DBS, the largest bank in Southeast Asia, has won multiple global banking awards, including being named Global Bank of the Year by The Banker, World’s Best Bank by Global Finance, and World’s Best Digital Bank by Euromoney. Data and technology have been crucial to the bank’s success. DBS employs around 1,000 data scientists, data analysts, and data engineers and has roughly twice as many technologists as bankers.
Sameer Gupta is DBS’s chief analytics officer, a pivotal role in a business that has long recognized the potential of AI. In his interview with McKinsey’s David DeLallo, Gupta describes the bank’s AI journey, the importance of balancing boldness with pragmatism, and his mission to industrialize AI across the development life cycle. An edited version of their conversation follows.
David DeLallo: How would you describe your vision for AI at DBS?
Sameer Gupta: A few years ago, when we started this journey, our vision was to maximize outcomes from data. That is still our aim today, only we refine the outcomes yearly based on future casting. This year, what we’re asking at the bank level is: “How can we be an AI-fueled bank?” What that means in practice is that AI is pervasive across all parts of the bank. Our business model needs to be built with AI at the center, with the outputs informing processes across the bank.
David DeLallo: How far along that journey do you think you are?
Sameer Gupta: Our vision has been a few years in the making, and I think we’ve come a long way. Changing the company’s mindset and building an entire architecture of responsible usage of data is not easy. Infrastructure, processes, and systems are only parts of the whole equation. We have amassed a significant base of technical specialists and data scientists, and they have been doing cutting-edge work, creating more than 250 analytics use cases. Skills and capabilities are just as critical for any organization to be able to tap AI to drive data-driven insights. We are focused on enabling employees to think of data first and empowering them with the capabilities and tools to effectively use data and analytics while making decisions. We have revamped our approach to data management and how we look at metadata, security issues, and responsible use. We call it “enable data.” In addition, data platforms are built from scratch, and we now have about 90 percent of our most used data in that single platform. We continually add more features, functionality, and data assets to that platform.
David DeLallo: How are you helping to embed these practices across the business?
Sameer Gupta: We started a program called “Managing Through Journeys,” where we focus on delivering on a joint set of outcomes as a team, with the customer as the cornerstone. We ran some pilots last year that proved successful, from both an employee standpoint and an outcome standpoint. Now we’re looking to scale that up and embed these processes into the way we operate. As we reimagine the way we operate and change the way we work, fundamental shifts in mindsets, behaviors, and practices will take time. However, we are confident that we will make substantial progress over the next three to five years.
The other goal we are working on is AI industrialization, namely, how we continue to reduce the time it takes to develop, deploy, monitor, and refactor AI models. When we began our journey seven or eight years ago, it took us roughly 18 months to deploy AI from start to finish. Now we are close to four or five months for fresh end-to-end deployment, and we can build some features in just a couple of months. Our goal is to bring AI development down to just a few weeks in most cases. Getting to that point will be essential for us to scale AI, and achieving that will take a lot of experimentation.
David DeLallo: How do you define AI industrialization? Is it synonymous with MLOps? And what are some of the aspects of your approach to it?
Sameer Gupta: AI industrialization encompasses a machine learning model’s entire life cycle and aspects that others might define as MLOps. It involves everything from identifying the problem we are trying to solve, procuring the data, building the model, and testing the techniques and the feature marts, to deploying the model and then continuously monitoring and refactoring.
We started with AI industrialization about three or four years ago. Our focus is predicated on reducing the time and effort it takes to go through this entire cycle. For that, we need to bring people, processes, and technology together. In the beginning, some of our data scientists were very focused on building models, but not deploying them, so they would pass the work along to a deployment team. And the coding wasn’t built for scale. So we had to work with our data science teams and provide them the tools and upskilling to adapt to this new way of working.
On the process side, we had to find ways to build on each other’s models. We created an in-house AI protocol that serves as a knowledge repository for all our AI use cases. It has more than 250 use cases, and anyone can search it. If someone is working on “attrition,” for example, they can search for relevant tools, techniques, and data features. They don’t have to build from scratch. Creating reusable assets was another focus. We wanted the low-code assets that the data teams have created to be leveraged in multiple use cases, including natural language processing.
