Google Duplex is back, and this week it moved into the Gemini app. Tasks in Gemini can now phone a business on your behalf to book a table, schedule a service or ask about availability. Google announced the launch on November 20 alongside a look at how the calling system developed.
Duplex first appeared in 2018 as a way for AI to hold natural voice conversations with local businesses. Since then it has run quietly behind several Google products.
Why phone calls are hard for AI
Booking a haircut or a dinner reservation sounds trivial, yet it eats time, and automating it by phone is a hard engineering problem.
Live calls don’t follow a script. The person on the other end can say almost anything and may change their mind mid-sentence. Tone tells people whose turn it is to talk. Wait half a second too long and the other side thinks the call dropped. Jump in too early and the two voices collide.
Google says a working system needs speech understanding, reasoning and speech synthesis running together with very low delay.
From restaurant bookings to Search and Maps
During the COVID-19 pandemic, hours and services changed fast, so Duplex began verifying business details by phone. Google said in 2023 that information collected through Duplex calls had appeared in Search and Maps more than a trillion times.
Starting in 2019, the conversational models moved onto phones too. Google rolled out Pixel Call Assist features over five consecutive Pixel generations. They still power Call Screen, which filters unknown callers, and Hold for Me, which waits in a queue and alerts you when a live agent arrives. In 2025 and 2026, Search and Ask Maps gained agentic calling that checks quotes, product stock and service openings, from pet grooming to auto repair.
Three eras of the technology
Over about a decade, the system behind Google Duplex went from hand-built neural pipelines to a single Gemini-based stack.
The first era came before modern language models and before the transformer existed. Google trained custom recurrent neural networks on anonymized phone conversations, one narrow task at a time. Interpreting a restaurant booking and checking holiday hours each needed their own large training sets.
In the second era, Duplex began using open-domain foundation models. In 2022 it used LaMDA in live calls. Those models handled tangents better, but they were less predictable and sometimes slower. Google built hybrid stacks that paired a generative model for flexible phrasing with deterministic state tracking and strict policy guardrails.
The third era dropped fixed task categories. The system now receives a free-text description of the job, which lets it cope with requests such as window seating or extra room for a stroller. It also covers new business types without building a separate dialogue model. Gemini handles understanding, planning and dialogue in one stack, which suits multi-constraint requests like booking car maintenance around calendar conflicts. It can adapt when a receptionist offers another slot.
Built for the person who picks up
Google says Google Duplex has to work as well for the business answering as for the user, and that shaped several design choices. Many errands and urgent tasks still need a direct conversation, especially when no booking form or API exists.
What Tasks in Gemini can do now
Tasks in Gemini now runs on the new Gemini 4 models, including Gemini 4 Hydrogen and Gemini 4 Neon. Besides calling, Google says it can:
- analyze utility bills,
- process transactions securely through the Agent Payments Protocol,
- coordinate with users over email and Rich Communication Services (RCS).
What to watch
Google presents the move as one more step toward closing the loop on errands that still need a real conversation.
The launch makes voice a fallback channel inside a broader assistant: the agent uses structured channels first and phones only when none exist. Google hasn’t said which regions get the feature or how it is priced in this announcement. Watch for how businesses respond to more AI callers, and for whether other assistants copy the approach.
Source: research.google