Newshot

Gemini makes calls for you, turning appointments from "speaking yourself" to "handing it to an agent"

AI summary

On November 20, 2026, Google integrated Duplex's calling capability into Tasks in Gemini: book a restaurant or a hair appointment with a single instruction, and it calls the business for you and negotiates the details. The shift isn't about whether AI can chat, but about handing over a real-world communication. It's more like a delegated assistant that can make calls than a chatbot: the former must catch real-time changes from the business, ask follow-up questions and confirm, while the latter mainly answers questions in a text interface.

Translation(model translation, not the source's own text)

Gemini Makes Calls for You, Turning Reservations from "Speaking Yourself" to "Handing It to an Agent"

Gemini is putting Duplex's phone-calling capability into Tasks, letting AI contact businesses on the user's behalf. It has evolved from an early narrow-scope restaurant-ordering model into a unified system combining Gemini's multimodal understanding, planning, and dialogue with specialized acoustic models; after a user submits a task, the system can question businesses, confirm conditions, and complete reservations.

Cover image from aiwrangler.co (news image)
Cover image from aiwrangler.co (news image)
  • Gemini can book restaurants and schedule haircuts on your behalf
  • Phone interaction depends on real-time understanding and generation
  • Duplex has evolved into a unified Gemini stack

This article is based on Google Research's official blog and the product capabilities it describes. All descriptions of features, architecture, and historical data come from vendor disclosures; no third-party hands-on testing was found.

On November 20, 2026, Google connected Duplex's calling capability into Tasks in Gemini: book a restaurant, schedule a haircut—give it a sentence, and it dials the business and negotiates the terms for you. The change isn't whether AI can chat, but that you've handed over a real-world conversation.

What calls can Gemini make for me?

Gemini can now do more than leave messages for restaurants—it can dial businesses on the user's behalf, handing over real-world chores like ordering food and scheduling haircuts.

Duplex isn't a new feature. It was originally the system inside Google for letting AI have natural conversations with local businesses, and it has since gradually entered multiple products; this time it's been built directly into Tasks in Gemini.

Tasks in Gemini is positioned as a personal AI assistant that coordinates real-world chores (an AI assistant that handles real-world matters for you). What distinguishes it from an ordinary voice assistant isn't whether it can dial, but what happens after the call connects: communicating with the business and following through on ordering and reservation requests.

The boundary of "making a phone call" has thus shifted: in the past, you asked the restaurant and Gemini relayed; now you hand over a reservation or haircut appointment to it, and it negotiates on your behalf to completion.

Whether it can clearly understand needs and handle changes between a human and an AI caller is the real test of this capability.

Why can a half-second pause ruin a phone call?

Reservation rules aren't necessarily complex; the hard part is that real people constantly change pace during a call. AI has to simultaneously understand, judge, and respond in time.

Real people rarely follow a script when calling. A business may change its mind mid-sentence, a requirement just stated may be replaced by a new idea; the other party may also interject, turning a simple confirmation into a string of improvised changes. AI can't just recognize fixed questions; it must understand from context what the other party is actually asking for now.

Even harder is judging "who should speak now." There are no buttons on a phone call to signal this; both sides often hand off turns through subtle shifts in tone. Pause for more than half a second, and the other party thinks the line has dropped; speak too early, and both talk at once, turning a polite conversation into mutual interruption.

So the system must control not just "what to say" but "when to say it." Having a text model generate a correct reply and then reading it aloud isn't enough.

High-quality speech understanding, situational reasoning, and speech generation need to work together in the same low-latency system (a system fast enough to respond and to handle sound cooperatively). Speech understanding handles hearing the content clearly, situational reasoning handles following changes of mind and interjections, and speech generation must pick the right moment and respond in a natural tone.

Google Research: Blog (webpage) official image 1
Google Research: Blog (webpage) official image 1

From being able to call, to handling things on the spot

Duplex's change is to let a phone system originally built for a single task gradually gain general capabilities for understanding conversations, responding on the spot, and accomplishing goals.

