An ordinary call to update a restaurant’s opening hours became an unusual AI moment in 2020: Google’s Duplex voice assistant reached a restaurant assistant made by PolyAI. The two systems spoke without either side being deliberately arranged for a demonstration.
The exchange was short, but it captured a change already underway. Voice assistants were no longer only tools used by people. They could begin encountering other automated systems in the everyday infrastructure of businesses.
Duplex Was Built to Sound Human
Google introduced Duplex as a system that could call a business for tasks such as making an appointment or reservation. Its natural pauses, pacing and acknowledgements made it sound strikingly similar to a human caller.
That realism impressed many people and unsettled others. A convincing voice could make a transaction smoother, but it could also create the impression that someone was speaking to a person when they were not.
The reaction helped establish an important norm: systems imitating human conversation should identify themselves clearly.
The Call Was About Routine Data
Duplex had been contacting businesses to confirm opening hours for Google Maps. That was a simple administrative task, but it required understanding a spoken answer and recording it accurately.
PolyAI’s restaurant assistant received the call. Its chief executive, Nikola Mrkšić, described it as the first naturally occurring conversation between voice assistants that he knew of.
The novelty came not from an advanced negotiation. It came from the fact that two deployed systems had met in the wild while trying to complete a real business task.
Automation Could Create New Kinds of Friction
When a person called a business, both parties could clarify an unusual request using shared experience. Two assistants had a narrower set of assumptions, different error-recovery patterns and no common understanding beyond the words exchanged.
That could produce loops, incomplete updates or escalating handoffs. A system might repeat a question that another system could not interpret, leaving neither able to decide when to stop.
Designers therefore needed to consider machine-to-machine interaction even when building products intended for people.
Human-Likeness Had Limits
Mrkšić acknowledged that Duplex sounded more human than PolyAI’s assistant, while also pointing to the “uncanny valley” effect. People often respond positively as a machine becomes more human-like—until it becomes close enough to feel deceptive or eerie.
For a transactional service, clarity could be more valuable than imitation. A short introduction such as “This is an automated assistant calling for…” gave the recipient context and a chance to choose how to proceed.
Natural speech was useful when it improved understanding, not when it concealed agency.
Businesses Needed an Opt-Out
Google had added disclosures at the start of Duplex calls and gave businesses a way to opt out. Those controls recognised that a shop owner might prefer to handle calls directly or use a compatible structured channel instead.
Consent mattered because voice systems could consume staff time, make mistakes or be difficult to distinguish from robocalls. A company should not have to adopt a new interaction model merely because a platform could generate calls at scale.
Opt-outs also created feedback that could reveal where automation was not delivering value.
Voice Agents Needed Shared Standards
As assistants became more common, they would need ways to disclose their identity, declare intent and exchange structured information. An opening-hours update could be better handled by a verified business feed than by a voice call.
Standards could reduce ambiguity. A caller might supply a verifiable organisation identity, specify that it was automated and offer a simple path to a human.
The goal was not to make every interaction robotic. It was to ensure that automation remained understandable when it crossed organisational boundaries.
Reliability Mattered More Than a Polished Demo
Voice systems had to cope with background noise, accents, interruptions and changing information. A glamorous demonstration did not prove that a product worked reliably across thousands of small businesses.
Evaluation needed to measure successful completion, correction rates, unwanted calls and the time required from staff. A system that sounded human but frequently created follow-up work was not efficient.
Human escalation was a useful safety valve. It could resolve edge cases and create training data without pretending that the agent was fully autonomous.
Privacy Applied to Spoken Data
Calls could contain phone numbers, reservation details and personal preferences. Providers needed to explain whether recordings were stored, how they were secured and whether they were used to train models.
Businesses also needed to know who was calling and why. Verification protected against fraudsters using a realistic synthetic voice to impersonate a trusted platform.
The more natural a system sounded, the more carefully identity and data handling had to be designed.
A Small Call Pointed to a Larger Future
The Duplex–PolyAI conversation was a curiosity, but it foreshadowed a world where software increasingly contacted other software on people’s behalf.
That future could reduce routine work, especially when agents handled simple, consented tasks through reliable interfaces. It could also create confusion if systems were optimised only to sound human.
The enduring lesson was that conversational AI needed etiquette as well as capability. Disclosure, consent, clear handoff and shared protocols were what made an automated voice useful in public—not how convincingly it could mimic a person.
