Generative AI Customer Service Copilot at a Neobank
A neobank with 4M customers receives 180,000 chat enquiries a month (fees, limits, blocked cards, transfers). 70% are answered with information that is already in 1,200 pages of terms, FAQs and internal procedures, which change every month. The COO wants to reduce the cost per contact without giving invented answers that lead to complaints to the regulator.
Where the money comes from
What to do now
- 1Organise the knowledge base and create 500 test questions with compliance.
- 2Launch as a copilot for agents (it suggests answers).
- 3Move to self-service only the topics with more than 95% accuracy.
For your committee
One page with the decision, the impact, the risks, the steps and the KPI to track.
The print dialog opens: choose "Save as PDF".
Watch out: more volume is not always more profit
Measuring only the savings: if the assistant answers badly about fees, churn and complaints rise. A closed account costs more than many enquiries saved.
KPI for the committee
Cost per contact, % resolved without a human and CSAT.
Detail for your data team
Proposed technique: LLM + RAG
Why: The task is not to classify, but to generate accurate answers about knowledge that changes every month. RAG (Retrieval-Augmented Generation) separates the knowledge (documents, updatable without retraining) from the language (LLM). Retrieval provides traceability: each answer cites its source, which is key for the regulator.
Cost of getting it wrong: An invented answer about fees can end in a formal complaint and a fine; escalating too much to a human keeps the cost up. The balance is managed with a confidence threshold and banned topics.
Data needed: Knowledge base: 1,200 versioned documents (terms, rates, procedures).; Chat history: 2M conversations with the human agent's resolution.; Real-time customer data (via API): account type, card status, latest transactions.; Post-contact surveys (CSAT) and reasons for escalation.