Early Duplex was like an operator who could only follow fixed procedures. It used task-specific recurrent neural networks (neural networks that process information in time order), trained on anonymized, task-specific phone conversations. The system could complete work within preset boundaries but couldn't flexibly handle all requests.

The turning point came in 2022. That year, Google Duplex was already using LaMDA (Google's large language model) in real-time calls.

LaMDA itself was a model launched in 2018; four years passed from its debut to being put into a real phone call. What changed over those four years wasn't just the model but what it had to face: changes on a call may come from the other party rather than preset paths in a test set, and the model no longer advances only along a specific task script but must handle unexpected turns and digressions.

Google's summary of the latest change is clear: Duplex has evolved from a custom neural pipeline (models and processing flows built around specific tasks) into a unified technology stack driven by Gemini (different tasks supported by the same underlying capabilities). Capabilities previously built separately per task are beginning to merge into a system that can support both conversation and task completion at once.

Google Research: Blog (webpage) official image 2
Google Research: Blog (webpage) official image 2

After Gemini takes over, call handling gains a division of labor

Gemini merges phone capabilities previously scattered across different models into a unified technology stack, then hands the error-prone line-level details to specialized models.

Gemini now handles the backbone of the entire call: understanding what the other party is saying, judging what the user wants to accomplish, planning how to communicate next, and generating appropriate responses. It places multimodal understanding (processing different forms of information such as speech simultaneously), planning, and dialogue capabilities in the same system, no longer relying on separately customized dialogue models for each type of reservation.

Specialized models handle the parts closer to the phone itself. Acoustic models identify speech amid background noise, line compression, and noisy calls; conversational rhythm models handle pauses, interruptions, and turn-taking, judging whether to keep listening, speak up, or leave the floor to the other party. Gemini decides "what to say and how to handle it," while the specialized models calibrate "when to listen and when to speak" to the rhythm of a real call.

Whether this division of labor can withstand real business calls depends on its performance under noise, interruptions, and on-the-spot changes of mind.

The same conversational capabilities have already crossed into Search, Maps, and devices

Duplex has entered devices, Search, and Maps, sharing the same conversational infrastructure. The agent doesn't just answer questions; it can advance real-world matters for users.

This infrastructure was first used for business information verification. During the COVID-19 pandemic, business hours and services changed frequently, and Duplex automatically verified this information and updated business profiles in Google Search and Maps. Information obtained from calls alone was displayed in Search and Maps more than 1 trillion times. Information generated from a single call thus entered the path users take to find businesses.

The call model then entered device scenarios, becoming the technical support for Call Screen and Hold for Me. Call Screen (a call identification and filtering feature) helps users handle unknown calls, while Hold for Me (a wait-on-hold feature) keeps the call connected on the user's behalf while waiting for a customer service agent. Both features rely on real-time processing of speech and conversation state—one guards the entrance to a call, the other stands watch in the middle of the wait.

In 2025 and 2026, foundation models further drove broader agentic calling capabilities in Google Search and Ask Maps. The same model capabilities can serve different entry points: managing calls on devices, initiating queries and tasks in Search, and handling place-related matters in Maps. The task entry points change, but the underlying capabilities remain reusable.

Our judgment: this capability has passed single-point validation. Whether it can scale next depends on users being willing to hand it the boundary of "speaking for me"; in scenarios involving price, time, and commitments, this premise does not automatically hold.

When AI negotiates terms with a business on your behalf, are you clear about how far it can decide for you?

Sources: Google Research, "How Google Duplex evolved to also call businesses for you in Gemini" (November 20, 2026), plus supplementary material from Gemini Drops, BPDATA, and TideReach; product capabilities, architecture evolution, and historical display counts are all Google official disclosures, and the related materials provide no third-party hands-on testing.

Glossary · Plain-language notes for this article

Full articles on deep dives, research解读, and tool tutorials, plus all podcasts, are in the member area.

The translation is machine-generated and individual terms may be off. Nothing is rewritten or cut; the original is authoritative.

Published by 听说了吗
This site only summarises and aggregates; copyright stays with the original author. Headlines and summaries follow the source — where they differ, the original wins.
Read the original ↗