Skip to content
8 min read

AI Telegram bot setup for order intake from your site in Uzbekistan

AI Telegram bot setup is easy to start and easy to get wrong: the goal is a bot that does not just reply but carries a website order all the way into your CRM. In Uzbekistan customers are used to messaging rather than filling in forms and waiting for a call. Below is the full path from a site button to payment, plus costs, languages and the places where bots break in production.

Why take orders in Telegram instead of a web form

A name-phone-comment form leaves the person in the dark: it is unclear when anyone will call back, or whether they will. A bot replies at once, confirms the request was received and keeps the conversation context. For many shoppers in Uzbekistan Telegram is the main way to communicate, and it is already open on their phone.

The second argument is mobile connectivity. Outside Tashkent the signal is often weak: a heavy form page loads slowly, while short messenger messages go through even on a poor connection. That is not a reason to delete the form entirely, but the bot should be the primary path.

The link is simple: a site button with a deep link such as t.me/yourbot?start=product_123. The bot receives the start parameter and knows which product or service the customer cares about, so it composes the first message itself. The parameter is limited to 64 characters (Latin letters, digits, hyphen, underscore), so pass a short identifier, not a product name.

AI Telegram bot setup: a step-by-step plan

Step 1: create the bot with BotFather and keep the token in server environment variables, never in code. Step 2: use a webhook instead of long polling; it is easier to host and reacts faster. Step 3: write a Node.js or Python server function that receives updates. Step 4: connect an OpenAI or Claude model to parse text. Step 5: add data validation. Step 6: write the order into amoCRM or Bitrix24 through the REST API and notify a manager.

The model is there to extract structure, not to chat. The customer writes 'need 3 units, delivery to Chilanzar, tomorrow before noon' and the model returns JSON with quantity, district and delivery_window. This is a structured output task, and it is reliable when the response schema is strict.

Never let the model write into CRM directly. A validation layer belongs in between: the phone matches +998 followed by nine digits, the product exists in the catalogue, the price comes from your database rather than the model's answer, and the quantity is a positive integer. A model can make mistakes; a validator does not.

Keep the scenario short: name, product, quantity, address, time, payment method. Do not ask for what the deep link already told you. If the model is unsure, it should ask one clarifying question instead of guessing.

Russian, Uzbek and mixed input: keeping customers on their language

Customers write in Russian, Uzbek Latin, Uzbek Cyrillic and mixtures of them. Separate language detection is usually unnecessary: modern models handle mixed input if the system prompt plainly says to answer in the language of the last user message.

Buttons and templates are harder. They have to be maintained in two languages by hand, with an explicit language choice on first launch. Auto-detection from language_code is unreliable because many people have their phone interface set to English.

Transliteration is a separate problem. Districts and streets are spelled in many ways: Chilonzor, Chilanzar, Чиланзар, Чилонзор. Keep a synonym dictionary on your side and match addresses against it instead of relying on the model's guess.

Payments in the bot: Payme, Click and cash on delivery

There are two routes: send the customer a Payme or Click payment link, or use Telegram payments if the provider you need is available for your bot. Check this in BotFather and in the provider's terms before development starts, because the architecture depends on it.

The usual link-based pattern: the bot assembles the order, creates it on your backend, generates a link with the amount and order_id, and sends it as a button. The provider calls your webhook on success, and only then does the order become paid. A customer saying 'I paid' is not proof.

Keep cash on delivery available. Many customers, especially first-time buyers, choose it, and if online payment is the only option you will lose part of your orders.

What a bot costs and how long launch takes

The price is built from dialogue scenarios, integrations (CRM, payment, catalogue), the LLM layer with its checks, manager notifications, testing and hosting. The studio's core service, Telegram bots, starts from 999,000 UZS; the final figure depends on the number of scenarios and integrations.

For a market reference: a simple bot that captures leads and forwards them to a sales channel runs roughly 6,000,000 – 12,000,000 UZS, and a bot with free-text parsing, a catalogue and CRM roughly 18,000,000 – 40,000,000 UZS. This is a guide, not an offer.

Do not forget monthly running costs. At about a thousand conversations a month and a sensible context length, model API spend is typically around 300,000 – 900,000 UZS (a rough guide). Caching the system prompt and moving simple branches to plain code with no model call reduces it.

Timeline: roughly 2–3 weeks for a base version and another 3–4 weeks of tuning against real conversations. Do not skip the second phase: live customers do not write the way you imagined at design time.

What breaks in production and what to check before launch

Timeouts. Telegram expects a fast webhook response while the model may think for several seconds. Answer the webhook immediately, queue the job and send the reply as a separate sendMessage call, otherwise Telegram retries the request and the customer gets duplicates.

Lost leads. Write the order into your own database before pushing it to CRM, and run synchronisation as a background process with retries. A lead the bot accepted and then lost costs more than any refactor.

No human fallback. Keep a command that hands the chat to a person, plus a notification to a manager. That branch is what rescues complex orders where the model starts to get confused.

Pre-launch checklist: run ten real conversations in Russian and Uzbek, then test a wrong phone number, a non-existent product, a repeated submission, CRM downtime and a payment that arrives twice. Turn on conversation logging so you can improve scenarios from facts.

FAQ

Do I need AI if I have few products? Not always. If the scenario is linear, buttons with no model are cheaper and more predictable; AI pays off when customers write in free text.

Where do orders go? Into CRM (amoCRM, Bitrix24) and into a Telegram group for managers; using both channels is safer.

What if the model makes a mistake? The validator rejects bad data, the bot asks again, and on a repeated error it hands the conversation to an operator.

Can the bot accept payment? Yes, through a Payme or Click link or Telegram payments if the provider is connected; change the order status only on the provider's webhook.

Who can build such a bot? The 4 Digital studio in Tashkent builds Telegram bots from 999,000 UZS: describe your order process and we will estimate the scenarios and integrations.

Need a website or ads? Let’s discuss your project.

Telegram Bot Development · от 999000 сум

We reply within 15 minutes during business hours. No spam.