AI Receptionist for Business in Uzbekistan: A Setup Guide

What an AI receptionist can and cannot do

It is good at repetitive work: saying when you are open, explaining how to find the office, finding a free slot, reminding someone about tomorrow's visit. It does not replace the person who negotiates a custom order, calms an unhappy client or sorts out a payment problem. The workable setup is simple: the assistant holds the first line, and your administrator takes everything difficult.

One caveat: a language model cannot see your schedule on its own. For booking to work, it has to be connected to a calendar or an online booking service. That is the technical part, handled by a developer or a ready-made platform. Your job is to define the rules.

• instant replies in Telegram, on the website and in Instagram Direct, including at night
• answers about services, addresses, opening hours and visit preparation
• collecting a name, phone number and the actual request into a clean lead
• booking, rescheduling and cancelling against a real calendar
• reminders the day before and on the day of the visit
• passing a hard case to a staff member with a short summary

Start with a map of your inbound requests

Export a week or two of conversations from Telegram, Instagram and your website chat, and add the calls your administrator made notes on. Sort them into groups: general questions, booking and rescheduling, payment, complaints, everything else. For each group, write down who answers now, what information a good answer needs, and when a human must step in.

That map becomes your brief. Usually a handful of repeated questions account for most of the volume. Start with those and leave rare, unusual cases to your staff.

A knowledge base and hard limits

Language models like to answer even when they do not know. So prepare a short document of plain facts: services, branches, working hours, which specialist does what, and what can only be settled in person. Write it as simple statements, not marketing copy. In the instructions, say the assistant must answer only from this document, and if the answer is missing, say it will check with a manager and take the client's contact details.

List the topics it must never handle: pricing of custom work, medical or legal advice, promises about deadlines, refunds. If you take prepayment through Click or Payme, the assistant can send the payment link, but any dispute about a payment goes to a person.

Booking and reminders without extra questions

A good flow does not make the client repeat themselves. If someone writes, "I want a cleaning tomorrow evening," the assistant pulls out the service and the day, checks the relevant specialist's free slots and offers two or three options rather than the whole list. After the client agrees, it creates the booking and sends the date, time, address and what to bring. Define the length of each service and the buffer between clients up front, otherwise a long consultation ends up in a short gap.

Keep reminders short: date, time, address and an easy way to reply "yes" or "reschedule." If the client stays silent, the task goes to the administrator for a phone call. A freed slot can then be offered to someone else. Do not overdo the number of messages, since an extra reminder starts to feel like spam.

Languages, tone and handing over to a human

A customer in Tashkent may start in Russian and finish in Uzbek, and Uzbek itself arrives in both Latin and Cyrillic script. Write a rule: reply in whichever language dominates the message. Test the Uzbek replies separately, because models differ in how well they handle it. Keep the tone polite, free of bureaucratic phrasing and emoji clutter, and say in the very first message that this is a virtual assistant, so nobody expects more than it can do.

Spell out the handover triggers explicitly:

• the client is unhappy or complaining
• the question is not in the knowledge base
• the client asks for a real person
• the topic is money, contracts or refunds
• the assistant failed to understand the request twice

How to tell whether it is actually working

A feeling that things got calmer is not a metric. You need numbers from your own systems: how many inquiries arrived, how many reached a booking, how many booked clients actually showed up, how many conversations went to a manager. It also matters where the client came from: a Meta ad, Instagram, the website or an OLX listing. That means tagged links and a source field saved in every lead.

When ads, conversations and visits sit in one report, you can see where people drop off: before the first reply, after the offered time, or before the visit. markdev builds this link between advertising, dialogue and booking as part of end-to-end analytics, so budget decisions rest on facts.

Where to begin

Pick one channel, most often Telegram, and one task, for example booking a consultation. Write the rules, connect the calendar, then test it by asking colleagues to play difficult and confusing clients. Only after that open it to real customers.

From there, read a sample of conversations every week and update the knowledge base the same day a schedule changes or a branch opens. If you want to work out which requests in your company to hand to an assistant first, and how to see the results in your reports, write to markdev and we will go through it using examples from your own conversations.