MMOOZZGG.com

How to build an AI agent

AI customer support agent: how to build one and where its limits are

5 min read11 sources

How do you build an AI customer support agent that answers customers from your knowledge base and passes hard cases to a person in time? You can build a simple version with an answer base and a Telegram channel on your own, while a link to the CRM and tickets needs a developer. Gartner forecasts that by 2029 agents will close 80% of common requests with no people.

On this page · 5 sections

A developer is needed

What a support agent does

A support agent takes common questions: where is my order, how do I return an item, how do I pay, when is delivery. It finds the rules in the knowledge base and takes the data on a specific order from the CRM or the order system.

The second half of the job is the hand-off to a person. A complaint, a dispute or an unusual case goes to an operator together with the chat history, so the customer does not repeat the question.

Market expectations are high: according to a Gartner forecast that Call Centre Helper reported in March 2025, by 2029 agents will close 80% of common requests with no people, and support operating costs will fall by 30%. This is a forecast, not a measurement: your share will be your own, and it is counted in the pilot.

How we build such an agent turnkey is described on the page on AI for customer support. Here we look at how it is put together.

Build it yourself: base and channel

A simple version is built from two parts: a knowledge base with answers and a channel where customers write. The base is built with the RAG method: before answering, the agent finds fragments in your documents and answers from them. How to build it is covered on the page on the RAG agent.

A Telegram channel in n8n is connected by the documentation as follows:

1. Write to @BotFather in Telegram, send the /newbot command and set the bot’s name and a username that ends in bot. You cannot change the username later.

2. Copy the token that BotFather gives you. It is the bot’s password; never share it.

3. In n8n, create Telegram credentials and paste the token into the Access Token field.

4. Add a Telegram Trigger node with the Message event (a new message): it starts the workflow for each customer message.

5. Add an AI Agent node and connect a model and at least one tool to it, for example search in the knowledge base.

6. Send the reply with a Telegram node and the Send Message operation: the Chat ID field holds the customer’s chat, and the Text field holds the agent’s answer.

Such an agent answers from the base, but it does not see order status and cannot hand the conversation to an operator. That is the next level, and it usually needs a developer.

Integration with the CRM and tickets

A useful support agent sees the customer: it finds them in the CRM by phone or email, reads the order status, opens a ticket and passes it to an operator. A ticket, that is, a request card with a history and an owner, is needed so that no question gets lost.

A CRM opens such access to programs through an API, an interface for data exchange. In Bitrix24, for your own scenario you create an inbound webhook: the Applications section, Developer resources, the Ready-made scenarios tab, Other, Inbound webhook. The webhook works with the rights of the employee who created it, and only within the selected access rights.

Then the developer’s work begins: which fields to read, how not to create a duplicate customer, what to do if the CRM did not respond, how to pass the conversation to an operator in the same window. Setups for amoCRM and Bitrix24 are collected in the integrations section.

The agent sees customer data, so 152-FZ, Russia’s personal data law, applies: under Part 5 of Article 18, recording and storing the data of Russian citizens in databases outside the country is not allowed. So the base, the CRM and the model that reads the chats run on servers in Russia. More on the page on data security.

Metrics

The main metric of a support agent: the share of requests closed without an operator. This is what we write into the pilot contract.

Three more numbers are counted next to it: time to first reply, the share handed to a person and the share of wrong answers in a sample check. Without the last one, a high share of closed requests may mean the agent confidently answers wrong.

Measure these numbers before launch, or there will be nothing to compare with. How to check an agent before launch is covered on the page on agent testing.

Few companies can measure the effect yet. According to McKinsey (November 2025), 62% of organizations are experimenting with AI agents, 23% are scaling them, and 39% see an effect of AI on operating profit.

When to call us

It makes sense to build an agent on a knowledge base for one channel on your own and see how it answers. If there are few questions and no CRM is needed, this may be enough.

You need a team when the agent must see orders in the CRM and 1C, a Russian accounting system, hand conversations to an operator, work in several channels and store customer data under 152-FZ.

Our pilot: one agent on common requests, 14 days, from $1,499, with the metric in the contract. An AI department of 3–5 agents, where support works together with sales and booking, costs from $4,999 and is built in 4–6 weeks. Support from $480 a month.

A typical support agent is built from open blocks. If you are asked to pay for it as a unique development, ask what the price consists of, and pay for the result by the metric, not for hours.

Sources and fact-checking

Checked by the editors on 16 October 2026.

  1. arXiv 2005.11401, 22 May 2020 (arxiv.org)
  2. IT Brief, 11 November 2025 (itbrief.co.uk)
  3. Call Centre Helper, 10 March 2025 (callcentrehelper.com)
  4. Bitrix24, REST API documentation, 2026 (apidocs.bitrix24.ru)
  5. ConsultantPlus, 152-FZ On Personal Data, Article 18, 2026 (consultant.ru)
  6. Telegram documentation: Bot API: From BotFather to Hello World, checked 16 October 2026 (core.telegram.org)
  7. n8n documentation: Telegram credentials, checked 16 October 2026 (docs.n8n.io)
  8. n8n documentation: Telegram Trigger node, checked 16 October 2026 (docs.n8n.io)
  9. n8n documentation: AI Agent node, checked 16 October 2026 (docs.n8n.io)
  10. n8n documentation: Telegram node: message operations, checked 16 October 2026 (docs.n8n.io)
  11. AI for customer support: the pilot metric, MMOOZZGG.com

Found a mistake? Tell us

Publications

Contact us