Skip to content
AI.info

The Pulse

Google brings Duplex business calls to Gemini Tasks

Tasks in Gemini can now use Google Duplex to call businesses on users’ behalf, with advanced features rolling out starting in the U.S. Google says the calling system has evolved from task-specific models to a Gemini-based architecture.

Google brings Duplex business calls to Gemini Tasks

Tasks in Gemini can now call businesses on users’ behalf using Google Duplex, extending the phone technology beyond its earlier use in Search and Maps. Google Research described the integration in a November 20 post, saying the calling system has evolved from models trained for narrow tasks to an architecture built with Gemini.

The advanced features are rolling out starting in the U.S. The wider Tasks update is rolling out free of charge in the U.S., Japan, Korea, Canada, Australia and India, using the Gemini 4 Hydrogen model.

Making calls for users

Duplex can handle requests such as calling a business about an appointment. Google says its current calling stack combines understanding, planning and dialogue generation, allowing it to navigate requests with several constraints and respond when a business suggests a different time.

In a separate product post, Google Vice President Josh Woodward described asking Gemini to call the two nearest dealerships to arrange a service appointment for his minivan. He says Gemini found the nearest shops on Google Maps, called both at once and put an early Friday appointment on his calendar in less than five minutes.

Users can switch from Chat to Task in the Gemini app to use the feature. Google says it has tested tasks to minimize mistakes but has not eliminated them.

From fixed tasks to prompts

Google introduced Duplex in 2018. Its early version used custom recurrent neural network models trained on anonymized, task-specific phone conversations, with separate training for jobs such as restaurant bookings and checking holiday hours.

In 2022, Duplex used the LaMDA model in live calls. Google says early generative models were paired with state tracking and policy guardrails because they were less predictable and sometimes slower than current models.

The current calling stack uses Gemini to unify multimodal understanding, planning and dialogue generation. Rather than building a separate dialogue model for each category, the system can be prompted with a free-text task description, according to the research post.

Google says its experience with goal-directed calls helps the system interpret context before and during a call, including whether a task is likely to work and how to read a polite, indirect refusal. The company also says custom acoustic and conversation-cadence models help it handle background noise, pauses and when to stop speaking.

What businesses hear

Each call opens with a disclosure that an automated system from Google is calling. Merchants can opt out through their Google Business Profile settings or by asking during a call for the business not to be called again.

Businesses do not need to install software or maintain a booking system to receive the calls. Information gathered during calls can also appear on Google Maps, including details such as patio seating, pet-friendliness or parking information.

Duplex has previously been used to verify business information for Search and Maps. Google says information sourced from its calls had appeared in those products more than one trillion times, a figure the company shared in 2023.

Sources

  1. Original source · Blog post Google Research research.google

Other pages read

  1. Google blog.google

Companies in this story

In the tracker

Related stories

Explore

More articles