Google launches Duplex in Tasks in Gemini to call businesses for you

Google Research has published a post by Yossi Matias (Vice President, Google & GM, Google Research), Yaniv Leviathan (Google Fellow, Google Research) and Janet Zhou (Group Product Manager, Google Maps), dated November 20, 2026. Its news: this week Google launched Duplex in Tasks in Gemini App, described as a personal AI collaborator that can help coordinate real-world chores, like calling businesses for you and handling real-world requests on behalf of the user.

The authors start from why phone calls are hard. Live conversations move fast and rarely follow a script. The person on the other end can say almost anything, often revising a thought mid-sentence, and people rely on subtle tone changes to signal whose turn it is. Pause half a second too long and it sounds like the call dropped; speak too soon and the two sides talk over each other. Doing this respectfully, they write, requires speech understanding, contextual reasoning and speech synthesis working together in a single, low-latency system.

Google introduced Duplex in 2018, when it pioneered natural voice conversations between AI and local businesses over the phone. Since then it has worked behind the scenes across Google products. During the COVID-19 pandemic, when business hours and services shifted rapidly, Duplex doubled down on verifying merchant information automatically to keep listings accurate. It remains an important data source: information sourced from Duplex's calls had been shown in Search and Maps over one trillion times, a figure Google shared in 2023. Working with Android and hardware teams, Google also brought conversational models on-device. Starting in 2019 it rolled out new Pixel Call Assist features across five consecutive Pixel generations. Those models still power Call Screen and Hold for Me, which help users filter unknown callers and speak with a live agent when a hold queue ends. In 2025 and 2026, foundation models unlocked broader agentic calling in Google Search and Ask Maps, checking quotes, product availability and service openings, from pet grooming to auto repair.

The post describes three eras in the stack behind these calls. In the first, work began before modern language models and even before the transformer was invented. Duplex was constrained to narrow domains and used custom recurrent neural network models trained on anonymized, task-specific phone conversations. With no pre-trained world knowledge, each task type, such as the rigid flow of a restaurant booking versus checking holiday hours, needed extensive training data of its own.

In the second era Duplex began using open-domain foundation models in production as they emerged; in 2022 it used LaMDA in live calls. These pre-trained models simplified comprehension and common-sense reasoning, so the system could handle unexpected turns and tangents while staying grounded in a standard context like a restaurant booking. But those early generative models were less predictable and sometimes slower than today's, so Google built hybrid stacks that paired generative models for flexible phrasing with deterministic state tracking and strict policy guardrails.

In the third era, as models improved, Google removed the constraint of fixed verticals entirely. It now prompts the system with free-text task descriptions, so a task is defined by a prompt. That lets the system generalize to new task types and nuances, such as asking for window seating or extra space for a stroller, and to new business categories, without building separate dialogue models from scratch. The current calling stack uses Gemini to unify multimodal understanding, planning and dialogue generation, so it can handle multi-constraint requests like scheduling car maintenance around calendar conflicts and adapt when a receptionist suggests another slot.

The authors state a design principle: Duplex has to work just as well for the business answering the phone as it does for the user, and they say that principle shaped several design choices.

On the product side, they argue that many errands, local services and urgent tasks still require speaking directly with another person. Voice becomes a modality the agent can deploy whenever structured digital channels, such as online APIs or booking forms, don't exist. Using the newest Gemini 4 models, including Gemini 4 Hydrogen and Gemini 4 Neon, Tasks in Gemini can now support more comprehensive workflows: analyzing utility bills, processing transactions securely via the Agent Payments Protocol, and coordinating with users over email and Rich Communication Services (RCS). The post closes by thanking the Duplex team and naming Yoav Tzur, Sasha Goldshtein, Rafael Goldfarb, Liran Peretz, Ritika Pathak, Sen Li, Ran Manor and Dekel Auster, among many others, for driving the move to Gemini-powered calling and making the launch possible.

Key facts

  • Google launched Duplex in Tasks in Gemini App this week, so the agent can call businesses and handle real-world requests for the user.
  • Duplex dates to 2018; its stack went from custom recurrent neural networks trained per task, to hybrid stacks with LaMDA in 2022, to free-text task prompts on Gemini.
  • Business information sourced from Duplex calls had been shown in Search and Maps over one trillion times, a figure Google shared in 2023.
  • Voice is framed as a fallback the agent uses when APIs or booking forms don't exist; the newest Gemini 4 models, including Gemini 4 Hydrogen and Gemini 4 Neon, also let Tasks analyze utility bills, process transactions via the Agent Payments Protocol and coordinate over email and RCS.
  • Pixel Call Assist features, including Call Screen and Hold for Me, ran on on-device conversational models across five consecutive Pixel generations from 2019.

Why it matters

A voice agent that phones businesses now sits inside a mainstream assistant, not only in Search and Maps. The post presents this as closing a loop: many errands still need a conversation with a person, and voice lets the agent act where no online API or booking form exists. The decade of history is the substance: per-task models trained on recorded calls have given way to a single Gemini-based stack steered by free-text prompts, which is what lets it handle new task types and business categories without separate dialogue models.

Who it affects

Users of Tasks in Gemini App, who can hand over chores like dinner reservations, hair appointments or car maintenance scheduling. Also the businesses that answer these calls: the authors say Duplex has to work just as well for them as for the user. Pixel owners already use the on-device descendants of the same models in Call Screen and Hold for Me.

How to use it

The post describes Duplex as part of Tasks in Gemini App, where the agent can call a business for you and handle a real-world request. Its examples include booking a restaurant or a hair appointment, asking for window seating or stroller space, and scheduling car maintenance around calendar conflicts. With the Gemini 4 models the agent can also analyze utility bills, process transactions via the Agent Payments Protocol and coordinate with you over email and RCS. The post gives no availability details for Tasks in Gemini calling: no countries, languages, platforms, subscription tiers or pricing.

How solid is it

This is a first-party post from Google Research, written by named Google executives and researchers, so it is an authoritative account of what Google launched and how the stack evolved. It is also a launch narrative: it gives no call success rates, latency figures or user counts. The one-trillion figure is a 2023 number about business information shown in Search and Maps, not a current figure and not a count of calls. Nothing is said about Gemini 4 Hydrogen or Gemini 4 Neon beyond their names.

Risks and caveats

The authors themselves note that live telephony is hard: pauses, interruptions and mid-sentence revisions break naive systems, and early generative models were less predictable and sometimes slower, which is why Google used deterministic state tracking and strict policy guardrails. The post gives no explicit statement of how businesses are told they are speaking with an AI, nor of consent and disclosure policy. It also has no reactions from businesses, users or regulators, and no detail on how Agent Payments Protocol transactions work or which payment partners are involved.

“Duplex has to work just as well for the business answering the phone as it does for the user.”

— Yossi Matias, Yaniv Leviathan and Janet Zhou, Google Research blog