Agent

AI manager for holiday cabins

It does not consult, it sells: it runs the conversation by their manager's playbook, handles objections, takes the deposit and comes back to those still thinking.

Live right nowSend a DM

@a_tmosfer_a

Their manager's playbookNine working situations were walked through with their best manager and turned into rules: how to open, what to ask and in what order, what to answer to «too expensive», what to offer when everything is booked. The agent sells in that voice, not in polite generalities.

22 knowledge blocksBuilt from the managers' playbooks and the good conversations of the past year, picked by the client. For a month the agent answered into a drawer, and its answers were compared with the way a manager answers.

16.7 secondsMedian time to reply, measured across 287 conversations in one week. Seven of those seconds are a deliberate pause: the agent waits to see whether the person is still typing.

91% of messagesClosed without a human. 27 handovers to a manager across 287 turns in that same week.

What they came with

A manager ran the DMs by hand: 8,447 conversations, and 6,805 of them asked about dates or price. In this niche a guest writes to several places at once and books with whoever answers first that the dates are free. In the evening, at night and on weekends there was no answer for hours, and that is exactly when people choose where to spend the weekend. Payments were matched by hand: the guest sent a screenshot of the transfer, the booking was confirmed manually, 537 screenshots over the same period.

What was built

The guest writes in the DMs as usual. The agent answers from a knowledge base built from the managers' playbooks and the conversations of the past year that the client picked as the good ones, goes into the CRM calendar itself for free dates and the price on those exact days, collects name, phone, dates and party size, shows a summary for confirmation, creates the booking and sends a prepayment link. After payment it checks the CRM and sends the arrival details in her own wording. A guest who goes quiet gets three reminders: half an hour before the invoice expires, at the moment it does, and a day later. The owner gets a card in Telegram for every new booking and for every message the agent did not take on, and runs the whole thing from there.

Telegram consoleEight commands and buttons under the cards: status, summary, list of pauses, a master switch. And the fork that matters more than the buttons: if a person should speak next, the agent goes quiet; if it is only missing one fact, it promises to check and keeps the booking going.

Database on the client's serverConversations, dialogue memory, created bookings and settings live in one database. The server is registered to the client.

IntegrationsInstagram Direct and post comments through ManyChat, TikTok direct on the same stream, BookMeNow for dates, bookings and prices, Monobank for prepayment links.

Two pass generationThe model returns a decision, ordinary code performs the action, and on the second pass the model puts the answer into words based on what actually happened. Dates, prices and bookings are never invented.

Five processesIntake, the brain, the console, daily reminders and a watchdog for invoice expiry. Each has one job.

Data passportEvery price and every rule is numbered on one page, and the owner either confirms it or corrects it. An error in the base would become an error in every answer.

Stack

n8nDockerPostgreSQLNode.jsPythonManyChatBookMeNowMonobank

What turned out to be hard

Selling in a borrowed voice

Polite generic phrasing sells worse than a real person who knows the niche and the guest. So the playbook was taken by interview: nine situations, the manager's answers, a pass over the contradictions between them and a run on a test stand against the same scenarios before release.

Pick the right cabin, not list them all

A couple, a group of eight, a birthday, weekdays or a weekend: every case has its own cabin, its own sauna and its own order of offers. The logic is written down in advance, so the agent offers one fitting option and an alternative to it, not a list of everything that is free.

Remind without waking anyone up

The follow-up is built not to look like a mailing: a soft reminder, an offer instead of a repeat, and daytime hours only, from ten in the morning to nine in the evening Kyiv time. At night nothing goes out, even if the invoice expired exactly then.

Someone else's booking system and their exceptions

Dates, prices and bookings live in BookMeNow, and the agent goes there for every answer. Their own cases are worked out separately: a single weekend night, cabins that are not given to a couple, holiday prices and services beyond the stay. The agent does not guess, it asks the system and answers by their rules.