Gorodenkoff - stock.adobe.com

How DBS is industrialising AI across its business

Southeast Asia’s biggest lender is building a strong data foundation and upskilling employees on data and artificial intelligence to realise its vision of becoming an AI-fuelled bank

With the fervour over artificial intelligence (AI) and growing pressure from corporate stakeholders to develop an AI strategy, it’s easy for organisations to fall into the trap of deploying the technology for its own sake.

That was not the case for DBS, Southeast Asia’s biggest lender, which, for the past seven years, had focused on making sure it had the right data foundation to support the big data and AI initiatives that were to come.

“We were doing things around big data even before big data was a term,” said Nimish Panchmatia, chief data and transformation officer at DBS. “And now with AI, we are focused very programmatically on developing the right skills and people to take us to the next level.”

That includes putting the right platforms and technology in place to not only drive value for the business, but also processes and guardrails to mitigate risks and ensure employees can access and use data in a responsible way.

To that, DBS has formed a Data Chapter that brings together 700 data professionals across the bank to scale up value from data analytics, AI and machine learning (ML). The chapter will also deepen domain knowledge and broaden exposure through cross-learning for the bank’s data professionals.

In a bid to future-proof employees with big data and data analytics skills, DBS has also developed a training curriculum that caters to different knowledge and skill levels across the bank – from novices whose day-to-day roles do not require them to have much interaction with data analytics, to data experts looking to sharpen their skills.

Employees have options to learn at their own pace and select what they need to improve on from a range of online courses, workshops and community programmes. Since 2021, more than 9,000 DBS employees have taken upskilling courses in data and AI.

Today, DBS is building on that data foundation to realise its vision of becoming an AI-fuelled bank. That means making AI pervasive across the business; reducing the effort and cost it takes to develop and deploy AI solutions; and delivering exponential outcomes through AI solutions and use cases.

So far, DBS has delivered over 350 AI use cases spanning customer-facing businesses including consumer and institutional banking, as well as support functions including finance and HR.

It has also accelerated the time to realise value from its AI initiatives from 12 to 15 months to two to three months, with a goal of reducing that to two to three weeks over the next few years. Over the past two years, DBS has more than doubled the economic impact of AI on its business from S$150m to over S$370m.

Generative AI initiatives

DBS’s early foray into big data and AI makes it well-poised to harness the benefits of generative AI (GenAI). The technology is already augmenting the way its employees work by handling routine tasks, allowing them to focus on more strategic and value-added activities, such as building deeper customer relationships.

“We’re working to see how we can help our developers speed up time to market, improve software quality and detect bugs, but this is augmenting rather than replacing the software engineer”
Nimish Panchmatia, DBS

It also launched DBS-GPT, an employee-facing version of ChatGPT, to help employees with content generation and writing tasks in a secure environment. Over 5,000 employees in DBS Singapore have used DBS-GPT, which is being progressively scaled up across the bank.

“We are using GenAI to extract information from documents, make sense of it, and then populate it into templates for new trade loans,” said Panchmatia, adding that relationship managers are also using the technology to pull together structured and unstructured data to build client proposals and analyse client portfolios.

Another GenAI use case is around software development, which Panchmatia said has been the most mature use case over the past year. “The other areas are still a bit experimental in nature … but it looks like a lot of people have made good progress with coding.

“We’re working to see how we can help our developers speed up time to market, improve software quality and detect bugs, but this is augmenting rather than replacing the software engineer,” he added.

For now, Panchmatia believes GenAI tools are not ready to be used autonomously, especially with clients, so the bank will experiment in “very contained environments” to ensure that sensitive information is not sent to the open web.

Other ways to mitigate issues such as model hallucinations include vectorising unstructured data and the use of retrieval-augmented generation to ensure outputs from large language models (LLMs) have references to source material.

“You can also set ‘temperature settings’ such that the model doesn’t try to create something for something it doesn’t know,” said Panchmatia, noting that the financial industry has a lower tolerance for inaccuracy than other industries like retail.

“While that takes away the elegance of an LLM because you’re restricting it to a small set of data rather than using world knowledge to augment your local knowledge, I believe that’s solvable over the coming months.”

Meanwhile, to address some of the risks associated with AI initiatives, DBS has developed an in-house framework to assess AI use cases before they are approved to ensure there is a purpose behind the use of the data, and that data use must be explainable, among other criteria.

It has also established a taskforce comprising senior executives who meet regularly to understand and address any gaps in the bank’s governance and control measures, particularly when it comes to GenAI initiatives.

“For things like copyright, let’s see how that evolves. But issues around toxicity, hallucination and appropriateness were not risks that existed in AI and ML before. We’re working through those, and in some cases, we’ll require humans in the loop,” said Panchmatia.

“We believe there’s enough technology to allow other models to run on top of the outputs of bigger models to ensure that things like appropriateness and toxicity are avoided,” he added.

Read more about artificial intelligence in ASEAN

Read more on Information technology (IT) in ASEAN

Data Center
Data Management