How One Bank Democratized Data Across the Enterprise

How do you transform a bank from data silos to 600 business users writing SQL?

In this episode of  ⁨@TheAIVisionaryPodcast⁩ Chief Data Officer Celine Le Cotonnec shares the playbook behind one of the most impressive data transformation stories in banking.

Instead of creating another centralized data team, Celine's team built a single source of truth, empowered business users to become self-service, and transformed data from an IT function into an enterprise capability. The result? More than 600 business users writing SQL, over 70 non-IT employees building Python applications, and a culture where data drives decisions across the organization.

If you're a CEO, Board Member, Chief Data Officer, Chief AI Officer, CIO, CTO, Head of Data, or business leader, this conversation provides practical lessons on building a truly data-driven organisation and preparing your workforce for the AI era.

In this episode you'll discover:

✅ How one bank enabled 600+ business users to write SQL and access trusted data independently.

✅ Why creating a single source of truth transformed decision-making across the enterprise.

✅ The hub-and-spoke operating model that scaled data access without creating bottlenecks.

✅ Why business users—not just data teams—should own analytics and reporting.

✅ How the organization evolved from exchanging Excel spreadsheets to sharing SQL queries.

✅ Why SQL, Python and AI literacy are becoming core business skills.

✅ How AI is reshaping relationship management in private banking.

✅ Why HR should play a central role in AI transformation and agentic organizations.

✅ The relationship between Data, IT and AI—and why collaboration matters more than ever.

✅ How to think about AI governance, guardrails, and measuring ROI from enterprise AI initiatives.

Chapters

00:00 Introduction

00:45 The challenge of multiple versions of the truth

03:00 Building a single source of truth

05:23 Why people don't want to share data

06:23 Why business users—not data teams—should build dashboards

09:40 From Excel to SQL: Creating a data-driven culture

11:36 How to get hundreds of business users learning SQL

12:49 AI's impact on relationship managers

14:08 Why HR is critical to AI transformation

17:25 Designing organizations for the age of AI agents

22:00 Why everyone should learn Python

23:24 Motivating non-technical employees to code

24:44 Data vs IT: Lessons from the transformation journey

27:24 Governing citizen development and AI at scale

39:41 Governance, culture and AI guardrails

42:18 Who should own AI governance?

43:49 Measuring ROI from AI and data transformation

50:43 Rapid Fire Questions


👍 If you enjoyed this conversation, please Like, Subscribe, and Share this episode with your colleagues.

#AI #ArtificialIntelligence #DataStrategy #DataLeadership #ChiefDataOfficer #DataCulture #SQL #Python #BusinessTransformation #DigitalTransformation #AILeadership #DataGovernance #EnterpriseAI #DataAnalytics #TheAIVisionaryPodcast

The AI Visionary Podcast is a platform where AI leaders share bold, unique and tangible insights about pathways to realizing meaningful ROI from AI, pitfalls to avoid and their success stories.

These conversations go beyond the hype to uncover what it takes for organizations to be successful in their AI transformation journeys from a leadership, talent, strategy and operating model perspective. 

Co-hosted by Joel Azariah and Milind, new episodes release every month via You Tube, Spotify, Apple Podcasts and various other Podcasts Channels

Links

YouTube Channel

https://www.youtube.com/@TheAIVisionaryPodcast

Spotify Channel

https://open.spotify.com/show/3Bt27yjIzXP9Yr1JyqzpTv?si=feq6M3RTR1ywJS_694-Q7A&nd=1&dlsi=ce2a21ba58db44b6

LinkedIn Podcast Page
linkedin.com/company/the-ai-visionary-podcast
Joel Azariah LinkedIn Page
...

Transcript

SPEAKER_01 0:00

So if you don't upscale your organization, yeah, you're gonna spend a lot on a useless data lake that we can often call data swarm. Right for everything related to data and AI. This is not a tech issue. You can't have a data team reporting to IT. I think you know, one of the roles that is often forgotten in organization today in this AI transformation is HR. Where is the HR in all this? Because we're talking about a real organization transformation.

SPEAKER_00 0:25

Welcome everyone to this episode of the AI Visionary Podcast. Our guest today sits at the intersection of AI, data, and private wealth in one of the most trust-based industries in the world. She's worked in very, very different industries across government, um, automobile, insurance, uh, and now as a chief data officer of one of the most prestigious uh private banks in Asia. Celine Le Kotonik. Welcome to the AI Visionary Podcast and great to have you on this episode.

SPEAKER_01 1:01

Thank you very much for welcoming me.

SPEAKER_00 1:03

Excellent. So, Celine, you've done some amazing things the last few years in your current role as chief data officer. And one of the things is how you've democratized data in a private bank where relationships are based on trust. Can you walk us through that journey? Like, how did you go about building that department in a private bank where oftentimes we hear things are very or people work in silos? How do we break that?

