70 % of enquiries resolved without an agent
Support bot for a bank
- Client
- Severny Bank
- Industry
- Finance
- Year
- 2025
- Type
- Bot
- Duration
- 4 months
- Team
- 5 people
01 — The problem
The call centre handled 14,000 enquiries a month with peak wait times of seven minutes. Yet 60 % of questions were about balances, application status, branch opening hours and product terms — all of it already written down in the knowledge base.
02 — What we built
We split questions into two classes
Simple, unambiguous ones ("where is my nearest branch", "what is my balance") are handled by deterministic logic: fast, free, no chance of being wrong. Free-form questions go to a language model with access to the bank's knowledge base and nothing else. That was a deliberate choice: in finance you cannot let a model invent loan terms.
The model answers only from documents
Claude receives the question plus excerpts from the knowledge base, and is instructed not to answer beyond them. If the answer is not in the documents, the bot does not improvise — it hands over to an agent. Every answer carries a reference to the source section.
Secure authentication
Balances and application status require an SMS code first. No personal data is stored bot-side and sessions expire after fifteen minutes. The project went through the bank's security review, which took five weeks and two rounds of changes.
Handover with context
When the bot cannot help, the agent receives the full conversation and an already-authenticated customer. Nobody has to ask who you are and what the problem is all over again.
03 — Result
- 70 %
- of enquiries resolved without an agent
- 14,000
- monthly enquiries through the bot
- 9 s
- average response, down from 7 minutes
- 0
- findings in the security review
«We originally wanted the AI to answer everything. Progress refused and explained what that risks in a bank. The model now covers a third of scenarios, and in a year it has never given a wrong answer about product terms.»