SalesAgent’s creator is giving away his AI salesperson: the bot passes 197 out of 200 catalog tests

A familiar scene: you want to check a price in a store’s chat, but the widget first demands your name, phone number, and email address. It then promises to respond “within a minute” and falls silent. While you wait, a competitor in the next tab has already answered. SalesAgent’s creator calculated the cost of losing visitors this way and made an unexpected decision: give the product away for free.

What a lost visit costs

Since the summer, the creator has spent about 200,000 rubles on Yandex Direct. Each visitor who performed any action on the website, from clicking a form to registering, cost more than 700 rubles. At the Belarusian plumbing store where his bot now operates, more than half of those who wrote in the chat arrived through paid advertising. If they closed the window when asked for a phone number, the money was wasted.

SalesAgent closes this gap in a straightforward way. You enter a website address, and the bot crawls its pages and builds a knowledge base, a priced catalog, a business profile covering delivery, payment, and warranties, and a sales flow. For a clinic, the conversation leads to an appointment; for a store, to an order; and for a manufacturer, to a contact request for a quote.

There is no need to write prompts: the bot answers in two or three sentences and does not turn the conversation into a questionnaire. It works in Telegram, MAX, WhatsApp, Instagram, and Avito, while leads can be sent by email or to Telegram, Bitrix24, or Hollihop. In four months, it has been tried on more than 300 websites.

Why it is not RAG

The most interesting part for engineers is how the answers are produced. The first version worked like most neural-network bots: search found similar fragments of the website, and the model wrote an answer based on them. That was enough for delivery questions. For product questions, however, the model mixed up prices of neighboring listings and recommended items that were not in the catalog.

The model no longer searches the catalog. For every user message, it writes a short program in an internal query language; the code then validates and executes it against the catalog and knowledge base. Conditions are represented as a tree of logical operations, much like SQL. The model formulates its answer only from the result: the found listings and calculated facts, such as how many options match and which is cheapest. Search understands Russian morphology, so the feminine adjective “belaya” matches the value “Bely,” as well as paraphrases: a visitor writes “lyuchok,” while the catalog calls it a “sewer access hatch.” Before every release, more than 200 test scenarios are run; 197–198 currently pass, or about 99%.

The practical lesson extends beyond one product: wherever an answer must be precise—prices, quantities, or comparisons—let the model formulate the query instead of composing prose from page fragments.

Why it is free

The creator’s explanation is direct: someone who arrives through an ad for an AI service has already seen dozens of similar products and does not trust their promises, while the bot’s performance on the customer’s own prices becomes visible only after payment. He needs thousands of real websites and conversations, so the entire product is being offered without reduced functionality or a trial period.

With a key for DeepSeek, GLM, or any OpenAI-compatible provider, responses do not count toward a limit, and you pay the provider directly for tokens. If the key stops working, the response is routed through the creator’s model within the free allowance, and you receive an email. Without a key, 100 responses are available every 30 days. These terms are permanent and have no end date; paid plans remain available for customers who find it easier to buy a response package.

What to do with it

If you are a developer, sooner or later a clinic you know or a client will ask you to “build us a chatbot.” There is no need to assemble RAG, a catalog, and prompts: a website address, an API key, and one line of code before </body> are enough. A referral program pays 30% of customer payments. For your own product, the bot reads documentation like a regular website, so you can test the idea of “adding AI” in real conversations before committing sprints to it.

There are limitations too. This is the product creator’s own account; figures such as 99% on internal tests and 300 websites have not been independently verified. Outcomes depend heavily on traffic: paying businesses average 41 conversations per month, while the rest average one. More than 40% of collected phone numbers remained in the dashboards because their owners had not configured integrations. And the bot knows only what entered its database: if a page blocks crawlers, the relevant fact must be added manually.