The Challenge
Repetitive questions consumed 70% of agent time, and hiring operators to keep up was neither affordable nor scalable.
The Solution
A retrieval-augmented assistant on Azure OpenAI that answers from the company knowledge base rather than generating, with a clean handoff to a human.
📋 Overview
A fast-growing fintech could not scale its customer support team at the rate it was acquiring users. A generative AI assistant automated first-line support while keeping a human, empathetic tone.
🎯 The Challenge
Business context As the customer base grew, repetitive questions — “how do I change my IBAN?”, “what are the fees?” — consumed 70% of human agents’ time. Response times were degrading, which hurt customer satisfaction and churn. Hiring operators in bulk was neither financially viable nor scalable.
Technical requirements
- Natural understanding: interpret questions in everyday language, including slang and typos.
- Context: hold the thread of a conversation.
- Safety: never hallucinate an answer, least of all on sensitive financial matters.
- Handoff: hand over gracefully to a human when the assistant does not know.
💡 The Solution
Strategic approach Azure OpenAI Service (GPT-4) in a retrieval-augmented generation architecture. The assistant does not invent answers; it consults the company knowledge base — PDFs, FAQs, manuals — to ground every response.
Technical implementation
- Knowledge base: all support documentation indexed in Azure AI Search.
- Orchestration: LangChain on Azure Functions manages the conversation flow.
- Guardrails: strict rules preventing the assistant from giving financial advice or straying from the brand’s tone of voice.
- Channels: deployed to the website and the WhatsApp Business API.
Timeline
- Phase 1 (data preparation and ingestion): 2 weeks
- Phase 2 (prompt engineering and testing): 3 weeks
- Phase 3 (channel integration): 2 weeks
- Phase 4 (live pilot): 2 weeks
📈 The Results
Success metrics
- 🤖 Deflection rate: 45% of contacts resolved without human intervention.
- ⏱️ Response: immediate, against a previous average of 2 hours.
- 😊 CSAT: satisfaction with automated support at 4.2/5, comparable to basic human support.
Business impact The human support team stopped answering trivial questions and moved to complex fraud and dispute cases, where human empathy cannot be replaced. Cost per ticket fell sharply.
🛠️ Technology Stack
- Azure OpenAI Service: the linguistic engine.
- Azure AI Search: the corporate memory, as a vector store.
- Microsoft Bot Framework: the channel layer for web and WhatsApp.
💬 Client Testimonial
“We were sceptical about putting a ‘robot’ in front of customers. But the quality of the answers is impressive. It is like having our best agent available at three in the morning.”
COO Fintech
Ready to achieve similar results?
Contact us to learn how we can help your business turn challenges into opportunities.
