DBS: Scaling AI in Singapore

DBS's Decade-Long Journey Towards Scalable AI Integration

Singapore's banking sector is at the forefront of technological adoption, and DBS Bank stands out with its strategic, long-term vision for artificial intelligence. For nearly a decade, the institution has meticulously built the foundational data infrastructure, robust systems, and comprehensive governance structures that now underpin its advanced AI initiatives. According to Eugene Huang, Group Chief Information Officer at DBS, the bank's foresight in preparing for the sustainable and scalable use of AI predates the widespread emergence of generative models. This proactive approach has positioned DBS to effectively leverage cutting-edge AI technologies.

Huang shared insights into the strategic decisions that guided this groundwork and how DBS is currently navigating the complexities of both generative and agentic AI.

Lessons Learned from Scaling AI: The DBS Experience

DBS embarked on its AI journey in 2014 with an initial pilot project involving IBM Watson for wealth management. This early exploration revealed a critical insight: data would be the bedrock of all future AI endeavours. Consequently, the bank dedicated several years to establishing the necessary architectural framework to support its data-centric AI strategy.

Key initiatives included:

• Technology Stack Modernisation: DBS focused on enhancing its technology stack to ensure greater scalability and stability. Automation was integrated into core processes, involving a significant shift from legacy systems to open-source technologies.
• Hybrid, Multi-Cloud Infrastructure: To bolster compute resources and flexibility, the bank invested in a hybrid, multi-cloud infrastructure. This move provided enhanced agility and capacity for its growing AI workloads.
• In-House AI Capabilities: Two pivotal in-house capabilities were developed to streamline AI deployment and management:
  • ADA (AI Data Architecture): This self-service platform functions as a unified source of truth for data governance, discoverability, quality, and security. It ensures that data used for AI is reliable and well-managed.
  • ALAN (AI Logic and Analytics Network): This AI protocol and knowledge repository standardises and makes repeatable the process of deploying AI models across various use cases, ensuring consistency and efficiency.

Today, DBS boasts an impressive deployment of over 1,500 AI models across more than 370 distinct use cases throughout the bank. While building this robust foundation was a significant undertaking, it has proven instrumental in scaling generative AI applications and preparing for the advent of agentic AI.

The bank has also been transparent about the economic impact of its AI efforts. Since 2021, DBS has been quantifying and disclosing these figures, with projections for the current year expected to exceed SG$1 billion. Beyond technology, DBS recognised that the human element was equally crucial. A concerted effort was made to ensure employees were actively involved and supported throughout this technological transformation.

Generative AI in Action: Impactful Use Cases at DBS

With its robust AI foundation in place, DBS was well-positioned to integrate generative AI capabilities in 2023 as the technology matured. These applications are now enhancing various customer and employee workflows across critical areas such as sales, advisory services, customer servicing, operational processing, and software development.

Notable generative AI use cases include:

• DBS Joy (Virtual Assistant for Corporate Customers): This virtual assistant leverages generative models to efficiently address customer queries, manage common requests, and facilitate routine servicing. For more complex issues, the system seamlessly escalates to a human service specialist, who is equipped with an internal co-pilot to provide comprehensive support and responses.
• Customer Service Officer (CSO) Assistant: Customer service representatives benefit from this tool, which automates transcription, call summarisation, and post-call documentation. This has led to a significant reduction in call handling times, by up to 20%.
• DBS-GPT (In-House Generative AI Platform): The vast majority of DBS employees, over 90% of the workforce, have access to DBS-GPT. This internal platform empowers employees with capabilities for writing, brainstorming, summarisation, translation, and efficiently retrieving information from the bank's extensive knowledge base.
• JIRA Assist (Software Development): Within the software development lifecycle, JIRA Assist aids developers and business analysts in refining code, generating documentation, and expediting bug fixes.

These generative AI tools are not merely about automation; they are designed to liberate employees from repetitive tasks, allowing them to concentrate on work that demands greater judgment, strategic thinking, and direct customer interaction.

