Key Highlights: What This Case Study Covers
- Transition from a rule-based chatbot to a Gen AI–led conversational assistant for financial advisory services.
- Dual-interface solution for customers and support agents to receive contextual intelligence.
- Integration of sentiment analysis and guardrails for higher accuracy and deeper ethical compliance.
- Real-time knowledge retrieval through Azure AI Search and RAG-enabled prompt flow.
- Scalable architecture supporting 100 concurrent users with sub-7-second latency.
Client Overview
The client is a leading mobile payment services provider based in Asia, offering digital transactions and financial management solutions to a rapidly expanding customer base. With growing user interactions and rising demand for instant financial insights, user engagement and agent responsiveness became their business-critical goals.
The Ask
The client wanted an AI-powered solution to improve customer experiences and agent productivity. They were looking to enable quick, accurate query handling, deliver financial insights, and provide consistent knowledge-based recommendations across all interactions.
Challenges
- Guardrail Integration Gaps: Lack of response controls made it difficult to manage competitor-related queries and maintain adherence to the approved knowledge base.
- Response Conciseness Issues: Challenges in maintaining brief, contextually accurate outputs affected user engagement.
- Custom Package Deployment Hurdles: Difficulty in implementing and integrating the Text Analytics Language API for sentiment analysis through Prompt Flow.
- Third-party Integration Limitations: Limited connectivity with the existing contact center application created integration constraints.
Our Solution: Gen AI-Powered Wealth Coach and Copilot
- Scope Identification & Prioritization: Designed for two user groups: end customers seeking investment advice and support agents using it as a copilot.
- Input Data: Data included historical contact-center interactions, SOPs, financial advisory content, company knowledge bases, best practices, and partner recommendations. Inputs were primarily PDFs and text files.
- Pre-Processing: PDFs were ingested using a pipeline with Form Recognizer for extraction and Azure AI Search for indexing and retrieval.
- Models Deployed: A sentiment model built on Azure ML evaluated customer feedback, while an intent recognition model using Azure OpenAI analyzed and classified user queries.
- LLM Tuning: Prompt tuning optimized LLM outputs for intent detection, responses, recommendations, and summarization. Prompt Flow orchestrated RAG and LLM nodes to retrieve documents and generate summaries within guardrails.
- System Interfaces: Two interfaces were deployed. The contact-center agent interface provided procedural knowledge, tracked intent and sentiment, identified similar tickets, and suggested sentiment-based responses. The end-customer interface delivered financial knowledge and advice by responding directly to user queries.
Impact Delivered
- Improved customer financial literacy and investment decisions through personalized Gen AI.
- Reduced average agent query resolution time, enhancing contact center responsiveness.
- Ensured scalability to support 100 concurrent users with an average latency of 7 seconds.
- Achieved 88% model accuracy across 6 parameters, with 100% compliance in ethics and behavior.
