AI agents in Telegram bots: what they can actually do
The bots we build for taking orders and leads mostly run on a script: buttons, fixed dialogue branches, and pre-written texts. This isn't about being behind the times — it's a deliberate choice: a script is predictable, doesn't make mistakes with numbers, and doesn't “make up” answers. But it has an obvious ceiling — the bot only understands what's built into the buttons, and real customers don't type like that.
Over the past year, we've been getting this question more and more: “let's add AI here so the bot can answer any question.” Let's honestly break down what an AI agent layer actually changes on top of a bot like this, and what is deliberately kept on a strict script — and why.
Button-based bots: what they can already do and where they hit the ceiling
A scripted bot handles standard tasks well: show the catalog, collect an order via buttons, confirm the total, and pass the lead to a manager. The customer moves through pre-planned branches, and at every step, the bot knows exactly what to say because there is a finite number of options.
The problem starts when the customer doesn't play along with the buttons. They type “will it be cheaper if I order two,” “can you deliver tonight,” or just describe an issue in free text — without using a single menu command. In this situation, a classic bot either doesn't reply or asks them to “select a menu item,” which annoys the customer.
Open-ended questions: what the AI layer actually changes
An AI agent on top of the bot adds an intent recognition layer: it doesn't look for an exact button match, but understands what the customer is actually asking about. It either answers on its own — to general questions like business hours, delivery to a specific area, or return policies — or gently guides the customer to the right script branch.
For the business, this feels like fewer “misunderstood” messages and fewer customers dropping the chat because they couldn't find the right button. It's especially noticeable with non-standard phrasing — typos, casual language, or a mix of Russian and Uzbek in a single message.
But this recognition is probabilistic, not exact. The model might misinterpret sarcasm, confuse questions with similar meanings, or give an off-topic answer to a rare phrasing. That's why buttons and scripts aren't removed completely — they are kept as a safety net and the primary path for important actions like placing an order.
Chat summaries and response drafts for the manager
The second thing that really saves time isn't answering the customer directly, but helping the manager. The AI can read a long chat or ticket and provide a short summary: what the customer wanted, where they left off, and what remains unresolved. This saves minutes on every inquiry, especially if the manager joins the conversation midway.
Response drafts work in a similar way: for a complex or non-standard question, the AI suggests a phrasing, and the manager reads it, tweaks it, and sends it under their own name. It's faster than writing from scratch, but that's exactly why a human is still the one “sending” it — the draft remains a draft, not a ready-to-go reply to the customer.
The weak point of summaries is that they are an interpretation, not an exact copy of the chat. The model might miss a detail the customer felt was important, or smooth out the phrasing so much that a nuance is lost — for example, that the customer has already complained about the same thing once. For disputed conversations, the summary should be checked against the original rather than relied on as the sole source.
Why prices, stock levels, and order totals stay out of the AI's free text
A language model generates text that looks plausible — which is not the same as text that is guaranteed to be correct. It might quote a discount that never existed, say an item is “in stock” when it sold out an hour ago, or miscalculate the order total by a couple of digits. For a business, this isn't an abstract technical risk; it's real money and ruined customer relationships when the bot's promise later has to be explained and canceled.
That's why in bots with an AI layer, the price, stock level, order status, and final total are never pulled from the model's “memory” — only via a direct query to the database or CRM at the moment of the reply. The AI can craft a nice sentence around the number, but the number itself always comes from the accounting system, not made up on the fly.
A separate issue is where the chat data actually goes. If you just copy chats with prices and customer contacts into a public web chatbot for the “AI layer,” that's already a risk of leaking commercial information. The working approach is accessing the model via an API with limited permissions and no training on your data, rather than manually pasting business chats into the first service you find.
Frequently asked questions
Can the AI itself tell the customer the exact price or stock level?
It shouldn't. In a properly built setup, the AI doesn't make up these numbers or pull them from general knowledge — it queries your database or CRM and inserts the real value into the response. If this middleware is missing, the AI will eventually give the wrong number.
Is it safe to share customer chats with AI?
It depends on how it's set up. Using an API with restricted access and no data retention for training — an acceptable risk for most businesses. Manually copying chats with prices and customer phone numbers into a public web chatbot — a bad idea.
Is it worth adding an AI layer to an existing button-based bot?
It makes sense if a significant portion of customers ask questions outside the script, creating extra work for your managers. With a low volume of inquiries, a button-based bot often handles things just fine on its own, while an AI layer adds complexity that doesn't pay off.