In Uzbekistan, and across much of Central Asia and the CIS, customers message businesses on Telegram first. So almost every automation conversation starts with "build us a bot". Then the confusion begins: one vendor offers a menu bot, another pitches an "AI agent", a third says they are the same thing. They are not the same, and they are not rivals either. A Telegram bot is the window inside the messenger through which a customer talks to you. An AI agent is a worker that can sit behind that window, understand free-form text and act inside your systems. This article walks through three practical setups, what each one demands from your integrations, a decision table by process and a checklist worth filling in before you write a brief.
The bot is the channel, the agent is the logic
Technically, a Telegram bot is an account controlled by a program through the Bot API. On its own it knows nothing: it receives a message, hands it to your code and sends back whatever the code returns. What it says depends entirely on what is connected behind it.
If a script sits behind the bot, you have a decision tree: a "Book" button, then service, then day, then phone number, then a lead in the CRM. Every step is drawn in advance and the bot never leaves the path.
If an AI agent sits behind the bot, the message goes to a language model that has instructions, a knowledge base and tools such as "find customer in CRM", "check stock" or "create request". The agent can parse something like "could I come Saturday afternoon, just not with the same technician as last time", decide which tool to call and write the reply itself.
So the real question is not "bot or agent" but what stands behind the Telegram window: a script, an agent, or both. The window itself is usually identical.
Three setups you will meet in practice
1. Scripted bot
Buttons, menus and forms. It works well when the customer journey is short and predictable: booking a fixed slot, sending a price list, collecting contact details, reminding subscribers. One example from our work is LUNA wellness: a visitor picks a class format on the website, the Telegram bot continues the automated funnel, collects the required details and sends the enquiry into ADDO CRM. No AI was needed, because the path was already clear.
2. A bot with an agent inside
The same bot, but free-text messages are routed to an AI agent. Buttons stay where they help, because tapping is faster than typing for frequent actions. The agent steps in when a customer writes in their own words, sends a voice note or asks something the menu does not cover. This is the most common setup for sales and support.
3. The agent answers from the company account
Plenty of customers do not write to a bot at all; they message a manager's or the company's own account. Telegram Business lets you connect a bot to such an account and choose which chats it may answer. Telegram has changed the eligibility rules for this feature over time, so confirm the current conditions in the app settings before you plan around it.
What each setup needs from your integrations
The difference in effort between setups rarely comes from Telegram. It comes from what the logic has to connect to.
| Component | Scripted bot | Bot + AI agent |
|---|---|---|
| Telegram channel | Bot token, webhook or polling | The same |
| Knowledge | Texts inside the script | A knowledge base of rules, prices and terms, with an owner who keeps it current |
| CRM | Writes a finished lead at the end | Finds the customer, creates and updates deals, adds a conversation summary |
| Accounting and stock | Usually not needed | Reads stock levels, order status, free slots |
| Handover | A "Talk to a manager" button | Rules for when the agent calls a person and what it passes on |
| Logging | Which buttons were tapped | Full transcript, tool calls and errors, for reviewing disputed cases |
If your CRM and accounting system cannot share data through an API, the agent can only repeat your policies, which makes it a more expensive version of a script. That is why data access is checked before the setup is chosen. For a concrete CRM example, see how an AI agent creates leads and moves stages in amoCRM.
Decision table by process
| Process | Best fit | What AI does | What stays with a person |
|---|---|---|---|
| Booking fixed slots | Script | Not needed if customers only tap | Exceptions and off-policy changes |
| Pre-purchase questions | Bot + agent | Understands the question in any of your languages, answers from the knowledge base | Special terms and discounts |
| Lead qualification | Bot + agent | Clarifies need, budget and timing, creates a lead with a summary | The call and the deal |
| Order status | Bot + agent with system access | Finds the order by phone number and reports the status | Complaints and compensation |
| Chats in the company account | Agent via Telegram Business | First reply and collecting details | Long-standing customer relationships |
A hybrid dialogue
An illustrative example: an electronics repair shop.
Customer (button): Check repair status
Bot: Please share the phone number the order is under, using the button below.
Agent: Found it: laptop, received on 21 October, status "waiting for a part". The system has no delivery date for the part yet.
Customer: can I at least get my files off it, I need them for work
Agent: That decision is up to the technician. I have passed your question on with the order number, and they will reply in this chat.
The button and contact request are the script; finding the order and understanding the follow-up are the agent; the non-standard request goes to a person. The agent does not invent a date the system does not have.
Telegram limitations to know upfront
- A bot cannot message first. The customer has to start the bot; reminders only reach people who already have.
- Voice notes arrive as audio and become text only after speech recognition. Test it on your real recordings, especially with mixed languages.
- Mixed scripts. Uzbek is written in both Latin and Cyrillic, often mixed with Russian. Keyword scripts break on this; agents cope better, but the knowledge base still needs every language you serve.
- Groups. By default a bot does not see every message in a group chat; this needs separate configuration.
- Personal data. Phone numbers and chat history are personal data, and local data protection rules apply to a bot and an agent alike.
Checklist before the brief
- Collect 20–30 real customer messages from the last month. If most fit three or four buttons, start with a script.
- Mark how many of them require data from your systems.
- Confirm that your CRM and accounting system expose that data via API.
- Decide where customers write today: the bot, a company account or a group.
- Name the person who will maintain the knowledge base and review difficult conversations.
- Write down the topics the agent must never handle alone.
If the checklist points to an agent, see how we build AI agents connected to CRM and accounting, or browse more articles in Integrations.
FAQ
Can an AI agent be added to a bot that already works?
Usually, yes. Scripts and buttons stay; messages that match no script start going to the agent. Customers keep using the same bot.
Which is cheaper to run?
A script mostly costs you in changes, since every new rule means editing the tree. An agent adds model usage per processed message and time to maintain the knowledge base. The total depends on message volume and which systems need connecting; a first conversation is how we work it out.
Could the agent promise a customer something it should not?
Only if it is allowed to. Agents get read-only tools where writing is not needed, topics such as discounts are off-limits, and every action is logged. Disputed cases go to a person.
Do I need a separate bot per language?
No. An agent replies in the customer's language within one bot. A script needs a language choice at the start and texts for each language.
Where should I start if I am unsure?
With the real-message review in the checklist above. If you would like to do it together, get in touch.