DBS Bank

DBS Bank, Southeast Asia's largest bank by assets, operates a large-scale contact center where customer service officers (CSOs) handle high volumes of inbound calls daily. Each interaction required agents to manually search for solution recommendations and document call details after the fact — a time-intensive process that reduced throughput and introduced inconsistency across the team. In an industry where a single error carries serious regulatory and reputational consequences, slow or inconsistent service creates both operational drag and compliance risk. As call volumes grew, the limits of manual lookup and documentation became a structural bottleneck the bank needed to address at scale.

DBS deployed a CSO AI Assistant that operates in real time alongside agents during live customer calls. Built on large language model (LLM) reasoning, the system automatically surfaces relevant solution recommendations as conversations unfold, eliminating the need for agents to manually search internal knowledge bases. It also generates automated call transcriptions, replacing post-call documentation work. Rather than replacing human judgment, the assistant augments it — keeping agents in the loop while reducing the cognitive load of information retrieval. DBS integrated the system through API-based infrastructure on Vertex AI, enabling model-agnostic deployment and making it straightforward to update underlying models as capabilities improve. The rollout was validated using a test-and-control methodology, with AI-treated groups measured against control groups to isolate actual performance impact.

The CSO Assistant delivered measurable efficiency and quality gains across the contact center:

• Up to 20% reduction in call handling time per interaction
• 95%+ accuracy in solution recommendations and automated call transcriptions
• ~$1 billion in total AI value reported by DBS across the bank in 2025, from combined traditional ML and generative AI applications

Beyond the headline numbers, the assistant enabled agents to focus more attention on the customer rather than system navigation. Consistency of service improved as recommendations became standardized and accurate. The test-and-control measurement approach gave DBS high confidence that reported gains reflected actual AI impact rather than confounding factors.

• Scope AI to well-defined workflows first: Achieving 95%+ accuracy is realistic when the task is bounded — solution lookup and transcription are structured, low-ambiguity problems well-suited to LLMs.
• Measure with test-and-control rigor: Splitting agents into AI-treated and control groups is the most defensible way to isolate real value and build executive confidence in reported outcomes.
• Invest in internal talent with domain context: External AI specialists without organizational knowledge struggle to apply models to specific processes — DBS built a 700-person data chapter embedded in business units to solve this.
• Design infrastructure to be model-agnostic: API-based deployment on platforms like Vertex AI lets teams swap models as the technology improves without rebuilding applications from scratch.