A Modular Approach to AI and LLM Deployment

DBS employs a modular AI architecture, a strategy that enhances flexibility and reduces dependency on specific technologies. The ADA platform has been expanded to accommodate generative AI use cases by introducing a generative AI marketplace. This marketplace offers applications that securely leverage Large Language Models (LLMs) as a service, operating under strict controls and governance.

This architecture ensures that DBS is not tied to any single LLM provider or technology vendor, whether on-premises or cloud-based. The key advantages of this approach include:

• Interchangeability of LLMs: The architecture facilitates the integration and seamless swapping of different LLMs with minimal integration effort.
• Integrated Governance: The marketplace incorporates essential safety guardrails, audit controls, cost management features, and pre-approved deployment patterns.
• Reusable APIs: A suite of reusable APIs supports the efficient and standardised deployment of AI solutions.

This strategic framework has dramatically reduced the time to value for AI and machine learning projects, shrinking it from an average of 18 months to approximately 2 to 3 months. In 2024 alone, this efficiency contributed an estimated SG$750 million in economic outcomes.

Balancing AI Innovation with Robust Governance

Responsible AI is a cornerstone of DBS's strategy, ensuring that AI is deployed ethically, transparently, and in alignment with clearly defined principles. While recognising AI's immense potential to enhance customer experience and operational efficiency, the bank insists that its application must be guided by shared ethical guidelines.

All AI and machine learning use cases undergo rigorous review through the PURE framework, which mandates that data usage must be:

• Purposeful: Data must be used for a clearly defined and legitimate purpose.
• Unsurprising: Data usage should not be unexpected or misleading to individuals.
• Respectful: Data must be handled with respect for individual privacy and rights.
• Explainable: The rationale behind data usage and AI decision-making should be understandable.

The PURE framework is not merely a compliance checklist; it is embedded within DBS's AI and machine learning processes. Use case owners are required to consider the ethical implications of data usage alongside legal and technical permissibility. To foster this culture, new employees are introduced to the PURE principles during their orientation, integrating them into the bank's fundamental approach to data management.

As generative AI and more sophisticated agentic systems emerge, DBS is actively extending its responsible data use framework to encompass specific guidelines for these advanced technologies. Insights derived from each use case review are incorporated into ongoing updates, ensuring the framework remains current and effective.

The Role and Limits of Agentic AI in a Regulated Environment

Agentic AI, capable of autonomous action and decision-making, is poised to play an increasingly significant role in customer interactions. Individuals may come to rely on these systems for tasks such as information retrieval, product procurement, and payment management. This presents new opportunities for automating certain banking functions, though critical decision-making will continue to involve human oversight for the foreseeable future.

DBS aims to provide customers with a secure and convenient platform for conducting these transactions. A dedicated working group is currently exploring various use cases, with a keen focus on critical aspects like observability, spend controls, accountability, and liability. The objective is to strike an optimal balance between enhanced convenience and the inherent responsibilities associated with advanced AI systems.

By automating routine tasks within predefined parameters and safeguards, agentic AI has the potential to significantly reduce manual workloads. This, in turn, allows employees to dedicate their expertise to activities that require nuanced judgment and more complex problem-solving.

Preparing the Workforce for an AI-Driven Future

DBS is proactively investing in upskilling and reskilling its workforce to ensure employees remain relevant in an evolving technological landscape. Since the beginning of the year, the bank has identified over 12,000 employees for targeted learning programs focusing on AI and data analytics. The majority of these employees have already commenced their personalised learning roadmaps.

To support this transition, DBS is enhancing its approach to product and customer experience design by making them more data-driven. The bank is also actively developing specialised roles in data analytics and governance. Furthermore, tech employees are undergoing training to adapt to new roles that will involve integrating generative AI tools into their daily work, fostering a collaborative environment between human expertise and artificial intelligence.