MMOOZZGG.com

How to build an AI agent

How to build an AI sales agent: lead, questions and deal in the CRM

6 min read6 sources

How do you build an AI sales agent that answers a lead in a minute, asks qualifying questions and creates a deal in your CRM? The simple part, replies in a messenger from your price list, you can build yourself, while the CRM link needs a developer. Our pilot of such an agent takes 14 days, with the metric written in the contract.

On this page · 5 sections

A developer is needed

The sales agent’s chain

A sales agent takes five steps. It receives a lead from the website, a messenger or email. It replies to the client within a minute, at any hour. It asks qualifying questions: what is needed, by when and at what budget. It creates a contact and a deal in the CRM. It hands ready-to-buy clients and unusual requests to a manager, together with the chat history.

Qualification is a filter: the agent finds out whether the client is ready to buy and is a good fit for you. The head of sales writes the questions and the filter rules, not the developer.

Each step on its own is simple. The difficulty is in the joints: the agent must understand free text, must not lose a lead when the CRM fails, and must not promise what is not in the price list.

There are usually several channels: a website form, Telegram, MAX, email and calls. The agent covers text channels with one logic, while calls need speech recognition and telephony. That is a separate build, described on the page about a voice agent.

If you need such an agent end to end, see the page AI for sales teams. Here we look at how it works and what you can build from it without a team.

What to build yourself

On your own you can realistically build the first half of the chain: an agent that answers in Telegram from your price list and terms and asks qualifying questions. It is an agent with instructions and examples, without integrations. How to connect the channel is covered on the page about an AI agent in Telegram.

The instructions hold the price list, payment and delivery terms, the qualifying questions, samples of good answers and a list of bans: no discounts, no delivery promises without a manager, no talk about competitors.

The agent can send the result of the chat to a manager as a message: name, contact, answers to the questions and a rating of readiness to buy. Even without a CRM, this saves the manager the first call.

Test such an agent on old leads: take a month of chats, ask the agent the same questions and compare its answers with the managers’ answers. This shows where it mixes up the terms and where it asks needless questions.

Most companies stop at this stage. According to McKinsey (November 2025), 62% of organizations experiment with AI agents, and 23% scale them. Moving from a trial to a working sales agent starts with the CRM.

CRM integration

Without a CRM, a sales agent stays a chat. In the CRM it creates a contact and a deal, sets the pipeline stage, writes the answers to the questions into the card fields and assigns a responsible manager.

For the agent, a deal in the CRM is a tool: an action it calls by itself when the client has answered the questions. How agent tools work is covered separately.

Both popular Russian CRM systems open this to programs. In amoCRM, the method that adds deals together with a contact and a company accepts up to 50 deals per request and checks for duplicates only against data already saved. In Bitrix24, you create an incoming webhook for your scenario, and it works with the rights of the employee who created it.

Hence the developer’s work: finding a client duplicate by phone, mapping fields, retrying after a failure and logging the agent’s actions. Schemes for amoCRM and Bitrix24 are in the integrations section.

Client contacts are personal data. Under 152-FZ, Russia’s personal data law (Part 5, Article 18), recording and storing them in databases outside Russia is not allowed, so the CRM, the agent’s server and the model work in Russia. How we do this is described on the page about data security.

The conversion metric

Agent conversion is the share of leads that reached the next pipeline step: a qualified client, a meeting or a payment. How to count it honestly, with a comparison group and the same period, we covered in the article about AI agent conversion.

In a pilot we write two numbers into the contract: the time to the first reply and the share of leads entered into the CRM. Both are easy to check in the log, and they depend less on the season than sales do.

Conversion to payment is checked later and compared with the period before the agent. If the agent replies in a minute but a manager calls back the next day, the gain in speed is lost on the human side.

Many look for an effect on profit, but few see it: according to McKinsey (November 2025), 39% of organizations see an impact of AI on EBIT, the company’s operating profit.

Risks to reputation

A sales agent speaks for the company, and its words count as yours. On 14 February 2024 a tribunal in British Columbia ordered Air Canada to pay a customer 812 Canadian dollars: the chatbot on its website had promised a refund under a discounted fare that the rules did not allow. The tribunal rejected the argument that the bot was responsible for its own words.

Hence the rules for a sales agent: prices and discounts only from the price list, terms and conditions only from approved text, and everything not in the rules goes to a manager.

Before launch, run the agent through provocations: a request for a discount, a question about a competitor, rudeness and an attempt to make it break its instructions. The head of sales checks the answers.

In the first weeks after launch, read a sample of chats every day. It is cheaper to catch a mistake in the log than to explain it to a client.

Sources and fact-checking

Checked by the editors on 16 October 2026.

  1. IT Brief, 11 November 2025 (itbrief.co.uk)
  2. amoCRM for developers, 2026 (amocrm.ru)
  3. Bitrix24, REST API documentation, 2026 (apidocs.bitrix24.ru)
  4. ConsultantPlus, 152-FZ “On Personal Data”, Article 18, 2026 (consultant.ru)
  5. Wikipedia, Moffatt v. Air Canada, 14 February 2024 (en.wikipedia.org)
  6. AI for sales teams: the pilot metric, MMOOZZGG.com

Found a mistake? Tell us

Publications

Contact us