Problem
Each Avito sales manager handles more than 100 sellers and offers them personalized paid products, such as search promotion. Offers come from heuristics and an uplift model, but managers call only the top and part of the middle of the list. The long tail expires, because every offer has a lifetime (TTL), and an expired offer is lost revenue. Also, sellers who agreed by phone but did not pay need another contact, which costs manager time.
Example from the article: a "competitive analytics" offer tells a seller that competitors in the category get on average 6.33 times more views and proposes promotion for 10000 rub.
Solution in brief
Avito added a voice LLM assistant that calls the segment managers cannot reach before the TTL ends, and follows up with sellers who agreed but did not pay. A router agent hands each call to a product sub-agent, which can ask an analytics sub-agent that has seller statistics and RAG. Quality is scored with the same binary Sales Quality criteria as human managers, partly by an LLM judge.
Questions this case answers
- How to run an outbound voice LLM assistant that contacts a long tail of users before each offer expires and handles interruptions, answering machines and hallucinated facts
- How to evaluate a conversational sales assistant with the same binary criteria used for human agents, partly with an LLM judge
- How to automate sales calls to the long tail of personalized offers that human managers cannot reach
- How to build a voice multi-agent LLM assistant for outbound calls
Tools and methods
- Uplift model with heuristic offer triggers — selects and ranks offers
- Scheduler — calling queue
- Voice platform (ASR, TTS, VAD) — real-time speech
- LLM multi-agent system — router, product sub-agents, analytics sub-agent
- Feature Store and RAG — seller stats, product docs
- Sales Quality criteria and LLM-as-a-judge — evaluation shared with human managers
Google or a one-time email code. No password.