AI news ·
Google Duplex evolves to call businesses for users through Gemini tasks
Google moved its Duplex calling AI into Gemini Tasks, letting users delegate phone chores like booking appointments to a single assistant. The system runs on Gemini 4 models and can handle multi-constraint requests such as scheduling car maintenance around calendar conflicts.

Google has integrated its decade-old Duplex calling technology directly into Tasks in Gemini, the company's personal AI collaborator. The move lets working professionals delegate real-world phone chores - booking appointments, checking business hours, or verifying availability - to an AI agent that calls businesses on their behalf, a capability previously scattered across Search and Maps. The shift matters because it consolidates autonomous voice tasks into a single assistant, removing the need for a human to wait on hold or navigate fast-moving, unscripted phone conversations.
The system runs on the latest Gemini 4 models, including Gemini 4 Hydrogen and Gemini 4 Neon, and can handle multi-constraint requests like scheduling car maintenance around calendar conflicts. It also supports workflows such as analyzing utility bills and processing transactions through the Agent Payments Protocol.
How the architecture moved from narrow scripts to open-ended tasks
The Duplex stack evolved through three distinct generations. The earliest version, built before modern language models and even before the transformer architecture, used custom recurrent neural networks trained on task-specific phone conversations. Each task type - restaurant bookings versus holiday hours - required separate training data to navigate domain-specific logic.
A second generation adopted early foundation models like LaMDA in 2022. These hybrid stacks paired generative models for flexible phrasing with deterministic state tracking and strict policy guardrails, since early generative models were less predictable and sometimes slower than today's versions.
The current architecture removes fixed verticals entirely. The team shifted to prompting the system with free-text task descriptions, allowing Gemini to generalize across new business categories without building separate dialogue models from scratch. "Our calling stack uses Gemini to unify multimodal understanding, planning, and dialogue generation," the research team wrote.
Technical realities of live phone calls
Live telephony creates hurdles that standard web datasets do not solve. The system must handle compressed phone lines, background noise from espresso machines and kitchen clatter, and the conversational rhythm of busy merchants. Custom acoustic and conversation cadence models assist with sub-second latency so replies land at a natural pace, avoiding awkward pauses or interruptions.
Pragmatic grounding also matters before the call begins. The system matches a free-text task with a business based on historical context - knowing, for example, that a specific vendor does not take large-group reservations on weekends - to predict whether a task will succeed. During the call, this grounding helps the model interpret polite indirect refusals and keep interactions brief.
Designing for the business answering the phone
Every call opens with a disclosure that the caller is an automated system from Google. Businesses can opt out through their Google Business Profile settings or by saying "Please stop calling my business" during a call. The system works through the phone a business already answers, with no software to install and no staff training required.
Details gathered on one call can answer many later questions. Google Maps can display information Duplex collects - happy hour specials, patio seating, pet-friendliness, parking tips - so a business is not asked the same thing repeatedly.
Why this matters for operations and customer-facing teams
For professionals in customer support, hospitality, sales, and operations, the integration signals a shift in how routine business interactions get handled. If an AI agent can reliably call to verify hours, book appointments, or check inventory, frontline staff may spend less time on repetitive phone queries and more time on complex customer needs. For small business owners, the opt-out controls and zero-setup design mean the technology arrives without requiring changes to existing phone workflows.
Teams exploring internal automation may find the architectural progression instructive: the move from closed-domain models to a unified stack built on Gemini mirrors the broader industry shift toward foundation models that handle multiple task types through prompting rather than bespoke training. Professionals evaluating AI Agent Courses or AI Productivity Courses can see a real-world example of how agentic systems are moving from narrow use cases to open-ended task coordination.