Illustration showing a phone call audio waveform and dining icon connecting through four colored pathways to the Gemini sparkle icon surrounded by calendar, map pin, email, receipt, and phone task icons. How Google Duplex evolved to also call businesses for you in Gemini

November 20, 2026

Yossi Matias, Vice President, Google & GM, Google Research, Yaniv Leviathan, Google Fellow, Google Research, and Janet Zhou, Group Product Manager, Google Maps

When we introduced Google Duplex in 2018, it pioneered natural voice conversations between AI and local businesses over the phone. Since then, this technology has worked behind the scenes across many Google products. These capabilities are integrated directly in Tasks in Gemini.

Quick links

Completing tasks like making dinner reservations or booking a hair appointment is time consuming. Live phone conversations move fast and rarely follow a script. The person on the other end can say almost anything, often revising their thoughts mid-sentence – and people often rely on subtle tone changes to indicate whose turn it is to speak. Pause half a second too long, and it sounds like the call dropped; speak too soon, and you end up talking over each other. Leveraging AI to complete these tasks respectfully over the phone required us to consider high quality speech understanding, contextual reasoning, and speech synthesis working together in a single, low-latency system.

We introduced Google Duplex nearly a decade ago to help address this challenge. Propelled by advances in foundation models, the Duplex architecture has evolved to power autonomous task completion across many core products, such as Google Search and Maps. This week, we launched Duplex in Tasks in Gemini App, 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. In this post, we look back at the decade-long architectural evolution that made this possible — moving from early, closed-domain models to a unified stack built entirely with Gemini.

From closed-domain calls to Tasks in Gemini

What began as a narrow, task-specific system quickly expanded into a shared conversational infrastructure, built in collaboration with teams across Google.

During the COVID-19 pandemic, when business hours and services shifted rapidly, Duplex doubled down on verifying merchant info automatically to keep listings accurate. It remains an important data source; for example, business information sourced from Duplex’s phone calls had been shown in Search and Maps over one trillion times, as we shared in 2023.

Simultaneously, we worked with Android and hardware teams to bring these conversational models directly on-device. Starting in 2019, we rolled out new Pixel Call Assist features across five consecutive Pixel generations. These models continue to power features like Call Screen and Hold for Me, helping users filter unknown callers and speak with a live agent when a hold queue ends until the user rejoins. 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.

Three generations of conversational architecture

Over the past decade, the stack behind these calls evolved through three distinct eras, moving from bespoke neural pipelines to a unified architecture using Gemini.

1. Pre-LLM closed-domain models

We started work on Google Duplex before modern language models and even before the transformer itself was invented. In these early days, to ensure businesses and users had a great experience, we constrained Duplex to narrow domains and trained custom recurrent neural network models on anonymized, task-specific phone conversations. Without pre-trained world knowledge, we taught our models how to handle every specific type of task efficiently. In practice, this meant relying on extensive training data for each task type — such as interpreting the rigid flow of restaurant booking versus checking holiday hours — to ensure the system could reliably navigate domain-specific logic.

2. Early generative models and hybrid stacks

Google Duplex started employing open domain foundation models in production as they emerged. For example, in 2022 Duplex used LaMDA in live calls. These pre-trained models simplified comprehension and common-sense reasoning, allowing the system to handle unexpected conversational turns and tangents, while remaining grounded in the context like a standard restaurant booking.

However, these early generative models were less predictable and sometimes slower than today's, so we built hybrid stacks that paired generative models for flexible phrasing with deterministic state tracking and strict policy guardrails. This helped ensure that both users and businesses had a great experience.

3. Unbounded tasks and Gemini integration

As foundation models advanced, we removed the constraint of fixed verticals entirely by shifting to prompting the system with free-text task descriptions. This allowed models to define a task simply via a prompt, enabling the system to generalize to new task types and nuances (like asking for window seating or extra space for a stroller) and new business categories, without building separate dialogue models from scratch.

Our calling stack uses Gemini to unify multimodal understanding, planning, and dialogue generation so the system can navigate multi-constraint requests — like scheduling car maintenance around calendar conflicts — and adapt when a receptionist suggests another slot.

Technical hurdles in live telephony

Even with frontier models, live telephony poses hurdles that pre-training on standard datasets like the web alone doesn't solve:

  • Pragmatic grounding for pre-call and conversational context, and real-world intent understanding: The web doesn’t have all the data needed to capture how calls unfold with busy merchants. Nearly a decade of experience with goal-directed calls helps our stack grasp pragmatic norms, starting even before the call. In practice, this means matching a free-text task with a business based on the historical context of past conversations (e.g., knowing a specific vendor doesn't take reservations for large groups over the weekend) to predict whether a task will work. During the call, this helps the model determine what makes sense to ask at a given moment, how to read polite indirect refusals, and how to keep calls brief and effective.
  • Sub-second latency in noisy acoustic setups: Calls run over compressed phone lines against loud backdrops like clattering in kitchens, espresso machines, street noise, and side chatter. Our stack is tuned for low-latency streaming inference so replies land at a natural cadence. Custom acoustic and conversation cadence models assist the core AI system with handling background noise and pauses, stop talking when expected, etc.

Designing for both ends of the line

Duplex has to work just as well for the business answering the phone as it does for the user. That principle shaped several design choices:

  • Clear disclosure. Every call opens by clearly disclosing that it’s an AI agent — specifically, an automated system calling from Google.
  • Businesses stay in control. Merchants can opt out of automated calls in their Google Business Profile settings, or simply say “Please stop calling my business” during a call.
  • No setup and zero adaptation. Duplex works through the phone a business already answers: no software to install, no booking system to maintain and no staff to train. Whoever answers can handle the call the way they would with any customer, even mid-rush with a line at the counter.
  • One conversation, many insights. Details gathered on one call can answer many later questions, so a business isn’t asked the same thing over and over. For example, Google Maps can show details that Duplex gathers about businesses, like happy hour specials, patio seating, pet-friendliness, a parking tip or even a weekly trivia night.

Conclusion

Many daily errands, local services, and urgent tasks still require speaking directly with another person to coordinate details reliably. By integrating Duplex’s capabilities into Tasks in Gemini App, we are taking another step toward closing this loop. Natural voice interaction becomes a seamlessly integrated modality that the agent can deploy whenever structured digital channels like online APIs or booking forms don't exist. Using our newest Gemini 4 models — including Gemini 4 Hydrogen and Gemini 4 Neon — Tasks in Gemini can now support more comprehensive workflows like analyzing utility bills, processing transactions securely via the Agent Payments Protocol, and coordinating with users over email and Rich Communication Services (RCS). Through these capabilities, we are bringing this research into the next generation of AI assistance.

Acknowledgements

We’re thankful to the Duplex team and to everyone across Google leading and involved in this decade-long journey.For driving the transition to Gemini-powered calling and making this launch in Tasks in Gemini possible, we're especially grateful to Yoav Tzur, Sasha Goldshtein, Rafael Goldfarb, Liran Peretz, Ritika Pathak, Sen Li, Ran Manor, Dekel Auster, and many others across the team.

Quick links