Google’s Gemini agentic calling is built around disclosure and caller identity

Google has turned Gemini’s phone-calling experiment into a disclosure-and-identity product. In the company’s Agentic Calling terms, last modified Sept. 24, 2026, Google says users can direct Gemini Apps to place outgoing calls from their own phone number for personal tasks such as reservations, business-hours checks and other information requests. But the same document also requires a mandatory AI disclosure at the start of every call, says users must provide accurate identity information, and bars emergency calls, business use, telemarketing, political outreach and sensitive health conversations.

What Google is actually shipping

Launch coverage from TechCrunch and The Verge says the feature is arriving first as a limited experiment for Pixel 11 owners in the U.S. who pay for Gemini, with the Phone by Google beta app also required. Google’s model is not a general-purpose calling bot. It is a narrowly framed assistant that can place a call, introduce itself, navigate phone trees, wait on hold and, where the user allows it, share approved personal details during the conversation.

That matters because the product’s mechanics are more specific than the headline suggests. The call is made over the user’s own cellular network and personal number, while Gemini speaks in an AI-generated voice and can be monitored through a live transcript. Google says users can take over mid-call. In other words, the system is designed as a delegated action with a human override, not as an autonomous agent left alone to represent the user indefinitely.

Why the disclosure rules matter

The most important change is not just that Gemini can make a call. It is that Google has made identity and disclosure part of the workflow. Every recipient is supposed to hear that they are speaking with a Google AI assistant on a recorded line, and Google says users may not try to hide or bypass that disclosure. The company also requires callers to provide accurate identity and contact information so Gemini can pass it along correctly. That is an editorial inference from the terms and the launch reports: Google is building this feature as a disclosed proxy, not an anonymous voice layer.

For users, that design reduces some of the obvious deception risk that surrounds automated calling. A restaurant, store or hotel front desk is not supposed to mistake Gemini for a person. For businesses, it means AI callers may become more common, but they are not supposed to arrive as silent or disguised automation. The trade-off is that this is still a machine acting on your behalf. If the user approves the wrong detail, if the call transcript misreads the conversation, or if the recipient’s processes do not work well with an AI intermediary, the failure lands in a messy human workflow rather than in a simple app error.

The limits are part of the story

Google’s own terms acknowledge the edge cases. The company says Gemini may misinterpret responses or provide inaccurate information during a call, and it does not guarantee that every task will succeed. It also says the feature is not for emergencies and does not support emergency calling at all. Sensitive health information is off-limits, and standard carrier charges still apply because the call uses the user’s phone network and number. Those limits are not footnotes; they define the boundary between a useful convenience tool and a system that could become risky if it is asked to do too much.

That boundary is the practical takeaway for AI product teams and consumers alike. The use case that fits is narrow: low-stakes, personal, non-urgent calls where the user wants a machine to handle hold music, menus and routine script. The use case that does not fit is anything that depends on discretion, urgency, or legal and regulatory ambiguity. The limited early rollout gives Google a chance to test whether disclosure, identity handling and human override work across variable real-world conversations before it treats the feature as more than an experiment.

The live transcript is the safety net the user can actually control. Before relying on Gemini for a call, check whether its AI disclosure, caller-ID and personal-use limits fit the task, and be ready to take over if the transcript goes wrong.