SPEAKER_01 1:32

I think there is there is the overall um you know origin of why we created this um data team, which was reporting directly to the CEO. I think this was one of the of the right things to do, especially at the beginning. The overall idea was to make data available to everyone because you can see a lot of banking institutions where today um data as siloed as the team, and um and data is an asset from the company. Everybody needs data, right? Uh, whether it's for regulatory reporting, you know, financial analysis, uh, portfolio review with a client. Uh so the first idea was to um create this source of truths. Um so when I joined the bank, there was like a different reporting system. And then uh my boss, the CEO, was like, yeah, we spent you know about half an hour last meeting to um uh to try to understand why uh AUM was 1.2 in finance, 1.6 in risk, you know, 2.8 in sales. And uh and so he threw me the challenge, right? Like, can you make sure that every time we come to a management committee, the numbers are just uh are just telling to each other, right? And um and so I look at um I look at this uh uh you know problems that he um uh threw at me. And the reality is like people were getting data from different systems and different timeliness with different aggregation methodology, right? And uh and I was like, of course you won't get the same number at the end. And um, and when I joined, I also did interview, uh I did interview 66 people, I remember. So I didn't did 66 interview, but I uh I kind of uh checked with Ether that there was about 60 persons that were doing any type of report in the organization, right? That would eventually go to a management committee at one time. And um, so I went to interview a few of them and I was like, what are your pain points? And they're like, we're spending so much time to try to access data, right? Uh so about 80% of the time were to try to access data. And then once they got the data, they were not even sure that those data could be trusted, right? And by the time that they generate their report and their analysis, you know, uh, it could have been it took them about two weeks to be able to generate a report on any type of analytics. And if you wanted to get updated number, then it's another two weeks. So, you know, we're so far from the truth and real-time information that are required for decision making. So we decided with uh Baron, the CEO of the bank at the time, uh, to have a hub and spoke approach. I told him those people know data, those people do analysis all day long. They just don't have the right level of tool. And what about we create this uh hub and spoke model, which is a hybrid model where the hub is in charge of creating and mentoring that source of truth that each and every analyst in the bank can actually tape on and be able to automate their report. So they always have refresh, trusted data that they can rely on and is the one source of truth for the organization. So we govern all the metric. No one is inventing AUM anymore. There's a clear definition of AUM, how it's calculated, and who owns that metric, and we ensure that this is actually maintained. And then after anyone in the organization can come and tape on their data, should you be a risk manager, an edge or uh analyst, a product analyst, marketing specialist, um to take from this single source of truth uh any type of data that you need. So it actually increased a lot the efficiency of the organization. And on top of that, it helped to democratize because went from 66 people, designated uh uh people to try to get access to the data whenever it was required, to more than 600 people now able to do SQL query. And I'm talking about assistant relationship manager, right? Like some people that would have, you would have never thought would be actually be doing uh their own query to get access to their own portfolio information.

SPEAKER_00 5:23

One of the things, just to build on that, Celine, that stops people from sharing data in a bank is losing power and control.

SPEAKER_01 5:31

This is exactly why I did what I did.

SPEAKER_00 5:34

How did you manage? Because in the beginning, like you said, people, there were different people owning different versions. How did you manage to get their trust?

SPEAKER_01 5:41

I took over the ownership of everything.

SPEAKER_00 5:42

Okay.

SPEAKER_01 5:43

So I think that was the, you know, that is uh the reality. I took over the ownership of all of the reporting system, yeah. Uh, the legacy one, the newer one, um, whether they were, you know, in finance, in risk, or or anywhere else. And uh, I took the ownership of everything and then I modernized everything. So we broke, you know, we did a commission eight system in the process by creating this single source of truth. Uh, it was only by taking the ownership that you can actually make the change.

SPEAKER_00 6:07

So for assistant relationship manager, their their main interest is their client. They're not so maybe uh you can walk us through the story of how you help them to understand, you know, why it's important to write an SQL query or understand what's in what's in it for them.

SPEAKER_01 6:22

You know, it it was the same thing as implementing this have and spoke approach, right? A lot of people were like, but you're the data team, you're gonna be the one creating dashboard for me. And I'm like, no, you're gonna create dashboard for yourself, right? Because A, I'm not in your head, right? I don't know what you're looking for. Everybody needs dashboard for different purpose, right? Um, the time that you're gonna spend writing me, writing me business requirements in Word documents, knowing that those requirements would change, uh, you know, it's it's just actually a waste of time, right? You better do it yourself. And let's not forget that analytic is an iterative process, right? Uh so you go and deep dive into the data, you want to look at a trend, you identify an anomaly, you want to dig deeper into that anomaly. You need to have the ability to be able to um interact with the data looking for answer by yourself. That's how you can take data-driven decisions. You should not outsource that to another team. And I didn't want my team to become a bottleneck. I don't think the data team should become a bottleneck to the organization. We're not a second IT team, right? Uh, we're actually here to empower people to be able to use and leverage data for their own decision making, right? Not take decision or tell people, you know, what is the trend on their own business. If if your business is to, I don't know, manage risk, then you should have the ability to do it by yourself without waiting for someone. And I think this is where the democratization did work because I told the people, well, you got two choice, you got you got two choices, right? One, you can um, you know, uh pick up a ticket and and and take the queue, or you can learn how to do it by yourself and you you can be self-serve anytime that you need data. And that's how we also reach this level of democratization of SQL query in the organization.

SPEAKER_00 8:09

That's a big step from doing for, like building for versus building with.

SPEAKER_01 8:14

Yes.

SPEAKER_00 8:15

So that's like, is it like a mindset shift that needs to happen to make that a reality?

SPEAKER_01 8:20

It is. People need to understand that upskinning is not an option anymore. I think more and more people realize it, especially with like AI, you know, but the data literacy and AI literacy, it's, you know, I would say it like um, you know, when computer uh arrive in an organization, right? Like um, everybody had to learn Exile, everybody had to learn PowerPoint, right? Everybody had to learn how to use um Outlook. Um I remember a joke from one of my ex-colleagues in another company, which was telling me her her boss was uh you know uh quite aged, and she was like, I'm printing all the emails so he can read them and then answer, and then I type. I'm like, okay, you know, like who would think about this today, right? Like you write your own email. You know, in the past you used to have assistants to write your letters, uh, but you know, things do change. And it's the same for data and AI today, right? There's a profoot profusion of data, and with data, you can actually drive your business. You need data to be able to, you know, um design your strategy. Uh, and all of this cannot be outsourced to the team that is not doing whatever you're doing.

