🤖 AI Assistant on Avito: How to Replace a Manager During Peak Hours

If you think AI agents are something distant and abstract, here's a real case with numbers from one of our students. A clothing store on Avito with a monthly turnover of 3 million rubles implemented an AI assistant to handle inquiries — and the results are impressive.

💡 What is an AI agent?
It's a program that can independently think, make decisions, and perform tasks. Unlike a regular language model, an AI agent has access to auxiliary tools and can consciously use them.

In this case, the AI agent:
- Reads customer messages
- Understands their questions (sizes, delivery, price)
- Searches for the needed information in a database of ~500 products
- Generates a response based on business data
- Sends the response to Avito

🎯 The Task
The manager couldn't keep up with 30-40 inquiries per hour during peak times. If you don't respond quickly, the customer leaves. The goal was to free the manager from routine tasks and automate inquiry handling.

⚙️ How It Works
Webhook from Avito → Message transformation → Data loading from DB → Token verification → Filter (only from customers) → AI generates response → Send back to Avito.

But there was a problem: a customer sends 3 messages in a row within 10 seconds — the assistant responds 3 times. It looks strange + extra token costs.

Solution: implementing Redis, which merges all customer messages within a few seconds into a single message — and the assistant responds once, but to all questions at once.


📊 Results
- Response time: 5-15 minutes -> 1-5 seconds
- 5-10 inquiries/hour -> 30-40+ inquiries/hour
- Manager: 10h/day -> AI agent: 24/7
- 50,000 rubles/month -> 10,000 rubles/month


Over 4 weeks of testing, "unanswered" inquiries disappeared while maintaining payment conversion. The number of payments even increased slightly due to full demand processing.

🔧 Bonus Feature
When a customer writes with a collaboration proposal or non-standard questions, the assistant redirects the conversation to the owner via a Telegram notification.

⚡️ The human is not completely removed from the process — they are simply enhanced by automation. Control is maintained, but peak loads are no longer a problem.

😎 Want to build such automations? Join our course.

👩‍💻 Data Flow