Google is testing the feature initially for U.S. Pixel 11 owners who have a paid Gemini subscription and use the beta Google Phone app. Calls are placed from the user’s own personal phone number, and the user can watch a live transcript and take over at any time. Google also says the agent can share personal information only after the user approves it.
The mechanism is a supervised voice-agent loop: Gemini identifies itself, traverses IVR menus, waits on hold, then conducts the business conversation to pursue a bounded objective. That matters because the phone system remains an integration surface even where a business has no agent API. The user’s number, explicit data approval, transcript, and takeover control provide a practical accountability layer around delegation.
The experiment arrives as agent claims are increasingly judged by whether software can complete a real-world task, not merely recommend a next step. Google says it is starting at small scale because real-world conversations are nuanced, which sets an evidence boundary: the source describes intended capabilities and a limited test, not independently measured reliability, error rates, or broad availability.
For operators, the implication is that voice workflows may become a viable fallback channel for agentic automation where web and API access are absent. What to watch is whether Google expands eligibility and publishes evidence on successful completion, failure handling, disclosure, and abuse controls before a wider release.