The development
What happened
Google’s 20 November technical account describes how Duplex evolved from narrow phone-call models into a Gemini-based stack for completing tasks with businesses. Gemini combines multimodal understanding, planning and dialogue, while specialized acoustic and conversation-cadence models help handle noisy phone lines and turn-taking. This is an architectural explanation following the week’s Tasks in Gemini launch, not Duplex’s first release in 2018 or a new worldwide availability announcement. Google says calls disclose the automated AI agent and businesses can opt out.
Key facts
Duplex began with task-specific recurrent models trained on anonymized phone conversations. Google says it used LaMDA in live calls in 2022, pairing early generative models with deterministic state tracking and policy guardrails.
The current stack uses Gemini for understanding, planning and dialogue. Free-text task descriptions replace separate fixed-vertical dialogue models, allowing constraints such as calendar conflicts or a receptionist’s alternative appointment.
Google identifies pragmatic business context and low-latency speech on compressed, noisy phone lines as separate challenges. Custom acoustic and cadence models assist with background noise, pauses and when to stop speaking.
Google says every call begins by identifying an automated AI system calling from Google. Merchants can opt out in Google Business Profile settings or ask during a call for the business not to be called again.
Businesses can use their existing phone without installing software or maintaining a booking system. Google says information from one call can serve later Search and Maps questions, rather than repeating the same inquiry.
The post names Gemini 4 Hydrogen and Neon in Tasks workflows. It does not provide a controlled calling-success benchmark or enumerate a complete country, account and plan eligibility matrix.
The context
Why it matters
TideReach assessment: the new detail is how an AI assistant reaches services that lack a usable booking API, not a claim that conversational fluency solves authorization or reliability. Business-side disclosure, opt-out and information reuse matter alongside the user’s task request. The engineering account explains design choices, but does not independently verify that every call succeeds, that all described controls work universally, or that access has expanded to every region. No calls or product tests were performed by TideReach.