Google has integrated Duplex’s calling capabilities into Tasks in the Gemini app. According to the company, the personal AI assistant can now call businesses on a user’s behalf and handle requests such as arranging a restaurant reservation or hair appointment.
Live phone conversations are particularly challenging: participants may depart from an expected script, interrupt themselves or revise what they are saying, while subtle changes in tone often signal whose turn it is to speak. Google says a natural exchange requires speech recognition, contextual reasoning and speech synthesis to work together in a single low-latency system.
Duplex’s path to Gemini
Google introduced Duplex nearly a decade ago. Initially designed for narrow tasks, the technology later became shared conversational infrastructure across several products.
During the COVID-19 pandemic, the system was also used to verify rapidly changing business hours and service information. In 2023, the company said business information obtained from Duplex calls had been displayed more than one trillion times in total across Google Search and Google Maps.
Google’s Android and hardware teams also worked on applications that run on devices. Starting in 2019, the company introduced Pixel Call Assist features across five consecutive Pixel generations. The models still support features including Call Screen and Hold for Me: the former helps filter unknown callers, while the latter helps handle contact with a live agent when a hold queue ends until the user returns to the call.
In 2025 and 2026, foundation models enabled broader autonomous calling in Google Search and Ask Maps. These services can check quotes, product availability or service openings in categories ranging from pet grooming to auto repair.
Three stages of development
Work on Duplex began before modern large language models and even before the transformer architecture. The first system operated in narrow domains, using custom recurrent neural networks trained on anonymized phone conversations for particular tasks. Because those models lacked general knowledge from pre-training, they needed separate training data to handle jobs such as restaurant reservations and holiday-hours checks.
With the arrival of generative foundation models, Duplex moved to hybrid systems. By 2022, it was using LaMDA in live calls. Pre-trained models handled unexpected turns and tangents more easily, but Google says early versions were less predictable and sometimes slower. The company therefore paired flexible generation with deterministic state tracking and strict rules.
In the current system, tasks can be described in free text, eliminating the need to build a separate dialogue model for every new type of business or request. Gemini combines multimodal understanding, planning and dialogue generation. Google says this allows the system to manage details such as asking for a window seat or space for a stroller, or arranging car maintenance around calendar conflicts.
Noise, latency and business control
According to Google, pre-training on web data alone does not teach a system everything it needs to know about calls with busy businesses. Duplex can also use the context of past conversations to assess whether a request is likely to work. For example, it may take into account that a particular business does not accept large-group reservations on weekends. During a call, it must also interpret polite, indirect refusals while keeping the conversation brief and effective.
Compressed phone audio, kitchen noise, street traffic and background conversations add to the difficulty. Google has tuned the system for low-latency streaming; custom acoustic and conversational-cadence models help it recognize noise, pauses and changes in speaker.
Google says every Duplex call begins with a clear disclosure that an automated AI system from Google is calling. Businesses can opt out of automated calls through their Google Business Profile settings or ask during a call not to be contacted again.
The recipient does not need to install software or operate a separate booking system, and can answer on the phone the business already uses. Information gathered once can also answer later questions; Google Maps, for example, can display Duplex-sourced details about patio seating, pet-friendly venues, parking or weekly events.
Broader tasks in the Gemini app
Google says calling enables an assistant to handle requests even when an online booking form or application programming interface is unavailable. The company also said Tasks can support more complex workflows using its newest Gemini 4 models, including Gemini 4 Hydrogen and Gemini 4 Neon. These include analyzing utility bills, securely processing transactions through the Agent Payments Protocol, and coordinating with users by email and Rich Communication Services (RCS) messages.