SPEAKER_04 9:28

There is one point that I'd um uh I'm curious about. Usually uh the data ownership lies with the uh business team, the team that is producing the data.

SPEAKER_03 9:39

Yeah.

SPEAKER_04 9:40

And one of the uh examples that you gave, AUM had different uh yeah, uh different numbers because there were multiple different definitions. And those definitions are usually owned by those business departments, correct. And aligning them is the biggest challenge.

SPEAKER_01 9:57

So you can't have an alignment for everything. So now we have a risk AUM, uh finance AUM. You know, sometimes metrics are different for diff for good for a good purpose and for a good reason. The same as you can't have one dashboard that does everything, right? Because people are looking at things differently and for a very good reason. They are looking at different trends uh depending on you know what is what is their interest. The only thing is don't call AUM uh different numbers that have different aggregation methodology, right? So let's be clear about which one is including loans, which one is including including, you know, um uh insurance uh premiums. Um otherwise, of course, there are going to be discrepancy in um in the numbers. And one of the things that um I realized I was particularly proud of um after fighting for this hub and spoke model and people not understanding that they needed to take ownership of their data and their learning journey uh was when I was um, you know, in the past, you would see different departments like finance and risk sending each other um Excel files regarding why are the numbers different. And then at a point in time, the organization started switching by a finance analyst and then a risk analyst sending each other SQL queries, right? And then I forward that to my boss and all, right? This is what a data-driven organization looks like, right? We're sending code to each other, and so very quickly we can see, oh, you did not include loan into you know your calculation. Um, whether when you send Excel file and you have you know the the the the data itself, you can't really understand what is the aggregation methodology behind. So yeah.

SPEAKER_04 11:35

And this upskilling part, how challenging was this uh uh to get people to learn and start using SQL?

SPEAKER_01 11:43

You know, when when when you actually remove the pain point, a lot of people have that pain point. I don't have access to the data. Uh when you start um uh relieving that pain point by telling them you can have access anytime you want, you don't need to ask anyone. Everything is fully, you know, access control managed, so you can only have access to what you're supposed to see. Uh, but you're relieving them from a big pain point, which is waiting for the answer that they need. Um, so they're usually quite keen on it, to be honest, because again, it helped them to stop waiting and having to resend an email and an email, you know, and follow-up just to get access to information that they should have. It's part of you know their job.

SPEAKER_00 12:25

Now that Celine, you've raised the bar quite high, right? Everybody has now a new expectation and a new uh feeling that okay, we we've already achieved this, so we can achieve something much bigger. But in private banking, most people say, okay, uh there it's the key to the relationship between the client and the relationship manager's trust. So where does AI and and data come into that to you know enhance that relationship?

SPEAKER_01 12:49

I think this is where the this is what we're trying to focus on today, right? In most of our use cases, how can we support our front office and relationship manager uh to serve their client better? Um so having access to the most accurate information um is uh is is uh is one of the, of course, idea, right? Recommendation, next best conversation, what kind of product can you propose? How do we ensure as well that the sales cycle embeds all of the risk control and compliance uh component that are required for private bank? Uh, but more and more, what is challenging for the relationship manager themselves is that all of that client are coming to the meeting with their own ChatGPT app, right? And then challenging what the relationship manager or the advisor uh is actually telling them on how do they need to rebalance their portfolio, right? So, how do we give them the right tool to be able to answer to that client with the same level of information, knowing that clients have, you know, uh web-grounded information, right? So real-time information from the markets.

SPEAKER_00 13:52

So the game has changed a lot. Yeah. The client also has a lot of information.

SPEAKER_01 13:55

I mean, it's the same for the teacher when you think about it, right? The teacher are struggling because you know the teacher kids something and everybody goes with their you know GPT app or other and then challenging the teachers, but whatever they're learning.

SPEAKER_04 14:08

In fact, that was going to be my next uh question. Now that uh both customers and our relationship managers are uh yeah uh completely AI tooled up, what requirements have changed from uh from the humans involved in this loop? Uh our relationship managers or people supporting the relationship managers? How apart from upscaling on the SQL query side, what is what additional um trainings or upskilling or uh or other requirements from uh from people that we hire for these roles?

SPEAKER_01 14:42

Specifically for the front office, is what you're asking. Yeah. I mean, it's you know, more and more you will have prompt engineering, but I don't think it's specifically for the front office. To be honest, it's for everyone in the organization. And um, and I think, you know, one of the roles that is often forgotten in organizations today in this AI transformation is HR. Because the reality, you know, in a lot of um AI governments committee, you would see like data people, tech people, security risk, right? And where is the HR in all this? Because we're talking about a real organization transformation. And when you're thinking about your agents, right, this is also an organization design skill set, right? We're talking about skills, right, for agents as well. Who is designing those skills? Today it's more on the tech side, right? Um, or the data scientist side eventually, right? But the reality is like, what is the role of HR becoming today? How is HR managing um, you know, the role of agent via real human, right? And what is the different type of workload? And and there's a lot of um economical question to it as well. I mean, we we're hearing more and more that it's not only the salary that would matter tomorrow, it's the salary plus your token budget, right?

SPEAKER_04 15:56

Yeah, for sure. But other than that, yeah, uh just taking two examples, uh, one from the vaccine company Moderna, uh, who merged their HR and IT departments last year, uh, around middle of last year. And uh around February, McKinsey said that uh they have 60,000 employees, yep. Uh what around 40,000 of them are AI agents, uh some uh such number. And this was already back in February. Uh in the uh age of AI and the speed of AI, this is uh AI agents have become uh far more capable. Uh to give you two uh examples recent examples with uh let's say Microsoft uh announcing in the build uh event uh and always on open claw-based agent uh which will have its own intra-ID as an uh independent identity, and it is basically AI is entering a uh well and truly into the employee layer of the organization. So how um uh exactly as you said, HR we cannot forget anymore. Yes, and this is uh this has, as you again correctly uh observed, uh, has uh huge implications for the organizational design.

