Google says around 25% of calls placed using its Duplex AI service started with a human; calls to ~12 restaurants found people made 3 of 4 successful bookings
In a free service, bots call restaurants and make reservations. The technology is impressive, except for when the caller is actually a person.
Context & Ripple Effects
A year after the controversial I/O demo — where Google declined to name the hair salon and restaurant on stage or say whether the calls were edited [[a:929694]] — the company opened Duplex to trusted testers and opted-in businesses and pushed it to Pixel owners in select cities. The NYT's reporting now puts numbers on what those calls actually were: about a quarter started with a human, and across roughly 12 restaurants people completed three of every four successful bookings.
First-order effects
- Restaurants that opted into Dupex calls were frequently talking to human operators rather than the AI, meaning the service's early booking volume leaned on staffed callers under the Duplex brand.
- Google's disclosure that only ~75% of calls even began as AI attempts undercuts the fully-automated framing from its 2018 demo and invites scrutiny of how it reports success rates.
Second-order effects
- Businesses evaluating opt-in voice-agent services now have to ask vendors for human-vs-bot breakdowns, making disclosure of automation rates a competitive differentiator.
- Rivals building phone-based assistants face a raised bar: demonstrating bookings completed end-to-end by software, not by a human fallback wearing the product's name.
Third-order effects
- If the human-in-the-loop pattern holds, 'AI' calling services converge toward blended human-machine operations where the brand is automated but the margin is labor — until disclosure norms or regulation force the split to be labeled.
- Duplex's later claim of 1M+ completed bookings shows the volume case, but this reporting suggests the durable question is not capability but how much of each task machines genuinely close on their own.
The trend: Consumer voice agents are launching as AI-branded services that quietly blend human operators with automation, pushing automation-rate disclosure toward becoming a standard trust metric.