How Google Duplex evolved to also call businesses for you in Gemini - Google Research

How Google Duplex evolved to also call businesses for you in Gemini - Google Research

November 20, 2026

By Yossi Matias, Yaniv Leviathan, and Janet Zhou

When we introduced Google Duplex in 2018, it revolutionized natural voice conversations between AI and local businesses over the phone. Since then, this technology has become an integral part of many Google products, now directly integrated into Tasks in Gemini.

Completing mundane tasks like making dinner reservations or booking a hair appointment is time-consuming. Live phone conversations are fast-paced and often lack a structured script. The speaker's tone and pauses convey important cues, making it challenging for AI to navigate these conversations. Our solution required advanced speech understanding, contextual reasoning, and speech synthesis working seamlessly in a low-latency system.

Google Duplex, launched nearly a decade ago, aimed to address these challenges. Driven by advancements in foundation models, the Duplex architecture has evolved to power autonomous task completion across core products like Google Search and Google Maps. This week, we introduced Duplex in Tasks within the Gemini app, a personal AI assistant that can handle real-world chores like making phone calls and managing user requests.

This post reflects on the decade-long architectural evolution of Duplex, transitioning from a narrow, task-specific system to a unified conversational infrastructure built on Gemini.

Initially, Google Duplex was a specialized, task-focused system, quickly expanding into a collaborative project with various Google teams. During the COVID-19 pandemic, Duplex played a crucial role in maintaining accurate business listings by automatically verifying merchant information. This data source has been invaluable, with over one trillion instances of business information sourced from Duplex's phone calls displayed in Search and Maps since 2023.

Simultaneously, we worked with Android and hardware teams to bring these conversational models onto devices. Since 2019, Pixel phones have featured Call Assist features, utilizing models like Call Screen and Hold for Me to help users manage unknown callers and interact with live agents. In 2025 and 2026, foundation models enabled broader agentic calling capabilities in Google Search and Ask Maps, allowing users to check quotes, product availability, and service openings for various tasks, from pet grooming to auto repairs.

Over the past decade, the underlying stack of Google Duplex has undergone three distinct eras, evolving from bespoke neural pipelines to a unified Gemini architecture.

In its early days, Google Duplex employed custom recurrent neural network models trained on anonymized, task-specific conversations to ensure excellent user experiences. As foundation models emerged, Duplex adopted them, simplifying comprehension and common-sense reasoning. For example, in 2022, LaMDA was used in live calls, enhancing the system's ability to navigate diverse domains and logic.

Leave a Reply