Customer serviceAI assistant / LLM
Task. The support team of a large service provider was overwhelmed by repetitive requests: customers waited up to 40 minutes, operators burned out, requests were lost outside working hours.
Solution
We built an LLM-based support assistant on the corporate knowledge base: it replies in chat and messengers 24/7, closes typical requests itself and hands complex ones to operators with full dialogue history.
Results
>90%assistant answer accuracy
3×lower load on tech support
24/7request handling without night shifts
Technologies: LLM (Qwen/DeepSeek), RAG, AIHub, FastAPI, Docker.
Project stages
- Request analysis (1–2 weeks) — support-ticket history mining and topic clustering.
- Assistant setup (2–4 weeks) — knowledge-base training, tone-of-voice alignment, channel connection.
- Escalation logic (1–2 weeks) — seamless handover to human agents for complex cases.
- Launch & learning — continuous improvement on real conversations.
Where this applies
AI support assistants serve e-commerce in Kazakhstan, banks and telecoms in Uzbekistan, and service companies in Kyrgyzstan and Tajikistan — with multilingual coverage in Russian, Kyrgyz, Kazakh and Uzbek. See AI assistants for business.