Google announced on November 20, 2026, that Duplex, the AI phone-calling system first demoed in 2018, is now integrated directly into Tasks within the Gemini app. This move allows users to describe real-world chores in plain text, prompting Gemini to place live phone calls to businesses that lack digital booking forms or APIs. The integration marks a shift from Duplex as a standalone novelty to a critical infrastructure component for agentic workflows, specifically targeting the 'last mile' of digital interaction where structured data is absent.

The Phone Call as the Fallback API

The core architectural insight here is treating live telephony as a fallback modality rather than a party trick. Tasks in Gemini now coordinates utility bill analysis, Agent Payments Protocol transactions, and email/RCS communication alongside voice calls. This design acknowledges that most small businesses will never expose an API, making the phone the only viable interface for automation. For developers building agents, the lesson is clear: an agent's utility is defined by its ability to handle unstructured, legacy channels, not just clean REST endpoints. If your agent fails when an API is missing, it is not truly agentic.

Technical Evolution and Stack Details

The current implementation relies on Gemini 4 Hydrogen and Gemini 4 Neon, which unify speech understanding, planning, and dialogue generation into a single stack. This contrasts sharply with the original 2018 Duplex, which used constrained recurrent neural networks trained on narrow, task-specific datasets. The modern approach eliminates fixed verticals, allowing the system to generalize to new business categories without custom model training. Google highlights technical hurdles such as pragmatic grounding for interpreting indirect refusals and sub-second latency on noisy, compressed phone lines, leveraging nearly a decade of goal-directed call data.

Privacy and Data Extraction Concerns

While Google emphasizes clear disclosure at the start of each call and opt-out mechanisms for merchants, a significant data extraction pipeline is now active. The 'one conversation, many insights' feature repurposes details gathered during callsβ€”such as patio seating or pet policiesβ€”for display in Google Maps and Search. Businesses may not explicitly agree to publish this information, raising questions about consent when offhand answers become public data points. This efficiency is a double-edged sword, turning every automated inquiry into a potential crowdsourced data update for Google's knowledge graph.

What Google Did Not Publish

Despite the fanfare, Google has not released a public Tasks telephony SDK or API keys for third-party developers, meaning this remains a closed product feature rather than an open platform. There is no disclosed per-minute pricing or usage limits for the calling functionality, and the specific country rollout list for day one remains ambiguous. Developers should not plan roadmaps around building on this stack, as it is currently exclusive to the Gemini experience. The announcement follows the expansion of Tasks to free users on Gemini 4 Hydrogen on November 17, 2026, indicating a strategy to push Gemini from a conversational assistant to an active agent.

Key Takeaways

  • Duplex is now a built-in fallback for Gemini Tasks when APIs or web forms are unavailable.
  • The technical stack uses Gemini 4 Hydrogen and Neon for unified speech, planning, and dialogue.
  • Google has not released a public SDK or pricing details, keeping the feature closed for developers.
  • Automated calls now feed data into Google Maps, raising questions about implicit merchant consent.

The Bottom Line

Google’s integration of Duplex into Gemini Tasks is a pragmatic admission that the digital world is not fully digital. By treating the phone call as a fallback API, Google is solving the last-mile problem for agents, but developers should note the closed nature of this stack and the potential privacy friction with merchants.