From dedicated models to free-text requests

Google describes three generations. Early Duplex used custom recurrent neural networks trained for narrow domains, such as restaurant reservations and business hours. The company says LaMDA entered production calls in 2022, followed by hybrid systems combining generative language with deterministic state tracking and policy rules.

The current approach accepts free-text task descriptions instead of requiring a separate dialogue model for each business category. Google says Gemini brings together understanding, planning and dialogue generation, allowing a caller to handle constraints such as preferred appointment times and alternative slots suggested by a receptionist.

Telephone conversations need more than a capable language model

The report highlights compressed audio, background noise and fast changes of speaker. Google describes streaming inference and custom acoustic and conversational-cadence models that help manage pauses, interruptions and turn-taking. Its sub-second latency discussion is an engineering target, not a published end-to-end success benchmark.

Context also begins before the call: previous interactions may indicate that a particular business cannot handle a proposed request. During a conversation, indirect refusals and corrections need to change the plan. These are task-specific interaction problems; adding a long system prompt alone does not demonstrate that they have been solved.

Disclosure and control for the business

Google says every call begins by identifying itself as an automated system calling from Google. A merchant can disable automated calls through Google Business Profile or ask during the call to stop receiving them. The company says businesses need no new software or booking-system integration.

The November 17 product announcement describes an initial free rollout across the United States, Japan, Korea, Canada, Australia and India, with advanced features starting in the US. Those general rollout countries do not establish calling availability everywhere: its footnote excludes calling in Indiana, Louisiana, Minnesota, Montana and Nebraska. Neither source confirms this calling feature for Ukraine.

What the report establishes, and what remains unmeasured

The sources explain Google’s architecture and stated controls, but provide no comparable completion-rate dataset, independent accuracy audit or measured error rate for bookings. We have not tested the service. A successful spoken exchange should be distinguished from a correct appointment entered at the right business, date and time.

For a practical evaluation, our proposed checks are task completion, confirmation accuracy, unwanted disclosures, interruptions and total time including corrections. The research report does not release downloadable model weights, a LoRA recipe or local hardware requirements; it describes a hosted product capability rather than an offline phone agent.