Three kinds of bot, not one
"A chatbot for business" covers three different things.
A button bot. The customer picks from a list: "Delivery", "Payment", "Opening hours". The bot answers with a prepared text. It is built in a no-code builder and never says anything you did not put in it. The limit: it does not understand questions that are not on the menu.
An AI assistant with a knowledge base. The customer writes in their own words, and an AI model composes the answer from your material: delivery terms, prices, return rules. It understands free-form questions but can get things wrong or invent what is not in the base. So it needs restrictions and handover to a person.
A bot with integrations. It pulls data from your systems: order status from the CRM, the waybill number, stock levels. That is development — by a programmer, or in a builder connected through an API.
A word on the channel. Telegram does not charge for a bot's messages to users; the limits concern only mass broadcasts — above 30 messages per second they are paid. What you pay for is the builder, the AI model and the people who set the bot up.
A decision path by type of enquiry
Take the last 100 customer enquiries across all channels and sort them with three questions.
- Are most enquiries the same few questions?Delivery, payment, hours, address — the wording does not matter because the answer is the same. Yes → a button bot
- Are the questions varied, but each answer is in your rules?"Can I return it if I opened the packaging", "do you deliver to villages" — people ask in different ways, but the rule is one. Yes → an AI assistant with a knowledge base
- Does the answer need data on a specific order?"Where is my parcel", "did my payment go through". The rules cannot tell you that. Yes → an integration with the CRM or the courier
If after this sorting you have 20 enquiries a day and all are different and complex, you do not need a bot. You need another manager or a better set of reply templates.
What AI answers cost
The model itself is the cheapest part. An illustrative calculation at OpenAI's API prices as of 7 October 2026: one assistant answer is about 2,000 input tokens (the question, the conversation history and knowledge-base excerpts) and 300 output tokens.
Prices per 1M tokens, standard mode, short context: Luna $0.10 input / $0.50 output, 6.1 Sol $2 / $10, Astra $10 / $50. One answer: 2,000 input + 300 output tokens. Caching the repeated parts of a request lowers the total.
The cheapest and the most expensive model differ a hundredfold, yet even the cheapest is usually enough for answering from a knowledge base. And the total is not the point. The point is what one wrong answer costs: a delivery date that will not happen, a discount that does not exist, a wrong return condition.
Testing on 50 real questions before launch
Here is a method that takes a day and catches most problems.
- Collect the last 50 real customer questions — verbatim, with typos and abbreviations. Do not invent convenient ones.
- Run them through the bot and sort the answers into four groups: correct; correct but the wording needs fixing; wrong; handed over to a person.
- Look at the wrong ones. Every wrong answer about money, dates or terms is either a gap in the knowledge base or a question the bot should not settle on its own.
- Fix the knowledge base and the handover rules, then run the same 50 questions again.
Count the share handed over to a person separately. If the bot hands over half the enquiries, it does not save time — it adds a step. If it hands over none, it is most likely answering where it should have stopped.
What a bot must not do
- Promise discounts, compensation or dates that are not in the rules.
- Settle disputes: complaints, returns with damage, conflicts — straight to a person.
- Ask for and keep unnecessary data: passports, full card numbers. Which data can be passed to AI services at all is covered in the article on data for an AI assistant.
- Answer on topics outside the business. Restrictions in the bot's instructions are part of the setup, not an option.
More broadly, which processes should not be handed to AI is covered in the article on what not to automate.
Our position
Start with a button bot, even if you plan to use AI. It shows which questions people really ask and gives you the knowledge base for a future assistant. Launch an AI assistant only after testing it on real questions, and with handover to a person for anything involving money or disputes.
The position does not hold if you get hundreds of enquiries a day, most of them free-form, and your knowledge base is already written down. Then a button bot is an intermediate step with no benefit, and it is worth building the assistant straight away, with testing.
API prices checked on 7 October 2026 on OpenAI's pricing page, Telegram's terms in its FAQ for bot developers. The cost calculation is illustrative: the number of tokens depends on the length of your knowledge base and conversations.