At the Oversea-Chinese Banking Corporation (OCBC), generative AI is not just hype or an experiment. “Generative AI has democratised artificial intelligence (AI). In the past, the business units/leaders may not understand how AI can help solve their problems but today, everyone has a greater understanding of what’s possible. So, the questions that I get from business units/ users now are much sharper as they realise the value of AI,” says Donald MacDonald, OCBC’s head of Group Data Office.
Generative AI is also transformational as it can help increase speed to market. He explains: “The large language models (LLMs) used in generative AI are general-purpose so one model can be repurposed to handle different tasks instead of having to build [dedicated] AI model for every task. It is also good at zero-shot learning, wherein it can solve a problem without being specifically trained or fine-tuned to handle that problem.”
Boosting employee productivity
OCBC’s generative AI projects focus on improving employee productivity. This includes OCBC GPT, a chatbot powered by ChatGPT’s LLMs and is now available to OCBC’s 30,000 employees globally via Microsoft Teams. Users can simply key in their query in natural language and the chatbot will produce a detailed response based on accessible text-based information on the web.
To safeguard its data, OCBC GPT is hosted in a secure and controlled environment, and information entered by OCBC staff will be kept within the bank. “We see this as an easy quick win because it democratises access to knowledge and simple content creation across all employees, but doing so in a secure environment [with little to] no risk of data leakage for the bank,” MacDonald states.
This bank-wide deployment comes after a six-month trial earlier this year, which saw 1,000 OCBC employees across multiple functions using OCBC GPT to write investment research reports and draft customer responses, among other use cases. On average, participants who leveraged OCBC GPT managed to halve the time they normally took to complete a task, even after including the time taken to check the chatbot’s outputs for factual accuracy.
See also: Alibaba anoints new chief in revamp of stalling commerce arm
An employee using OCBC GPT. Photo: OCBC
MacDonald shares that OCBC has also developed and deployed its own generative AI productivity tools. One of them is OCBC Wingman, a coding assistant that auto-generates, debugs and improves computer code. This standardises code quality, ensures the code does not leave the bank’s environment, and saves 20% developer effort during code build. The latest version of the tool (i.e., version 3) leverages a fine-tuned LLM that allows developers to use plain English to talk to, refactor, debug and document code. Today, OCBC Wingman writes about half a million lines of code per day.
See also: Break up Google? What’s at stake in antitrust action
The next tool is Document AI, which extracts and summarises key information from documents (like financial and sustainability reports) through a drag-and-drop process. As such, frontline employees can find the information they need in 1 minute per document, compared to 30 minutes previously. “They can ‘talk’ to their documents and pull out the necessary data quickly. Since the information will be automatically presented in a structured table, employees no longer need to manually copy and paste information or links into say a Word document,” says MacDonald.
As for OCBC Whisper, it is a speech-to-text tool that can analyse all sales calls with customers to automatically identify potential anomalies in the sales process. Besides that, the bank is currently trialling the use of OCBC Whisper to transcribe and summarise calls in real-time at the bank’s contact centre. MacDonald claims that the tool can transcribe with at least 90% accuracy. He adds: “If the customer switches language midway through the call, OCBC Whisper can still handle that and translate into English. We’re also working with Nanyang Technological University to look at how we can fine-tune the model to be even better at understanding Singlish.”
OCBC Buddy is another notable generative AI-powered tool. It is an internal knowledge base that functions like a chatbot for OCBC’s employees to get quick answers on their company policies and information such as medical claims and annual leave. “OCBC Buddy crawls over 150,000 pages of the bank’s intranet to give relevant answers to employees’ questions, which may include links to source documents,” states MacDonald. On average, it is used more than 30,000 times a month by all employees.
In addition to being an internal chatbot, OCBC Buddy can also record face-to-face meetings when it is accessed via the bank’s mobile app for employees. It will email a full transcript of the meeting to the meeting owner at the end of the meeting.
Foundation for AI
According to research firm International Data Corp, two-thirds of organisations in Asia Pacific (excluding Japan) are exploring potential use cases or are already investing in generative AI technologies this year. Yet, not many banks have deployed generative AI across the entire organisation.
MacDonald attributes OCBC’s success in scaling AI enterprise-wide to the bank’s early and committed investment in developing a solid foundation for AI. “Building our own modern data infrastructure and deep talent pool over the years has helped us to scale quickly, especially when generative AI technologies started to pick up speed after ChatGPT became widely popular. In fact, our coding assistant OCBC Wingman was built and launched in a short turnaround time of five days. Organisations without [the right infrastructure and skills] already in place would struggle to move as quickly.”
To stay ahead of the latest tech trends, click here for DigitalEdge Section
The bank’s hybrid LLM strategy — wherein it uses a mix of private models and open-source models — is also crucial. MacDonald shares that his team has built an LLM sandbox that brings in open-source models as they are released, which can be as frequent as every week. That way, OCBC’s data scientists can test those new models (for effectiveness, accuracy and more) and replace existing LLMs that power different applications when necessary, without users knowing that the application is using a new LLM.
Besides that, OCBC’s model management platform (MMP) ensures the bank can effectively manage LLMs at scale as AI is increasingly being deployed and used across the organisation. MacDonald explains: “MMP allows us to monitor LLMs deployed in OCBC and see all the documentation and approvals associated with the models. MMP is also critical in ensuring responsible AI. Every time a model is running, the logs and the results are streamed automatically to MMP so we’re able to analyse things like the fairness of every model. For example, are we being biased against any groups in the population? We also use the platform to monitor the performance of LLMs, and get alerted if a model starts to drift from expectation, [whereby its predictions get less accurate over time so that we can quickly investigate and rectify the issue.] When you’re doing AI at scale, you need to automate all these processes.”
He adds that the bank has also put up additional guardrails to prevent AI hallucination, in which an AI model generates false or illogical information but presents it as fact. This includes having its own LLM evaluation framework to test models before putting them into production, prompt sensitivity testing, and using LLMs to check other LLMs to ensure the accuracy of their results.
Since a tool is only as good as its users, OCBC launched two virtual training modules for all employees on the basics of generative AI in September 2023. “We also organised training sessions for staff who want to learn about using OCBC GPT effectively, such as how OCBC GPT works, the optimal way to write prompts and the limitations of generative AI,” says MacDonald.
Specific to data or AI professionals, OCBC has a Data Certification Pathway to groom data professionals in-house through 12- to 18-month training. More than 300 staff have completed the programme since its launch in 2019. The bank also offers postgraduate AI scholarships to nurture a pipeline of AI talent.
