Google and Meta are putting AI agents on the phone. Both want people to hand over everyday errands, from checking stock to making reservations. My view is that Google has the stronger long-term position in this particular contest if it makes user control and disclosure reliable. Meta may win downloads and attention today. Yet an agent that speaks in your name needs a level of trust Meta has repeatedly made difficult to grant.

TheStreet’s September 27 comparison is the starting point for this opinion about Google’s limited phone beta and Meta’s Muse test.

Google offers controls at the moment of the call

Google’s new Gemini Call for Me trial is narrow. It is gradually reaching adults in the United States who have a Pixel 11, a Google AI subscription, and the public beta of the Phone app. The agent uses the user’s phone number and mobile network. Google says it identifies itself as an AI assistant on a recorded line, shows the task and information to be shared before dialing, permits live listening or transcript monitoring, and lets the user take over. It also rules out payments and sharing sensitive financial or health information.

These controls matter more than a polished voice demo. A person can see what an agent intends to say, who it is calling, and what happened afterward. A business can hear who is calling. This gives both sides a chance to correct an error before it becomes a mistaken reservation or an unwanted disclosure. TheStreet’s comparison frames the launch against Meta’s Muse. The actual test of the comparison will be whether these controls survive wider use and whether users find them easy enough to exercise.

Meta’s trust deficit is relevant to the product

Meta’s Muse also promises to act across a person’s services, and its launch materials describe a protected virtual machine, user controls, and an audit trail. Those features deserve evaluation. But earlier Newsroom reporting on Muse and Meta’s privacy record documented a problem close to this use case. Internal testers encountered calls described as AI work that were completed by human callers without the human handoff being clear to at least one tester. Meta said the feature was still being tested and that a public launch would include appropriate disclosure. The share of calls routed to people remains unknown.

The prior article also examined Messenger audio sent to contractors and Meta’s stated use of some AI conversations for advertising. Those histories do not prove that Muse’s current calling product mishandles every call. They do mean that a broad privacy promise carries less persuasive weight without visible, prior consent at each handoff. A call about an appointment can expose names, schedules, and other personal context. For a user, the difference between a model and a human operator is material.

Google has its own burden of proof

Google should not receive a free pass. Its agents also process personal details, and its calling service can retain a transcript and recording in the Phone app. Users need a clear account of retention, review, deletion, and any use of call data beyond completing the task. Google’s existing Search agentic calling already shows that the company has experience connecting calls to local commerce, but it does not prove this Pixel beta will outperform Muse. The current hardware and subscription limits sharply constrain reach.

Meta has a huge distribution advantage, a fast-growing Muse app, and time to repair its disclosures. Google has a mature phone surface and a design that presently makes the agent’s identity and user intervention explicit. My forecast is that the latter approach will win more sustained permission to make calls on people’s behalf. The prediction fails if Google obscures its own data use, if its agent repeatedly makes costly mistakes, or if Meta demonstrates a clearer and more reliable consent path. In the agentic AI race, trust becomes an observable product behavior when a user can stop, inspect, and correct the action.

Why the phone changes the trust calculation

A chat assistant can make a bad suggestion that a user reads and rejects. A calling agent acts in a live conversation with a business that may assume the caller has authority to commit to a time, disclose a detail, or speak for the customer. The user may see only a summary after the fact. That makes the approval screen, live transcript, and takeover control more than optional conveniences. They define where a person can interrupt the delegation.

Google’s published limits also draw a boundary around the kinds of tasks it will attempt. Barring emergency calls, payments, and sensitive information may frustrate some users, but it narrows the consequences of an early beta mistake. The challenge is whether the agent reliably honors those limits when a real conversation wanders into an unexpected question. If a restaurant asks for a card number to hold a booking, the software needs a clear way to stop or return control, not improvise around the rule.

Meta could build similarly strong safeguards. Its declared virtual-machine separation and audit trail are relevant designs. But they answer only part of the question if the user cannot tell who will actually place a call or see the request. The earlier Muse human-caller report makes that handoff a concrete test for Meta. Its fastest route to trust would be to make the handoff visible before any information is shared, give users a genuine choice, and show the resulting record. Those are product behaviors users can inspect rather than assurances they must accept.

What would change my forecast

I would look for four observations as these products mature. First, do calls identify the agent to businesses consistently? Second, can a user review the information and task before the call, then monitor and interrupt it? Third, are errors, cancellations, and human handoffs disclosed in a usable record? Fourth, does each company give a credible account of retention and secondary use of call data? Public evaluations and user experience over time matter more than initial app rankings.

Meta’s distribution may be decisive if it fixes the trust problem. Google may squander its advantage if it expands the feature too quickly, hides data practices, or makes the controls cumbersome. This is why the judgment is a forecast, not a claim of measured superiority. A trusted agent wins permission to act again tomorrow. For business calls, repeat permission is the advantage that can outlast a flashy launch.