On November 17, 2026, Google officially made Gemini Tasks free to users in six countries—the United States, Japan, South Korea, Canada, Australia, and India. The underlying engine is Google Duplex phone agent technology, which has been running for nearly a decade. Users only need to switch to Tasks mode on the Gemini app’s home screen, and the system can autonomously call local businesses to complete everyday matters such as restaurant reservations and hair appointments, and can complete payments within a user-preset spending limit through Google Pay’s Agent Payments Protocol (AP2). Throughout the entire call, it is mandatory to state, “I am an automated AI system from Google.” Merchants can verbally decline during the call or permanently disable such calls in the Google Business Profile backend.
The system runs on Google’s latest Gemini 4 Hydrogen model. In terms of technological lineage, when Duplex was released in 2018 it relied on custom recurrent neural networks (RNNs), and at that time could only handle extremely vertical, closed scenarios such as restaurant reservations and checking business hours; the system’s “understanding” depended on pre-written dialogue branch scripts. According to a technical retrospective published by Google Research on November 20, 2026, Duplex’s architecture went through three stages of iteration afterward: after introducing LaMDA in 2022, it adopted a transitional architecture combining generative models with deterministic state tracking; between 2025 and 2026, foundation models freed the system entirely from the constraints of fixed vertical domains, first enabling cross-industry invocation based on free-text prompts in Google Search and Ask Maps; and the release of Gemini Tasks marks the formal replacement of the previous domain-specific architecture by a fully unified Gemini technology stack. Engineers no longer need to train separate dialogue models for “restaurants” and “hair salons”; a single Gemini 4 foundation model can handle open-ended conversations with any merchant.
Making this system usable in real-time calls is far more engineering-intensive than text interaction. According to Google Research, real phone lines face acoustic interference such as kitchen noise and street background sounds; speakers change their minds mid-sentence; a half-second pause is taken by the other party as a dropped call; and interrupting means a failure of politeness. The system therefore specifically introduced acoustic models and conversational rhythm models to handle turn-taking timing and latency tolerance, and the overall inference pipeline was optimized for sub-second streaming responses over compressed phone lines. Before a call, the system also uses historical call data to make feasibility pre-assessments—for example, determining whether a restaurant accepts weekend group reservations—to reduce futile calls. Each task runs in a hardware-isolated virtual machine dedicated to that user account, layered with prompt-injection filters and Safe Browsing detection.
The payment process is another engineering-heavy part of this launch. Agent Payments Protocol currently supports Mastercard and Visa, and American Express is reportedly set to be added. The system requests user confirmation before triggering any purchase, and receipts are automatically stored in Google Wallet. For trusted routine operations, users can set an “Always allow” policy to reduce interruptions. This protocol effectively defines a standardized authorization interface between AI agents and payment infrastructure—the AI acts on the user’s behalf, but the authorization boundary for each transaction is set in advance by the user, and the agent may not exceed its authority.
For consumers, Tasks provides, beyond Chat, an interaction channel designed for long-running tasks and recurring matters, and it is completely free. More than 60 third-party apps are integrated in this launch, and developers can apply to join. But geographic limits remain significant: within the United States, due to differences in state-level privacy legislation, calling functionality is entirely blocked in five states—Indiana, Louisiana, Minnesota, Montana, and Nebraska. Germany, Austria, Switzerland, and the entire European Economic Area are likewise not in the first batch of covered markets—directly related to the complexity of EU compliance requirements for AI systems.
For small and medium-sized merchants lacking digital booking systems, this mechanism provides a zero-cost customer access channel: the AI screens appointment intent on their behalf and syncs information to their Google Maps business profile, without the merchant needing to deploy any dedicated software. According to data Google published in 2023, business information collected during Duplex calls over the years has been displayed more than 1 trillion times cumulatively in Search and Maps—a figure showing that merchant-side data collected through the phone channel is itself one of the important sources of Google Maps freshness, and the scaled rollout of Tasks will further strengthen this data flywheel.
For other players in the AI industry, this launch sets a compliance reference point. The mandatory AI identity disclosure mechanism, merchant-granular opt-out controls, hardware-isolated task sandboxes, and spending-limit authorization—the combination of these four designs effectively outlines the minimum technical foundation that an “accountable autonomous agent” should have. Doing these does not mean the system will not make mistakes (Google itself also acknowledges that errors cannot be fully eliminated), but failing any one of them will increase pressure at the regulatory and public-opinion levels.
To understand the historical position of this launch, it is necessary to return to the PR crisis of 2018. In that year’s Google I/O demo, Duplex played a recording of an AI call with a real restaurant host, with no AI identity disclosure throughout, and the system simulated highly humanlike pauses and filler words. The demo drew legislative attention in multiple U.S. states, and media and public criticism focused on “deceptively simulating humans.” Google later added AI disclosure in actual deployments, but the scale at that time was extremely limited, and compliance enforcement standards for AI transparency remained voluntary in the industry for a long time.
Eight years later, this launch writes mandatory AI disclosure into the system architecture as a non-bypassable technical constraint, rather than as an optional compliance add-on. The significance of this change lies not in a rise in Google’s moral standards, but in the fact that technological and market conditions now allow “transparency” to be embedded with zero additional friction—Gemini 4’s low-latency synthesis capabilities are sufficient to insert a clear AI disclosure at the beginning of a call without disrupting conversational rhythm, while the structured merchant opt-out mechanism also gives the system a sustainable business logic.
As for strategic direction, first, whether Agent Payments Protocol can become a cross-platform standard: if platforms such as Apple, Microsoft, or Samsung successively adopt AP2 as the unified interface for AI agent payments, this will evolve from a Google product feature into a contest for influence at the infrastructure layer. Second, the EU access threshold: Tasks currently bypasses the European Economic Area as a whole, and how the EU AI Act’s risk-tier requirements for “high-risk AI systems” apply to AI phone agents will directly determine the feasible path for such services in Europe; this discretionary outcome will also have demonstration significance for other regulatory jurisdictions worldwide. Third, actual merchant acceptance: Duplex call data was previously used mainly to check merchants’ business hours, whereas Tasks handles full commercial booking conversations—the merchant opt-out rate in actual use will be the core metric for whether this system can truly run the data flywheel of “phone agent → completed booking → map data update.”
© 2026 Winzheng.com 赢政天下 | 转载请注明来源并附原文链接