SPEAKER_01 17:15

Yeah.

SPEAKER_04 17:15

Yes. Uh how are you thinking about that, considering that this is now the frontier of where AI is and you've already put uh all the foundation in place?

SPEAKER_01 17:25

I mean, I I think it's a journey, right? But we we're we're we're turning toward an organization where silo has to be broken, right? Like you can't replicate, and it's the same with your Agentic AI framework, right? Like you can't replicate siloed agents, um, uh the same way as you would. And and this has always been the story, even in IT, right? Like uh you have a team that is creating their own system, uh, don't want this to be shared, you know, with another team because different purpose, everybody thinks that they have different purpose. Um and this is why the agentic organization of the organization needs to be thought through far before the agent are actually created, right? A lot of time we just go, okay, let's go, let's create agents, let's experiment. And then after we realize, oh, actually we need to harmonize and unify because you know, we have a couple of things that are doing pretty much the same, you know, similar um activity. And and now with the costs of running agent and model, you know, again, those token costs, we were not talking about them in the past, right? But now it's something that every every organization is starting to look at, right? Is what is my token cost? Stop using AI for anything and everything, right? So we we're going to the extreme, and I think it's gonna um we're gonna rebalance uh the role between uh AI and and and human and and try to actually um uh be clearer about what AI can do and is supposed to do, and then what task uh should be actually met by human, because you know it's also very costly to run AI for each and every activity of your organization.

SPEAKER_00 19:00

And that's why I think Celine, the point that you made before about HR being involved in this transformation as a one of the key players is so important because oftentimes organizations see it as a technology problem, exactly not a human-related problem. And so HR says, okay, it's not exactly in our area, so we'll take a step back. But what you said is right, I think why most companies struggle is because they view it very differently and they tend to take it more as a technology problem rather than HR.

SPEAKER_01 19:31

And this is the problem, right? For everything related to data and AI. This is not a tech issue, right? And I keep I keep talking about that because you know, a lot of time I see data projects and I see the SDLC lifecycle. Like we are not developing software, right? Like there is no SIT UAT concept when you develop a model. We need production data, we need real data, and we need to have the capability to deploy much more quickly. And that's one of the problems also uh in organization where data teams do report to IT. I think this is really an organizational issue. You can't have a data team reporting to IT. A data team is here to be able to generate insight and solve business problems. Yes, you have coding requirements. Uh, and I think the the and you have a deployment process, you are using tools that are similar to what your software uh developer would use, but used in a very different way, which is to solve business problems and to bring instant insights, right? Uh and um and so we need to have a certain level of flexibility. And even we're talking about potentially, you know, um, why do we need UI anymore, right? Like now with um uh MCP being launched, and I don't know, look at a bank. Uh 60% of the time people are spending uh their time on a Microsoft-based environment, right? Everybody's on Outlook, you're doing your Team School, you're preparing your PowerPoint, and the rest of the time you're on Excel, right? Pretty much. And then you have your trading system, uh, front-end system. Uh, but now with um data and uh and model all being embedded, right, within your uh Microsoft environment, um then uh yeah, everybody is able to create agent now, right? And and then uh we we're hearing about this one people, uh one-person company. Uh everyone with some you know uh vibe coding and and ability to prompt properly is able to create their own app within uh within a day or two. I I do believe that the barrier uh between business and IT is getting blurrier and blurrier. And and this is what we did in Bank of Singapore. Um we spend a lot of time creating that foundational source of truth. One of the strengths as well of this unified data model is that it's taking from our financial approved number. You will see a lot of organizations where data team produces a number, and then when you want the real number, you go and check with finance, right? Because those are the ones that are actually holding the PN. And the board reporting. So we wanted to make sure that everything matched. So we took all of the financial months end STEM data as the basis of our data model. And then we layered on top of it, you know, employee dimension, uh, transaction dimension product, you know, uh risk compliance. Uh, so we have a proper uh one source of truth. And now everyone is able to build their own application. Um, so we're promoting Python within the organization. This is the language of the business uh in the bank. Uh, and we have more than 50 people that are committing code today, which are not in IT.

SPEAKER_03 22:37

Oh, wow.

SPEAKER_01 22:38

So we have more than 70 people committing code and developing their own web application in Python. Uh, and those are the people that have the ability to kind of understand how how do you code, so how do you prompt as well, right? Uh, so it's also a learning journey. You know, I was told why do you want people to learn Python? And I'm like, well, because it's it's more for learning purpose um than uh on how do you build a proper application? Uh, what is the the quality of the code that that should be expected? Uh, because the better you learn how to code, the better you're going to be able to prompt tomorrow.

SPEAKER_00 23:14

And how do you incentivize that? Because for somebody who's not a natural or doesn't have that background of coding in Python, how do you incentivize them? Okay, this is something that you can do.

SPEAKER_01 23:24

And uh, I mean, it really depends on the people. Again, the incentive is that to get your pain point resolved, right? Like either you wait, either you keep your manual process, either you develop your own Python application. I I guess also um the rest of the business, especially the management, have understood how important it was to get those skill set, SQL, Python. So now it's kind of the basic uh requirement for any job in any um, you know, business line. So the young generation are the ones that are helping to transform the bank because they arrived, they got those skill sets, you know. Um Python is a mandatory uh uh requirement in a lot of high school today, um, you know, even in Singapore. And so everybody has basic knowledge. And then through training and hands-on learning, uh, you know, it's it's it's easy to um uh so it's it's also coming from the new generation and the new recruit.

