranking that combines several tests measuring how well an AI holds a natural conversation and completes tasks by voice. Notably, Google highlights strong scores on a benchmark built by Sierra, a ten billion dollar customer service AI company founded by Bret Taylor, the former Salesforce co-CEO who now chairs OpenAI's board. Beating a rival on the rival's own test is a deliberate signal: Google is chasing the same customer service phone call market that Sierra, OpenAI, and a growing list of startups are fighting over.
The other new capability worth noting is language switching. The model automatically detects and moves between 97 languages in the middle of a single conversation. For any company that runs a multilingual support line, a travel booking desk, or an international sales team, this removes a real cost: routing calls by language or paying for separate language-specific staff.
The business impact here is not abstract. Research firms tracking this space estimate that a large share of routine customer service work will shift to AI agents within the next few years, with junior support roles facing the steepest cuts. That does not mean every call center job disappears. It means the routine, repetitive calls, password resets, order status checks, simple bookings, get automated first, while harder or more emotional calls stay with humans, at least for now.
There is a real risk side to this too. As voice AI gets better at sounding human, voice cloning scams targeting executives and families have already become a documented fraud pattern. Google addresses this by watermarking all generated audio with something called SynthID, an invisible marker meant to prove the audio came from AI. That is a reasonable safeguard, but it only works if every company adopts similar labeling, and today many do not.
For any business that operates a phone line, a booking system, or a support desk, the practical move is to start testing these tools now on low-risk, repetitive calls rather than waiting. The technology is arriving fast, the price is dropping, and competitors in your industry are very likely already testing it.