It looks ordinary enough: someone opens a chat with Alice and asks to book a dental appointment this week or a dog grooming session near home. Since November 12, 2026, "Act!" mode in Alice AI triggers an entire chain of actions behind the scenes right after that prompt. At launch, the feature opened to several thousand "Alice Plus" subscribers, with full rollout to all subscribers scheduled for December.
An isolated cloud Yandex Browser spins up on a remote Yandex server, where the agent opens tabs, reads pages, and fills out forms on its own. If a login is needed, control temporarily returns to the user, while partner tasks use plugins like Aviasales and external sources via MCP servers, as noted on the "Act!" mode page. When online booking isn't available, the agent picks up the phone and calls the receptionist directly.
On the clinic or salon's caller ID, an honest label appears: "Alice, user request." No more than ten calls per hour go out from a single number so business lines don't get flooded. Meanwhile, the user sees a live text transcript of the call in the chat and decides whether to confirm the booking or payment. The agent never closes a deal without human approval, and legal responsibility for the agreement stays with the client.
For local businesses, the real problem starts before the phone even rings. While receptionists are still learning not to hang up on the call as spam, the AI has already filtered out 99.9% of the market while building its call list.
The Moscow Experiment: 22 Companies Out of 19,592 Locations and Zero From Ads
On November 17, 2026, marketer Alexander Krasnobaev tested in Moscow who actually makes that shortlist and published his test breakdown on vc.ru. Because voice calling is still in closed beta for select subscribers, he tested the first and decisive step: Yandex AI's live answer to five client prompts like "find and book" or "find and compare." That answer is exactly where the agent pulls phone numbers to call.
To see the scale of the filter, the author compared the number of named businesses against total addresses in the same niche on 2GIS maps. The funnel was brutal: only 22 businesses made the shortlist out of 19,592 locations, roughly one company in nine hundred.
| Niche and Moscow Prompt | Named by Alice | Locations in 2GIS | What Made the Results |
|---|---|---|---|
| Dentistry "for treatment this week" | 4 | 5,363 | Plus two public clinics on a line below, not counted by the author |
| Auto repair | 3 | 7,541 | The answer advised calling a couple of places for an open slot |
| Dog grooming "this week" | 4 | 1,912 | All four salons pulled from their own sites with a booking button |
| Apartment renovation, estimate comparison | 5 | 2,501 | Contractors gathered from third-party ratings and reviews |
| Accounting for LLCs on USN, price comparison | 6 | 2,275 | Only companies with open price lists made the list |
These numbers have two fair caveats. First, the test measured Alice's search response step, so actual call completion and booking conversion weren't tested—though Yandex itself confirms the "call the ones I named" logic. Second, the denominator came from 2GIS, whereas Alice pulls local business profiles from Yandex Maps, and AI outputs vary from query to query. Even so, the scale of elimination doesn't get any softer: instead of thousands of pins on a map, the client and their agent look at three to six names.
The most painful finding hits paid traffic. Under each of the five answers sat one or two ad blocks marked "Promo"—nine advertisers in total. Not a single one of those nine made it into the AI's own list. While a human scrolls the screen, a "Promo" ad can still earn a click, but once an AI agent takes over the "find and arrange" task, the call goes to the organic shortlist above. That same week, OpenAI opened a ChatGPT dashboard for businesses showing mentions and referral traffic—both ecosystems turned AI answers into a standalone sales channel at once.
Four Paths to the Shortlist: What the Machine Actually Reads
When buyers used to open tab after tab, they wanted two things: a direct answer and outside proof that a company wouldn't let them down. Now an AI runs that exact check in fractions of a second, hunting for facts where businesses are used to writing vague estimates or hiding everything behind a callback form. Across the five Moscow answers, four distinct selection mechanisms stood out, depending on the service type and the cost of a mistake.
For simple local errands near home like dentistry and auto repair, Yandex Maps became the main gateway. Three clinics in the shortlist held a 5.0 rating backed by 3,200, 816, and 341 reviews. The AI even read the opening hours and told the user directly that the clinic "is open daily until 11:00 PM, so you can book even this evening." Evening and weekend hours turn into a direct argument for the algorithm.
In dog grooming, things worked differently: Alice ignored aggregators and pulled all four salons straight from their own websites. The answer almost word-for-word quoted the address, hours, and "How to book: online on the site or by phone" block, adding a "Book" button below and a note: "I checked: all of them have open slots right now." If your booking button and schedule are hidden inside an image or a pop-up widget that bots can't parse, you're left off this list.
In accounting for LLCs on the simplified tax system (USN), a plain-text price list made or broke everything. On the price-comparison prompt, the AI cut off the market in one line: "Below are specific companies with published prices." All six finalists kept their rates openly in page text. The common habit of writing "price on request" or "ask a manager" becomes a hard stop for agent traffic: the algorithm literally has nothing to put in a comparison table.
In apartment renovation with multi-million budgets, the high-ticket rule kicked in. The cost of a mistake is too high here, so the machine trusted neither map pins nor promises on contractor sites and went to third-party articles: "Who to compare: 5 companies from ratings." Sources included Dzen, Pikabu, Sovershenno Sekretno, and a banking portal. The pricier the service and the higher the client's risk, the weaker your own landing page becomes compared to independent reviews and comparisons on external platforms.
The Call Moment: What to Fix Before December So You Don't Lose Leads
Even a spotless map profile and an open price list can be wiped out in a second when the front-desk phone rings. An untrained receptionist sees an unfamiliar "Alice, user request" caller ID, hears a synthetic voice, and hangs up out of habit, assuming it's another spam robocall. At that exact moment, a ready-to-buy client who picked your business a minute ago and is waiting in chat moves on to the next competitor on the list.
Pay close attention to how your staff handles that call. Since the client reads a live text transcript of the dialogue, any hesitation, annoyance, or "call back later, we're changing shifts" shows up right on their smartphone screen. Your receptionist needs a short, clear script: greet calmly, state the nearest open slots, give a price range, and explain what's needed to lock in the booking.
To rebuild your funnel before the feature rolls out broadly in December, run through this short checklist.
Preparing a Local Business for Alice Calls Before December
- Test yourself in Alice with real customer prompts like 'Find a service in city where I can book this week' and 'compare prices' to see which platforms it cites.
- Put prices, real opening hours, and booking steps on your site as plain text, removing image-based price lists and 'ask a manager' placeholders.
- Upgrade your Yandex Maps and 2GIS profiles: verify your service list, highlight evening or 24/7 slots, and set up steady review collection.
- If your average ticket is high, build a presence in external comparison reviews, ratings, and articles on content platforms.
- Brief phone receptionists on the 'Alice, user request' caller ID and rehearse a quick reply covering open slots, prices, and booking terms.