SPEAKER_04 24:19

So we've heard a lot of your successes, yes. Uh, and uh what I'm hearing is just uh totally amazing uh about the transformation. Tell us some about some of the roadblocks uh yeah or speed breakers that you came across this journey and maybe uh something that uh our uh viewers and listeners can keep in mind as a learning that okay, this these are some of the mistakes to definitely avoid.

SPEAKER_01 24:43

I don't think we did any mistake about that. No, I think the challenge will always be you know the relationship between data and IT, especially in bank, right? Because you got this big like Chinese wall and all of the regulation that it used to technology risk management. And um, and there will always be um, you know, challenges uh because it looks like we're doing the same thing, but we're not doing the same thing. But we have you know some common skill sets, and so collaboration between both teams is critical. But the way that we operate is very different. We operate at different speed, right? Like, uh, and for very good reason as well, uh, because you know, uh technology department are handling a critical trading system, booking systems, you know, we're handling dashboard inside, you know, like we have a level of accuracy that is not the same as what you would expect for um uh, of course, trade reconciliation and booking and portfolio management. Um so I I don't know. I I to be honest, I don't have an answer, but how can you ensure that there is more collaboration between uh you know the data uh driven people in the organization and then the IT folks who have the knowledge, right? And when you put a data guy and an IT developer, a data engineer, you know, a software engineer developer talking together, they understand each other. But you know, I believe the organization again, the siloed and the perception that, you know, one team is supposed to do something like deployment should all only be done by IT, right? This is one of the rules um that is uh often seen in financial services, an organization. The reality is like today, everybody should have the capability to deploy um in a certain framework, right? Uh, and so I think uh um yeah, I don't know if it's a failure, but it's definitely a learning, right? Um as part of this journey.

SPEAKER_04 26:36

That point. Um, in fact, uh just to support the point you made, uh, it's not just banks. Uh uh I think uh one month back uh or uh two months back, there was an article in The Economist uh written by Ethan Malik, uh who is a professor at Wharton. And the title of the article was uh the IT department where AI goes to die. Exactly. Uh because the remits are different and speeds are different.

SPEAKER_02 27:04

Yeah.

SPEAKER_04 27:05

Having said that, yeah, uh you uh correctly pointed out now that everybody is uh able to write code, deploy code, build their own things.

SPEAKER_02 27:16

Yeah.

SPEAKER_04 27:16

How do you decide what goes into production, what is scaled up, and how do you contain the sprawl?

SPEAKER_01 27:24

We don't decide, we see things that and how are they taking up? I mean, again, the the definition of in production, yeah, I think it's it's outdated as well, right? Like the reality is when you do data, you're always on the production segment.

SPEAKER_02 27:37

Right.

SPEAKER_01 27:38

Right?

SPEAKER_02 27:38

True.

SPEAKER_01 27:39

So we have this concept of pre-production or also aka sandbox, and then you know, deployed in a running environment where you have like monitoring, quality check, you know, and a certain SLA, but we're all in production. You know what I mean? Like there is no in production. So when I hear this about data, I'm like, guys, we're not doing it's not a software development lifecycle, right? There is no SIT UAT prod that does not exist anymore in our world. So the definition of in production doesn't really make sense for me uh in that context. You don't decide, you see the adoption rates of whatever you have built. And all of the things that we have built in Python, um, you know, I like to think it's also a game changer because it is the user uh defining through code what are the requirements for a functionality that we're gonna implement into another target system. So we're working very closely with our architecture team. Our Python application, uh, the business requirement in code that needs to be translated in React.js or JavaScript into another target application. Um so rather than having this overall cycle of let me write down my requirement in Word document, and then you know the requirement arrives to a developer that developed whatever he understood, right? Because there is mystic in words, right? Now you have a Python script that has been um iterated on what the process have been you have been able to improve because through our Python application we're also collecting the data. So then you can see uh some of the bottleneck in your current process. And there's a lot of process redesign in whatever we do, right? It's not only about data, it's about thinking about how um teams collaborate with each other and what are the bottleneck. And you know, for example, we're doing um we're um doing a tracker uh for a process in the organization and we realize, oh, there's too many checkers and not enough makers, right? So when the boss saw that, he's like, okay, we're gonna revamp the resources and make sure that, you know, like the checker are not waiting for the maker, which don't have enough resources to do their job, right? So um you can't improve what you don't track. So start by tracking, um automating uh this process, and then after developing this functionality up to the point of stabilization, right? When you see that you don't make any more, you know, um uh increment or improvement on your code, then it means that it's mature to be handed over uh to an IT team that would develop this functionality as a new functionality into a target platform.

SPEAKER_00 30:06

And I think Celine, what you mentioned earlier is because you've democratized the whole process, there's nobody owning the production and saying, okay, these folks, we need to go to them to put things into production. And then that's the whole point you made earlier about not being the bottleneck.

