Alice AI Starts Making Calls for Users: 10-Calls-per-Hour Limit per Number

Many people dislike making phone calls: waiting on hold, explaining the same thing over and over, booking a haircut by voice. Yandex has decided to hand this routine over to an assistant. "Alice AI" has started making outgoing calls on behalf of users in test mode. Product manager Sergey Mikhailov told the newspaper Vedomosti about it. Technically, this is one more step toward agents that act in the real world rather than merely answering in a chat. But the most interesting part of this story lies not in the model but in the rules around it.

How it works

The user sets a task, for example booking a table or scheduling a service. The assistant calls, reads out a pre-prepared list of questions, listens to the answer, and asks a follow-up if necessary. While the call is in progress, the recipient sees the caption "Alice, request from a user" on their screen. In other words, the person on the other end immediately understands that they are not talking to a live human. After the conversation, a transcript arrives in the chat, and the assistant asks the user to confirm important actions.

There is also protection against abuse. Alice can call a single number no more than ten times per hour. Yandex explains this by saying the tool must not turn into a means for cold calling and marketing campaigns. The company has not yet disclosed overall limits on calls and tasks per period. Pricing has not been revealed either; it is promised to be set later.

Test access will be given to several thousand active subscribers of the paid version. A mass launch is planned for the end of 2026.

Who is responsible for all this

This is the main problem. The terms of use classify calls in the new mode as the user's own actions. The user is obliged to supervise the assistant, check the results, and stop unwanted actions. Yandex describes itself as a provider of a technical tool, does not guarantee error-free operation, and refuses to compensate for losses. At the same time, the assistant is not supposed to perform legally significant actions fully autonomously, without human involvement.

Lawyers warn that an agreement the agent reaches during a conversation may be recognized as valid. In this logic, the assistant is simply a technical means through which a person expresses their will. One expert adds a nuance: if the agent agreed to terms that differ from those confirmed by the human, the error remains on the service's side. The transcript of the conversation would become key evidence in such a dispute. How this squares with the user agreement is not yet clear.

A separate question concerns telephony. For calls, Yandex uses a ready-made pool of numbers from a mobile operator that it does not name. This is not "Yandex SIM." If the company provided telecom services itself, it would be subject to the requirements of the Law "On Communications": SORM equipment and data storage in Russia under the Yarovaya law. For now, Yandex formally remains a user of someone else's network.

Competitors and trust

Telecom operators are heading in the same direction. MegaFon tested outgoing calls by its AI secretary "Eva" with free-form dialogue. T2 has a secretary called "Lina," but it only handles incoming calls. T2 calls proactive voice assistants the next stage of development. Beeline currently sees limited interest in delegating tasks to assistants and links it to distrust: people do not understand what data the AI will receive and what it will agree to on their behalf. MTS says that such tasks will be treated by its protection systems as ordinary calls from a private subscriber, and abuse will be blocked.

Takeaways

If you are building agents that act in the outside world, this case offers several practical solutions:

• Label the agent for the person on the other end. A caption on the screen removes some of the transparency concerns.
• Build limits in at the product level. Ten calls per hour per number is a simple way to cut off spam scenarios.
• Don't give the agent irreversible steps without human confirmation.
• Keep transcripts: they are both a service for the user and evidence in a dispute.

The limitations are also obvious. This is still a closed test with several thousand people, and nothing is known about accuracy or complex scenarios. The conflict between the user agreement and the lawyers' position is unresolved. For product managers the conclusion is simple: the calling technology already works, but the boundaries of responsibility and user trust will have to be built separately, and they will determine whether such a mode takes root.