Agency Turns Telegram Chats into a Steady Lead Channel with a 5,000-Ruble AI Bot

Telegram chats can be more than a place for idle chatter: they can also serve as a working sales channel. The author of the case study describes how an agency built an AI bot that tracks open business chats, picks out requests for services and delivers a steady stream of qualified leads for about 5,000 rubles a month.

Where to find demand

The idea grew out of a simple observation: industry Telegram chats regularly feature requests for contractors, advice and help. For B2B teams, it is almost a live bulletin board where entrepreneurs, managers and decision-makers talk. The problem is that these requests drown in thousands of messages. Reading dozens of chats by hand is expensive and slow, and assigning a dedicated employee solely to monitoring is not cost-effective. So almost everyone has the channel, but few use it systematically.

Instead of buying a ready-made service with unclear payback, the team decided to test the hypothesis in-house. A business analyst built a bot in a single working day. It listens to open chats from a shared database, runs messages through a set of filters and sends managers only what looks like a real need for a service. A human then decides whether to join the conversation and make an offer in a private message.

How the selection works

The main value of the system turned out to lie not in the mere fact of reading chats, but in the logic that cuts out noise. Telegram has too much junk: job postings, long advertising posts, small talk, jokes and mass mailings. So the bot works in several stages and does not try to push everything into the CRM funnel. First it cleans the stream, then it checks the meaning of the text, and then it assesses whether the message contains an actual request for help rather than just a discussion of a topic.

"The bot listens only to open business chats from my personal database

and from the database that employees compiled."

• The anti-spam filter removes messages with excessive emoji, advertising style and templated job lists.
• Lemmatization reduces words to their base form and helps catch the meaning even if the author writes with errors or in slang.
• A database of roughly 200 key lemmas links messages to the agency's services, from performance marketing to audit and analytics.
• Intent analysis looks for phrasings such as "need", "looking for" and "recommend" to separate real demand from chatter.

After scoring, the message goes through a final check with a prompt that lists the agency's services. If the model considers the request relevant, the bot sends it to a corporate messenger rather than straight to the CRM. This detail matters: managers get a compact queue for manual validation and can change the logic without a developer if needed. According to the author, commands and prompts live in Google Sheets, so employees add new services and key lemmas themselves and control whether the bot is switched on or off.

Economics and results

The project's budget is almost microservice-sized. Launch took about $100 of the business analyst's work, the server costs about $5 a month, and OpenAI expenses run $20–50 a month. Converted, that comes to roughly 5,000 rubles in monthly costs. By comparison, a dedicated manager or an outside developer would cost noticeably more, especially given the need for constant monitoring of tens of thousands of messages.

In output terms, the bot reads tens of thousands of messages a month, leaves about 800–1,000 after two levels of filtering, and then helps managers find 15–22 qualified requests a month. Of the messages that reach manual processing, 5–8% prove relevant, and among those 50–60% turn into leads for the agency. Because of the large volume of noise, the authors deliberately did not connect the tool directly to the CRM: otherwise junk would skew the statistics and eat into sales time.

What this means

The case shows that Telegram chats can become an inexpensive B2B lead-generation channel, provided you do not try to automate everything and leave the final decision to a human. The approach works best where clients sit in topic-specific chats: marketing, advertising, real estate, consulting. The main asset here is not the model itself but well-tuned filtering logic and a cheap operating setup that the team can maintain without constant developer involvement.