Donald MacDonald, head of group data office at OCBC, discussed with The Asian Banker the extensive integration of generative AI and large language models (LLM), the technology behind OCBC since 2019, emphasising their evolution from first-generation models like BERT and FINBERT to more advanced systems like GPT.

OCBC has its inclusive approach, enabling all employees to access GPT within Microsoft Teams, thus enhancing the capabilities of 30,000 staff members. The introduction of 'Buddy,' a chatbot that efficiently navigates the bank's vast internal resources of more than 150,000 pages, and 'Wingman,' a coding assistant that increases developer efficiency by 20%, exemplify OCBC's innovative use of generative AI. These tools, combined with a strategic focus on internal productivity and role-specific applications, allow OCBC to leverage GenAI for material process improvement and organisational transformation.

ChatGPT was a eureka moment not so much for data scientists but for the business said MacDonald. “The greatest benefit of the GenAI boom was the impact on business users who understood and were able to interact with AI through an interface they could understand, realising its true power.”

MacDonald regards productivity, risk management, and revenue generation at the front end as the three major areas in looking at GenAI use cases. However, he clarifies that 80% of the work at OCBC is still traditional, using machine learning tools for rational decision-making such as propensity modelling or financial crime and attrition.

"GenAI is helping us more with the right side of the brain tasks, creative decision-making, which is a smaller subset of problems in particular based on unstructured data, where banks have seen rather limited success in mining those prior to GenAI. Where he really sees the potential is in combining traditional and GenAI. The latter is good at making sense of unstructured data, which can then be turned into more structured signals to feed into traditional machine learning."

OCBC aligns GenAI's potential with ROI and prototyping capabilities by focusing on the substantial incremental value it can bring over other AI tools, rather than the technology itself. The bank has applied machine learning tools for years, dedicating significant time to identifying and prioritizing use cases, considering the time required to develop models and their return on investment.

To manage costs, OCBC uses GenAI largely on-premise using open-source tools. MacDonald highlights that based on OCBC’s large GPU infrastructure, the incremental cost of running GenAI projects is effectively zero. He sees cost as less of a consideration, and in most cases, GenAI reduces project costs.

"Because we use general models rather than custom-made models, use cases are faster to work on compared with projects where traditional machine learning is involved, hence reducing overall manpower costs and the cost of analytics. The first version of ‘Wingman’ was built internally with open source and deployed in less than two weeks at near zero cost. We’re rolling it out progressively to about 2,000 developers across OCBC", he said.

What is the strategic roadmap for genAI and its impact beyond productivity and speed? Currently, MacDonald says it's early days for this technology, and focusing on internal productivity use cases are quick tactical wins. He believes that GenAI is not yet fully ready to be exposed to customers because of the issue around hallucination.

MacDonald says he will be systematically equipping all of OCBC's workforce with these foundational capabilities, whether for the contact center, the compliance team, or sales. Eventually, he believes GenAI will be applied to customer-facing use cases such as smart IVR and voice banking—once better controls are in place to manage existing flaws.