SPEAKER_01 30:21

Exactly. I mean, we are the one deploying, yes, uh, but we don't have the same level of, let's be honest, right, rigor that what IT would have, because it's a single platform, it's content, it's only internal data. It doesn't go to client. There's nothing that we do that goes to client. It's only for pure international internal efficiency and collaboration. So automating, um, you know, a lot of the people in in bank, um, you know, we talk we talk a lot about this shadow IT or EUC end user computing. Uh the reality is like if this exists, it's because there is a need. So let's stop uh being blind about it and let's try to empower the people that are developing uh end-user computing into their own laptop, you know, which is a very dangerous way of operating, uh, and then get them a proper platform with proper resources, with proper, you know, GPU, model access, API access to external data. Um, so all of those knowledge can be shared, you know, and people would stop, you know, redeveloping the same code and the same script within the each and every computer, right? So yeah, it's more of a collaboration platform. And yeah, you can call this Shadow IT, right? But the reality is like it is a proper platform with, you know, the same um IT um uh tools, right? We have a Bitbucket repository, we have a single sign-on integrated, we have you know uh proper uh data entitlement, uh, the same as any other IT system, but it's managed by the business and we deploy three times a day.

SPEAKER_00 31:52

Excellent. Did you have to, Celine, put any specific governance frameworks and you know, guardrails in place to ensure that this happens in a way that's you know, everyone's it meets the standards and policies of the bank. Because I'm sure like once you go out and you share with your business that this is something you can do as well, you need to put in place certain structures and frameworks to make that possible.

SPEAKER_01 32:17

No, I mean, the the main thing that we did, to be honest, right, uh, on the governance of this data model is just to mask any PI because when you're doing analytics, you don't need the name, you don't need the mail address. Of course, there are things that you need, such as age, residency, uh, nationality uh for reporting purpose. Um, you need we had a long debate about um portfolio number, is it a PI or not? I'm like, no, I give you a portfolio number, you don't know who's behind, right? So um there was a whole we had to uh align all of the control function of the definition of PI and not only the definition in a policy, but really um at the data element level, is this a PII, yes or no? And uh we just mask everything uh that is you know a client ID, uh any client identifier. Um and then after, yeah, democratize uh the access to the information and then we let people come with their own uh idea on how to solve their own uh problem. And after we got you know um similar um deployment process than um any IT, uh I mean Jira for requests, uh Bitbucket for code repository and versioning, uh we implemented single sign-on. Um so we as the data hub take care of implementing all those frameworks, but they're automated framework, right? Okay, um there is no policy or form that you need to sign or you know, everything has been automated, regression testing.

SPEAKER_04 33:42

Um whenever people think of uh banking, the first thing they say, oh, it's a very regulated industry. And yeah, it is this is possible, that is not possible. And maybe change will always be slow. However, uh from what I'm hearing, yeah, you've done a fantastic job. Which parts? Uh so just to uh uh bring focus and complete understanding of uh when we say banking is definitely very regulated, but from a data and AI perspective, what are the parts that are affected by regulations and what are the other parts uh where we you uh we have much more freedom to uh implement AI? So one part you already answered.

SPEAKER_01 34:26

You get any freedom. It's highly regulated, as you just say. Um that being said, so there is no freedom, there is no part that have more freedom. I mean, of course, internal efficiency, right? So the use of copilot and stuff, but after all, there's always a uh a question of the sensitivity of the data that you can put in those models, right? Um the use of cloud has always been a long time debate. Now I think every bank realized that cloud is not a discussion anymore, right? Like you need cloud if you want to be able to be up to speed to the frontier model, and you know, you can't keep buying your GPUs and it's just too long. That being said, I think banks are quite in the forefront of being able to regulate AI much more than any other industry. Um, it comes from the fact that banks have been, you know, all of their business model is built on model, right? So there used to be GLM model, uh, different from uh uh the current AI model, you know, the type of data that we use is different, but they already have framework in place in order to uh validate uh the materiality of a model. Um they have proper committee, they're already proper policy, uh, proper understanding, ownership compared to probably a lot of other industries. So thanks to also uh BCPS239, which is the Basel Tree Regulation, uh looking at the data quality, uh the lineage of all of your data. Um I think banks have good foundation. Now the question is how do you transform this um governance from a policy and a committee base uh to a code-based approach? Because the the truth is in the code, right? Your controls need to be part of your CICD deployment pipeline. It can't be left to human, right? Human are here to give the you know the guiding principle, but not to come and review each and every model to see if they are fit for purpose or fit for assessment, or it's it just you know, it's not it's not covering the speed that AI needs to be able to be fully implemented.

SPEAKER_04 36:29

So govern uh regulations and governance has been not a speedbreaker, uh any kind of blocker.

SPEAKER_01 36:38

I mean, it it you know, this is bank in general. So of course, every time that it's regulation, every time that it comes to to uh and and and for a very good reason, right? I mean, you don't you you don't want you don't want your your your bank to be managed the same way as your Netflix, right? Like we were talking about you know your retirement, we're talking about your savings, we're talking about things that do matter to people. So in terms of security, the standards are quite high, and for a very good reason. Uh integrity of the data as well. Uh, you don't want your number in your bank account to change from one day to another. And yeah, double your network within overnight, but you would be happy, but you know, it would mean that we have done something very wrong. So, yeah, the reality is like things are slow, but for a good reason. Uh, but when it comes to internal efficiency, the use of AI for internal efficiency, banks have been quite at the forefront already for quite some time. I mean, you know, we are the first one concerned by fraud management, for example. So anomaly detection and you know, being part of the OCBC group, which were also had a quite advanced data science team, uh, is also something beneficial because we we work with them quite closely, whether it's in terms of platform or whether it's in terms of use cases. Um, bank and financial services in general have always been looking at predicting the future, right? This is what your portfolio manager is trying to do, your equity analyst, right? So the use of data has always been here. Now the question is how do you um govern those data and modernize the way that we work? And um this is why Shadow IT in the past used to be VBA-based macro on your computer. Uh, let's move everything to Python with a proper UI and access control. Uh, and and realize that, yeah, anyone in the organization is able to um leverage on data. Uh now my main concern is about the second line of defense, right? So you get this concept in financial services of first line of defense, second line of defense, third line of defense. Um, the reality is like you can see that the first line are going fast. The third line, even the audit function now is being transformedly transformed by the use of AI, the access to data and the ability to actually identify any anomaly uh uh coming from the data. Now the question is how do we transform our line to, right? How do we ensure that the policy are not just um the policies on the committee are not only human-based, but they're actually translated into code that can uh monitor on real time. How can those uh line two function be monitoring function rather than approving function?