Technology, of course, is the third linchpin. With our data platform, we’re continually refining our analytics and operational clusters, as well as our data factory, in order to increase the throughput and productivity of all our data teams.
Feature marts are also a big part of AI industrialization for us because they cut down on time and increase reusability. A feature can include anything from behavioral data to product data. In the past, every use-case team built their own. Then we said, “Look, you can leverage this element here.” Over the years, we have fine-tuned these features and made them available in a repository so everyone can use them.
David DeLallo: What prompted you to decide to build your own data platform?
Sameer Gupta: When we started four years ago, we had a precise vision of the platform. There were no end-to-end solutions available. We then decided to purchase the components we felt were mature, build the other capabilities in-house, and stitch everything together. We created a modular architecture that allows us to adapt as we grow. We take out the old stuff, add the new, and continually adapt it.
David DeLallo: How do you determine which new technologies to bring on and which old ones to keep?
Sameer Gupta: That is one of our most challenging decisions, and we occasionally get it wrong. We initially had a tool we thought could handle most of our data orchestration. It worked great when we were small, but once our data volumes began to grow and our compute levels started to spike, the orchestration tool could not handle the load. We spent a lot of time working with the toolmaker to refine and adapt the software to our needs. But finally we decided to remove it from our architecture altogether because it was creating a lot of scalability issues. That was a tough decision for us. So now we make it a priority that any tool we bring in can grow with us and give us the scalability and stability we need.
David DeLallo: What have been some keys to your success thus far?
Sameer Gupta: Making effective trade-offs has been critical. For example, tech firms preach, “Do metadata for everything.” But of course, getting metadata for everything requires a lot of effort. Our bank may have half a million attributes in one area of our business and three million in another. But we don’t use many of those attributes, so what’s the point of getting metadata for them? We try to approach data collection through a pragmatic lens. We start with the use case and work back from there, asking ourselves questions such as, “What are the attributes that make sense?” We work with our senior leadership team and CEO in making these “must-have” and “nice-to-have” trade-offs. This approach and mindset have really helped us.
David DeLallo: Where do you find your AI talent?
Sameer Gupta: We work closely with universities and hire many of our people there. One example is our Analytics Capability Enhancement (ACE) program, which we launched to build an enterprise-level data science talent pipeline, with a targeted curriculum for data scientists and data analysts. We do a lot of training and upskilling with these hires—and with our own people, too—because the reality is that it’s rare to find the exact skill set your business requires. We found that if people have a basic understanding of data models, then upskilling them is easier. Machine learning is more challenging, since data analysts and traditional data engineers are generally more used to the relational database management world and need to learn how to work with the modern stack we have built into our platform.
David DeLallo: How do you organize your teams? Do you have separate teams for deployment, livemodel operations, and so forth?
Sameer Gupta: We use a federated model, with a center of excellence at the core and cross-functional squads radiating out from there. The consumer banking squad, for instance, will work on intelligent banking, which has hundreds of models. That team consists of product people, marketing people, data scientists, data analysts, technologists, and, in some cases, operational people. It functions as a persistent group with a focus on experimentation and deployment. Experiments that prove successful are then scaled. Data scientists are embedded in every squad and work to improve AI models continuously. Our center of excellence supplements these teams, providing advice and specialist resources such as naturallanguage-processing experts, forecasting experts, and other specialists as needed. It’s a structure that works really well for us.
David DeLallo: How is your organization approaching change management? What are the pain points, and how have you overcome them?
Sameer Gupta: Change management is never done, of course. We work closely with the business, and when we started, there was an expectation that AI would give a correct answer 100 percent of the time and that, with those answers, our cross-sell rate could jump from 1 percent to 95 percent. By working together, we’ve achieved a better shared understanding. Our goal is continuous improvement. We may not get from 1 percent to 95 percent straight away, but jumping from a 1 percent cross-sell rate to a 2 percent cross-sell rate in itself is huge. Change management is also about building trust. You can send the best leads to your relationship managers, but if they don’t act on them, there’s no value. They need to feel comfortable that recommendations generated by the models are accurate. That trust is built over time. Change management is always a work in progress.