HLB now has a chatbot for customers to ask questions and have their banking needs addressed quickly and efficiently. Named Marina, the chatbot is a significant AI use case that moves digital banking beyond simple customer support into true transactions. The 24/7 chatbot reduces agent workload and enables faster resolution times, resulting in a higher transaction volume and improved customer satisfaction.

Deploying the chatbot was challenging at first; the team's initial open-source Large-Language Model (LLM) choice led to a model constrained by a static knowledge base and pre-defined answers. This approach resulted in slower response time and lacked the rapid processing required for a multi-agent workflow. HLB made a strategic pivot to Gemini 2.5 Flash for its dynamic Retrieval-Augmented Generation (RAG). This move immediately boosted the chatbot's accuracy from 75% to an industry-leading 99%, building customer trust and driving a 3x increase in digital engagement of monthly chat transactions.

Marina's architecture uses Gemini for enhanced intelligence, enabling delivery of accurate, contextual, and multilingual responses in English, Bahasa, and Chinese. The system ensures high-accuracy responses and transactional capabilities through two complementary processes: Dynamic FAQs and real-time transactions.

For pre-login general questions, Marina instantly accesses the most current information by dynamically searching HLB's public website using Google Search, instead of relying on a slow, manually updated database. This capability frees developers from painful manual knowledge base indexing, ensuring the bot's public information is always current.

For post-login personalized, complex tasks like checking a customer's credit card balance, Gemini acts as the control "brain" within the Agentic RAG framework to analyze the user's intent and determine the response. Using the Agent Development Kit (ADK), Gemini commands specialized agents to execute secure calls to HLB's backend APIs and pull real-time data directly from the core banking system. With this real-time access, the chatbot can then analyze interactions and credit history to proactively recommend credit card promotions. As a result, it has been able to handle 70% of chat volume, allowing human agents to shift their focus from high volume work to high complexity work.