AI Chatbot for Your Website: a Consultant That Knows Your Catalog and Never Makes Things Up

A familiar picture: 30 leads a day land on the site, and twice as many questions land in the chat. Managers are drowning, evenings and weekends the chat goes silent, and the scripted “pick from the list” bots only annoy buyers. Foreign services like Intercom got expensive, handle Russian poorly, and expect payment methods from another world.

We build a different thing: an AI chatbot for your website on large language models, Claude, GPT or GLM, picked for the task and the budget. Not a button script, but a consultant that reads your knowledge base, answers naturally, walks the customer to a lead, and hands it to your CRM or messenger. Build and setup: from ₽30,000.

What it solves

• Response speed. The bot answers in seconds, 24/7, no weekends and no lunch breaks. While the manager sleeps, the bot qualifies the lead and records the question.
• Catalog knowledge. The bot answers from your knowledge base: prices, specs, stock, delivery terms. If the answer is not there, it says so honestly and calls a human instead of improvising.
• Language quality. Claude- and GPT-class models write naturally, with no machine-translation smell, and the tone is tuned to your brand.
• No lost leads. Every contact goes where your team works: Bitrix24, Automata CRM, Telegram, MAX or email.
• Independence. The API key stays with you, the code and the knowledge base are handed over, and the bot is not tied to us.

How it works

At the core is a large language model with a RAG pipeline (retrieval-augmented generation): before answering, the bot fetches the relevant chunks of your knowledge base and answers strictly from them. That is the difference between a working consultant and a “hallucinating GPT”: the bot quotes your price list, not the average temperature of the internet.

The knowledge base is your material: site pages, FAQ, product descriptions, internal documents, a Google doc or JSON. Updates take minutes: change a price, fix a term, and the bot already answers the new way.

The widget is a floating button or an embedded window, full-screen on mobile, with the greeting and color scheme matched to your site. It installs on 1C-Bitrix, WordPress, Tilda and headless front ends on Astro and Next.js, which means on any of our Automata Site platforms.

What the work includes (from ₽30,000)

• API connection for the chosen model (your key, or ours with a prepaid balance);
• collecting and structuring the knowledge base: site parsing, FAQ, your documents;
• prompt design: tone, topic boundaries, lead-walking scenarios;
• the website widget, customized to your brand;
• lead delivery: Bitrix24, Automata CRM, Telegram, MAX, email, webhook;
• test runs across dozens of scenarios until the answers are clean;
• documentation and a warranty period of fixes, with the terms fixed in the contract.

Complex integrations (1C exchange, custom CRMs, multilingual setups) are quoted separately after the brief; we honestly show what fits the base price and what does not.

What the API costs

Model usage is paid separately from the build, directly to the provider. At an average question flow it is usually comparable to a couple of coffees a month; the exact number depends on the model and the dialogue count, so at the brief we show an estimate for your traffic and pick a model that does not eat the budget. If the flow grows, the model can be swapped with one setting.

Which model to pick

We are not loyal to one brand. On real projects we use Claude, GPT and GLM, with three criteria in mind:

• Language and tone. Claude and GPT hold a natural dialogue; GLM wins on price.
• Context. For large catalogs the context window size and the price per million tokens matter more than a loud name.
• Availability and payment from your region. The keys have to work steadily and be payable without a dance. We will suggest a working option for your situation.

The final pick comes after the brief. Sometimes the honest and cheaper answer is a pair of models: a cheap one for the first line, a heavy one for the hard questions.

Timeline

A base bot with a knowledge base and lead delivery takes 5 to 10 business days. Bitrix24 or 1C integrations and multilingual setups add time; we give the exact estimate after the brief, and it does not change mid-project without your decision.

Four steps to a working bot

1. Brief and knowledge base. We collect the material and agree on the scenarios.
2. Model connection. API, prompt, RAG pipeline.
3. Training and tests. Dozens of scenarios, fixes until the answers are clean.
4. Integration and handover. Widget, leads, documentation, and the knowledge base under your control.

The bot is not a manager replacement, it is the first line: it clears the routine questions, qualifies the flow, and wakes a human only where a decision is needed. The customer gets an answer in seconds; the team gets ready leads instead of raw chat.

Want to see how the bot answers questions about your catalog? Send a link to the site, and we will build a demo on your data.

Phone: +7 (906) 311-77-69, email: hello@automata.sale, Telegram: @automatasale, site: automata.sale

Evgeny Uryadov, sole proprietor (INN 645112058391, OGRNIP 312645301900058). Working across Russia, Belarus and Kazakhstan.