Google expanded its conversational Duplex assistant to the UK, Canada and Australia in April 2020, moving beyond its earlier deployments in the US and New Zealand.
Duplex could call businesses for a user to request services such as a restaurant table or salon appointment. Its human-sounding voice made the product memorable, but it also created a debate about disclosure, consent and whether automation was actually replacing people.
The Feature Solved a Narrow Problem
Booking a service by phone was an everyday friction point. A capable assistant could save a user time, particularly when they were unable or unwilling to call during business hours.
The task was deliberately constrained. It involved a known goal, a short conversation and an outcome that could be confirmed. Even so, businesses varied in procedure, language and availability.
The narrow scope was a sensible starting point for agentic AI: solve a real problem without pretending that the system understood every kind of conversation.
Human-Sounding Speech Created an Obligation
Duplex’s pauses and natural phrasing were designed to make calls feel less mechanical. The same feature raised concern that a staff member might believe they were talking to a customer.
Google responded by adding a disclosure at the beginning of calls and allowing businesses to opt out. Those controls gave recipients context and agency.
The lesson extended beyond phone calls. Any assistant acting externally should state its identity and purpose before requesting someone’s time or information.
Human Support Was Part of the System
Google said around a quarter of Duplex calls were started by human operators, while 15 percent began with the AI but later involved a person. Intervention could occur when a conversation became difficult or the recipient did not want to talk to automation.
Those figures showed that the product was a hybrid service, not a fully autonomous replacement for human work.
That was not a failure. Human fallback improved completion, reduced frustration and created a way to handle unusual cases responsibly.
Scaling Across Countries Was More Than Translation
The UK, Canada and Australia shared English with the US, but phone etiquette, accents, business practices and privacy expectations still differed. A system trained on one set of conversations could misunderstand another.
Local rollout needed testing with representative speakers and clear support for businesses. Availability also depended on local telephone infrastructure and consumer-protection rules.
International expansion was a product and governance exercise, not merely a switch turned on from a central office.
Businesses Needed Control
Small businesses often had limited staff to answer calls. An automated caller might save time by collecting details, or create work by repeating questions and producing an incorrect booking.
An opt-out was therefore essential, but it had to be easy to use and honoured consistently. Businesses should not be forced to navigate a complex platform account merely to avoid an unwanted AI interaction.
Structured booking integrations could offer a better long-term alternative to voice calls for many services.
Reliability Was the Core Product Metric
Public demonstrations focused on natural speech, but users cared about whether a reservation was correct. A wrong date, number of guests or location could be worse than no automation at all.
Evaluation should track completed bookings, corrections, cancellations and the share of calls handed to people. Systems also needed clear confirmation sent to the user and business.
Trust grew from predictable outcomes, not from a voice that happened to sound convincing.
Data Protection Was Not Optional
Calls could contain names, phone numbers, appointment details and health-related information. Google needed to explain recording, retention and training practices, while businesses needed to know who controlled the resulting data.
Consent rules varied by market, particularly around call recording. Design choices that were acceptable in one country might require a different approach elsewhere.
Privacy had to be considered alongside the conversation design, not after launch.
Automation Changed Work Rather Than Erasing It
Duplex illustrated a common pattern: AI handled repetitive portions of a task while people remained responsible for exceptions, service quality and accountability.
Workers needed training and clarity about when they could take control. Customers needed a route to a human when the automated path failed.
Measuring job impact solely by the number of calls initiated by a bot missed the new coordination and oversight work created behind the scenes.
The Expansion Set a Template for Agents
Google’s international rollout demonstrated useful restraint. The product had clear limits, disclosure, an opt-out and human backup.
Those design choices should be normal for agents that act on behalf of users. Autonomy was valuable only when the person receiving the action could understand it and decline it.
The durable lesson was that successful AI agents were not the ones that hid their automation most effectively. They were the ones that completed a bounded task reliably while keeping people in control at the moments that mattered.
