SIAgentNews.com
← All news

Google integrates Duplex calling into Gemini Tasks

Google is expanding its AI agent capabilities by allowing Gemini to make phone calls to book appointments and handle real-world errands.

· First reported by Google Research

Isometric diorama: picture for the story 'Google integrates Duplex calling into Gemini Tasks'
AI-made isometric diorama for this story. No real people or products are shown.
one trilliontimes business info from Duplex was shown in Search and Maps

The short version

  • Duplex is now integrated into the Tasks feature within the Gemini app.
  • The AI can handle multi-constraint requests, such as scheduling around calendar conflicts.
  • The system uses Gemini 4 models to manage speech understanding, planning, and dialogue.
  • Duplex has already powered business data for Search and Maps over one trillion times.

Why it matters

This moves AI from just answering questions to performing physical-world actions. It allows users to automate time-consuming phone tasks that lack online booking forms.

From narrow tasks to Gemini

Google is bringing its Duplex technology into the Tasks feature of the Gemini app. This update allows the AI to act as a personal collaborator that can coordinate real-world chores on behalf of the user. The primary goal is to handle time-consuming tasks like booking a hair appointment or making a dinner reservation.

The technology was first introduced in 2018 to pioneer natural voice conversations between AI and local businesses. Since its inception, Duplex has worked behind the scenes across various Google products. It was originally designed to navigate the unpredictable nature of live phone calls where people often revise their thoughts mid-sentence.

The transition to the Gemini app represents a shift toward more autonomous task completion. Instead of just providing information, the AI can now deploy natural voice interaction as a modality. This is particularly useful for errands where structured digital channels, such as online APIs or booking forms, do not exist.

The evolution of the calling stack

The architecture behind these calls has evolved through three distinct eras over the last decade. In the early days, Google used custom recurrent neural network models trained on anonymized, task-specific conversations. These models were constrained to narrow domains because they lacked the pre-trained world knowledge found in modern systems.

In 2022, the system began employing open-domain foundation models like LaMDA. These models improved common-sense reasoning and helped the AI handle unexpected conversational turns. However, because early generative models were less predictable, Google built hybrid stacks that paired them with strict policy guardrails and deterministic state tracking.

Today, the stack is unified entirely with Gemini. By shifting to prompting the system with free-text task descriptions, the AI can now generalize to new categories without needing separate dialogue models. This allows the system to handle specific nuances, such as asking for window seating or extra space for a stroller.

Technical hurdles of live telephony

Making an AI sound natural on a phone call requires more than just good speech synthesis. The system must manage high-quality speech understanding and contextual reasoning in a single, low-latency environment. A pause that is half a second too long can make it sound like the call dropped, while speaking too soon leads to talking over the other person.

Google Research notes that the AI must work just as well for the business answering the phone as it does for the user. This principle guided the design of the unified Gemini stack. The system must be able to navigate multi-constraint requests, such as coordinating car maintenance around specific calendar conflicts.

The current implementation uses the newest Gemini 4 models, including Gemini 4 Hydrogen and Gemini 4 Neon. These models help the system manage multimodal understanding and planning. They allow the agent to adapt dynamically when a receptionist suggests a different time slot or a different service.

Data and ecosystem impact

Duplex has already played a significant role in maintaining the accuracy of Google's local data. During the COVID-19 pandemic, the technology was used to automatically verify merchant information as business hours shifted rapidly. This ensured that listings remained accurate for users relying on Google Maps and Search.

The scale of this impact is significant, with business information sourced from Duplex calls appearing in Search and Maps over one trillion times. This data serves as a vital source for keeping the digital map of local services up to date. It bridges the gap between offline business operations and online discovery.

Moving forward, the integration into Gemini Tasks connects these capabilities to broader workflows. The system can now analyze utility bills and process transactions securely via the Agent Payments Protocol. It also coordinates with users over email and Rich Communication Services (RCS) to complete complex, multi-step requests.

The future of agentic calling

This launch marks a step toward closing the loop on AI assistance for real-world actions. By integrating Duplex into Gemini, Google is moving toward a model where AI can handle the 'messy' parts of life. This includes tasks that require human interaction but lack a standardized digital interface.

The move from bespoke neural pipelines to a unified Gemini architecture suggests a broader trend in AI development. Instead of building specific tools for every task, Google is building a general-purpose conversational infrastructure. This infrastructure can be prompted to perform a wide variety of roles depending on the user's needs.

As the technology matures, the focus will likely remain on reliability and respect for the person on the other end of the line. The goal is to ensure that the AI remains a helpful tool for coordinating chores without creating friction for local business owners. This balance is essential for the widespread adoption of autonomous calling agents.

How it works

The system functions by converting a user's text or voice request into a structured task. Gemini 4 models then interpret the requirements, such as specific dates, times, and preferences. It generates a dialogue plan based on these constraints to initiate the phone call.

During the live call, the AI listens for cues from the human receptionist, such as tone changes or specific availability. It uses low-latency speech synthesis to respond in real-time, mimicking human conversational flow. If the receptionist offers an alternative, the AI reasons through the new information to see if it fits the user's original constraints.

Once a confirmation is reached, the AI can update the user's calendar or provide a summary of the booking. This entire process happens autonomously, requiring the user to only provide the initial prompt and review the final result. It effectively acts as a digital proxy for the user's voice and intent.

Who is affected

The primary users are individuals who find manual phone calls for bookings to be a burden. This includes people managing busy schedules who need to coordinate appointments across multiple businesses. It also benefits users who prefer to stay on their devices rather than making active calls.

Local businesses are also affected by this technology. While they interact with an AI, the system is designed to be respectful and efficient. For many merchants, this means more accurate listings and a streamlined way to handle routine booking inquiries without increasing their own workload.

Finally, the broader ecosystem of local services is impacted. By automating the flow of information between businesses and consumers, the technology helps maintain a more accurate and accessible map of available services. It creates a more seamless connection between the digital world of AI and the physical world of local commerce.

What comes next

The next phase involves expanding the types of workflows Gemini can handle autonomously. With the inclusion of the Agent Payments Protocol, the AI may soon be able to handle secure transactions during these calls. This could mean paying deposits or settling small balances directly through the AI agent.

Google is also looking to deepen the integration across different communication channels. The system is already being prepared to coordinate with users over email and RCS. This means a single task could involve an AI making a call, sending a follow-up email, and updating a shared calendar.

As the Gemini 4 models continue to improve, the complexity of the tasks the AI can handle will likely increase. We can expect the system to manage even more nuanced requests, such as negotiating specific terms or handling multi-step logistics for complex events. The goal is to move toward a truly proactive personal assistant.

Summary of key changes

Google is moving from a specialized calling tool to a general-purpose AI agent. By integrating Duplex into Gemini Tasks, they are making it easier for users to delegate phone calls. This is powered by the Gemini 4 family of models, which handle the complex reasoning needed for live conversation.

The technology has evolved from simple, hard-coded scripts to a flexible, prompt-based system. This allows the AI to handle a massive variety of tasks, from booking a haircut to checking product availability. It is designed to bridge the gap where online booking forms are missing.

This shift represents a major milestone in agentic AI. It moves the technology out of the

Sources and links

Read the original at Google Research

First reported by Google Research. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.

Liked it? Say why

Add a line for other readers.

No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.

More AI news