SPEAKER_00 39:18

And how important is culture? Because obviously you spoke about the democratization part, but also in the governance part. If it's already part of the culture, then it makes a big difference, I'm assuming. If they the culture of governance already exists, that people feel that yes, we understand it. It's a it's a rule, but we also understand the importance of it when we're developing these products and solutions.

SPEAKER_01 39:40

Yeah, I mean, again, right? One of the benefits of being in an already highly regulated industry is that uh governance, risk, compliance are part of everyone's day-to-day responsibility. Uh, it's part of the KPI from the front office to any any any first line um employee. Uh, so it's always in the back of the mind of everybody, right? Like we got all of those compulsory training that everybody has to go through on a yearly basis, um, which is mandated by the regulator. And and it's something good because then we, you know, one of the of the issues and of the fear that I have with the upcoming, you know, wave of let's go everything with AI right now is the lack of governance, right? We're not we're not stopping to think about, okay, what are the risks, right? So everybody is just diving into the adoption of all those frontier models and then building iagent, but no one is thinking about what could go wrong. And how do we need to start thinking right now about what could go wrong and what are the guardrails? We talk about guardrail, but what are they? If you don't uh spend time to sit down and define those clearly and implement them into your deployment pipeline, um yeah, this is this is where things could get out of hand, I guess.

SPEAKER_00 40:56

I think one of the reasons for that is a lot of boards and CEOs are incentivized to drive AI in whichever form or way that drives their share price.

SPEAKER_01 41:06

Yeah. Yeah, I think it's a trend, right? But that the reality is like everybody's running, um, but sometimes we need to stop and look at do we even have the right found. Uh and uh everybody wants to work on the shiny stuff. Nobody wants to work on the, you know, a real foundational work of metadata documentation, you know, the data about my data, uh, and uh the access control. We're talking about access control. You can't replicate access control that I use for human to an agent, right? Like things can go very wrong. You need to have a level a level of access control to a granularity that is much lower than what you are giving to human, right? And uh otherwise this could create potential risk. Um, so there's a lot of foundational guardrail that needs to be set right now. And a lot of time, you know, a lot of the uh consulting company will come and see you and say, Oh, you're gonna make 40% efficiency with AI. Yeah, but who is costing the cost of implementing compliance for those um AI energetic solutions that are being set up in organization?

SPEAKER_00 42:08

And where does that responsibility lie, you think, Celine? Because a lot of people say, okay, it's the chief data chief AI officer, or is it the CEO of the board? Or it's a combination of both.

SPEAKER_01 42:18

I mean, at the end, it's at the board level, right? Like it's an organization, you can't have a single person. I mean, and this is why one of the roles of the of the chief data officer, chief AI officer, is is um is the setup of those committees, right? So uh my role as chief data officer is to align our management on the strategy, bring the issue, propose some solution. Uh, but we have a committee. I can't be the one taking decisions for all of the data of the organization because you know I don't know all of the data, to be honest, uh, and I'm not the one using them on a day-to-day basis. So this needs to be a collaborative approach about how do you govern um the data in your organization. So where the CDO sits is always a big question. Uh, my view is um wherever you want, as long as it's not in IT. And you know, I did report to finance, I'm now reporting to uh the CEO. Um it doesn't really matter. I could report to risk tomorrow, even to HRY not. Um, but uh the reality is that uh uh making sure um that you have a unanimous uh governance of the data set of the organization is is is critical in any organization today.

SPEAKER_04 43:32

So we've been talking about the transformation journey till now. And we uh hear this particular statistic quoted everywhere that uh MIT says 95% of all AI initiatives don't drive or deliver any ROI.

unknown 43:48

Yes.

SPEAKER_04 43:48

So with all of these efforts over the last few years that you have uh achieved this uh this amazing transformation, how is it that you uh measure the ROI? From all of these AI initiatives that have been realized. How do what are the metrics that we uh we have to keep in mind and uh or that you have learned to uh keep in mind and track?

SPEAKER_01 44:11

I mean, the reality is like it always comes down to um the number of hour saves, right? How did you do something before? And then how quicker will you be able to do it tomorrow, thanks to agent? But it's not only AI, right? There's a lot of process reengineering behind, right? Of course. It's uh it's uh it's a constant um question about uh how can it's it's continuous improvement. How can I do better? A task that I used to do yesterday in a different manner today. And um so yes, there might be a model in the loop, right? Uh and uh it's very difficult to quantify insight, right? It's the same thing, right? For dashboard, how do you value dashboard, right? Like I was blind before, now I can see how do you value that? What is the metric that you're putting? Uh so you could check, okay, this dashboard is highly used, right? It means that it's useful. So, you know, we kind of look at the value framework for data. I mean, um, you know, the table that we created, for example, how many times have they queried every day, right? And then I put an API call number, right? Like I say, okay, the same as how you are paying for external data today through API call, then I would say, okay, this table, you know, this query querying this specific table, this table is actually an important data asset because it's being queried by everyone every day, right? So what value does it have for the organization? That's also how you're able to value your data engineering work, right? A lot of people don't really appreciate enough the job that is done on the data engineering side of things. You know, they can see the dashboard, the data scientists, you know, when there is no UI, you know, people don't understand and they don't understand the value that it has. Um so yeah, I mean, you have different ways. Um, to be honest, as long as we're bringing value to the organization. This is why I think it's also important that it's look, it's a it's a function that is supported, um, located in the business and financially supported by the business. So you don't get into those uh details of, you know, how much how much do I invoice you back for the service that have rendered you, or how much do I um uh value, you know, how much do I charge back? Um uh the same as what IT department needs to do.

