Problem
A cloud provider's website chat loses visitors outside working hours: someone writes after 11 pm, nobody answers, and the lead is gone. A rule-based bot does not help much: users do not leave their data with it, and nobody checks that the visitor is a real prospect. The chat also receives very different questions, from "what is inference as a service" to support problems, so every message needs sorting before anyone can act on it.
Example from the article: at Cloud.ru a visitor asks which database to choose for RAG. The assistant asks qualifying questions, then creates an opportunity in the CRM with everything the user said, so a sales manager can call with a specific offer.
Solution in brief
Put an LLM assistant behind the chat webhook. A classifier splits requests into sales and support. Sales requests go to a LangGraph state machine that follows the SPIN question list, answers product questions through RAG and writes the result to the CRM. Support goes to the existing support bot.
Questions this case answers
- How to let an LLM assistant run a fixed qualification dialogue with inbound visitors when no operator is available and write the qualified lead to a downstream system
- How to route incoming messages between sales and support with an LLM classifier before a stateful dialogue starts
- How to build an LLM assistant that qualifies website chat visitors and creates CRM leads while sales is offline
- How to route inbound chat requests between sales and support with an LLM classifier
Tools and methods
- LLM classifier (GigaChat API) — labels each request as sales or support
- LangGraph FSM with SPIN — runs the qualification dialogue and calls the tools
- RAG tool: bge-m3, Qdrant, bge-reranker-v2-m3 — answers product questions from the multilingual knowledge base
- FastAPI gateway, Redis, ClickHouse — auth and logging, session state, metrics
- CRM and Naumen service desk APIs —
create_leadandcreate_tickettools
Google or a one-time email code. No password.