A chat assistant in Azerbaijani
An assistant that answers from the bank’s own documents, runs in the bank’s data centre and hands over to an agent.
- Client
- A bank’s contact centre with 140 agents
- Industry
- Banks and lenders
- Went live
- May 2026
- Duration
- 5 months
- Team
- 5 people
What was there
Most chats asked about card limits, transfer fees and branch hours, and at peak times the wait passed 3 minutes. The bank would not let client data reach an outside service, which ruled out ready-made cloud products.
What we built
- An assistant on an open-weight language model, on the bank’s own servers
- Search across tariffs, rules and branch documents, with the source under every answer
- Hand-over to an agent for card and account questions, with the chat history
- A weekly review panel where two agents mark wrong answers
- A test set of 1,900 questions in Azerbaijani and Russian
Results
34%
of chats closed without an agent
June to September 2026
3:40 → 0:50
average wait in chat, minutes
peak hours, 2026
0
client records leaving the bank
since launch
The assistant says when it does not know and calls an agent. That was our main condition.
Head of the Contact Centre, a bank’s contact centre with 140 agents
Technologies
- Python
- PyTorch
- PostgreSQL
- Kubernetes