SPEAKER_00 46:22

I think too many times companies talk about ROI without focusing on the right data strategy, the right data foundation.

SPEAKER_01 46:29

And that's the thing, right? Like we're talking about ROI, but the thing is like you need to invest a lot for uh your data quality to be, you know, data accessibility again, right? Like what I've just told you. We have more than 20% of the employee that can access data anytime. And um, and and this is valuable. How do you put a number to that, right? Uh I mean, I could take all of the salary and tell you that, you know, all those people are uh um are being more efficient by I I don't know what percentage, because now they have access to data, whether before they were struggling to have access to data, right? Um the question is do you spend your time measuring or do you spend your time doing? And the reality is like I prefer to do to do change and transformation rather than um, but um one of the objectives, if I if I can give it coming from a procurement background, one of the of the objectives that we um align with the team is that we need to bring four times more than what we cost. So that's our target, right? So I take the cost of the team and I'm like, okay, guys, this is how much we cost, this is how much we need to bring back, right? So yeah, but we don't, apart from uh time saved, um, the value that you can put on insights, the value that you can put on a whole business transformation and engineering uh is hard to track. And when you know, when you're doing machine learning and recommendation, you're doing A-B testing and it's okay. But I mean private wealth, right? Like I don't sell credit card, right? So it's it's very difficult. I'm not selling standard product to mass customer. We're uh designing a portfolio uh for one client, right? How do you value that, right? BF, uh VS when there was no portfolio, right? Before you know what I mean. So exactly um it's it's always a big question. Uh, but for me, uh, if you are asking the value that it brought, it means that you don't see the value that it brought by your data team. So yeah, um, so far.

SPEAKER_04 48:20

Your answer is very interesting because uh, and uh this is uh part of the reason why we asked this question. Because uh in a lot of AI and data discussions, this statistic or similar statistics get quoted. We have we spent so much money, we did so many um But they spend money on what?

SPEAKER_01 48:38

On the IT infrastructure or on the upskilling the people? That's the question, right?

SPEAKER_04 48:42

Good point.

SPEAKER_01 48:43

Uh you spend money on building a Hadoop data like uh whatever that costs you million and that no one has access to. What's the point? Right?

SPEAKER_00 48:49

Yeah, I think they deploy the technology before building the business use case with the business.

SPEAKER_01 48:54

Before before granting access to the people, right? So if you don't upskill your organization, yeah, you're gonna spend a lot on a useless data lake that we can often call data swamp because you know no one has access to it, right?

SPEAKER_04 49:06

Call them data puddles. Yeah, exactly.

SPEAKER_01 49:08

You just end up like pouring and pouring data, and you're like, okay, what's the value? Well, no one has access, you know. Where do you get the value from if it's not from your business um improvement?

SPEAKER_04 49:17

This uh question was one of the main drivers why Joel and I yeah started this podcast. We got uh we used to get uh asked this so many times and said we said we invite the business leaders who are actually making the change and ask them, yeah, and get uh because you can give any answer to this, but uh coming from people who have actually made the transformation, it's extremely valuable.

SPEAKER_01 49:40

Yeah. I think it's you know, if the rest of your organization don't see the value that you're bringing and it's starting to ask, right? And and I got this question a lot. I mean, and I I like there's um another CDO uh fellow here in Singapore who says, you know, um, you can uh torture the number to make them tell whatever you want them to tell, right? So um we know how much uh some people are very creative in in doing business case, telling you that, okay, it has been one billion of benefits, you know. I don't know how it's calculated and what is the again, uh, the veracity of this metric who attempt that this is actually bringing you know a million efficiency or a billion. Uh the reality is like, yeah, if people see the value, they don't ask how much it is because they know that there is value and we don't want to spend time on tracking what we do, we prefer to solve business problems.

SPEAKER_00 50:34

Excellent. So on that note, we go to the next stage of our episode, the rapid firearm.

SPEAKER_01 50:40

Really? Wow.

SPEAKER_04 50:42

2035. Yeah, who will own the customer? Algorithm or relationship manager in private banking.

SPEAKER_01 50:51

Still relationship manager.

SPEAKER_00 50:52

Who do you trust more with your data? Your smartphone or your bank?

SPEAKER_01 50:56

My smartphone.

SPEAKER_04 50:58

Regulations. Singapore, US or EU?

SPEAKER_01 51:02

Uh Singapore. It's a good place in between.

SPEAKER_00 51:07

What's the biggest lie the AI industry is telling itself right now?

SPEAKER_01 51:11

That everything will go to agents.

SPEAKER_04 51:14

One person company. Should AI be allowed to manage money autonomously?

SPEAKER_00 51:20

No. Next one. If you could pass one AI law tomorrow, what would it be?

SPEAKER_01 51:26

I don't know. It should be to make sure that AI is not impacting too much the human race, right? Like, what is the impact on the society?

SPEAKER_04 51:37

That brings us to uh to the end of our rapid fire section. Thanks, Celine, for joining us. It has been a pleasure. Uh the conversation was scintillating and extremely uh insightful.

SPEAKER_00 51:48

Fantastic. Thank you so much, Celine. Excellent. Thank you for everyone for joining this episode of the AI Visionary Podcast. Please remember to like and subscribe our YouTube, Spotify, and Apple podcast channel. Thank you.