# The Foundation First - ING's Global CIO on Why AI Pilots Fail to Scale

- **Show:** CIO Leadership Live: ASEAN
- **Published:** 16 Jul 2026
- **Source:** https://chrt.fm/track/3B598B/traffic.megaphone.fm/FCI8486216415.mp3
- **Transcript:** https://podcasts.opennash.com/shows/cio-leadership-live-asean/the-foundation-first-ings-global-cio-on-why-ai-pilots-fail-to-scale-db8d47/
- **Provider:** OpenNash Podcasts (machine-generated)

## In brief

ING's Global CIO of Wholesale Banking argues that AI succeeds at scale only when the boring groundwork is done first. A decade spent modernising core banking systems, standardising products and building shared data platforms is what lets a pilot that works in one country be copied to others. Without that base, he says, pilots stall.

The central tension is that generative AI is easy to pilot and very hard to scale. ING's answer is to test whether a use case can scale before investing in it, to follow the strictest regulators, and to keep humans involved for large or risky decisions. The guest is open that agentic AI is still at an early stage.

## Key insights

1. **Generative AI pilots are easy; scaling is hard** Hoving says a pilot is simple to launch but very hard to roll out widely. ING now asks whether a use case can scale before it invests in the proof of concept.
2. **A decade of core-bank modernisation enabled scaling** ING replaced core banking systems that were around 35 years old and standardised its products and data. A pilot that works in one market on the new platforms can then be copied to others, such as chat support moving from the Netherlands to Belgium and Spain.
3. **ING follows the strictest regulators as its baseline** Rather than building to each country's rules, ING builds to the high bar set by regulators such as Germany's BaFin and Singapore's monetary authority. Hoving says meeting those standards lets it comply across its whole footprint.
4. **Humans stay in loans, with AI doing the prep** Machine learning decides small, low-risk loans, while analysts review large or risky ones. For those, AI agents or co-pilots extract data from annual reports and pre-fill the approval flow, and agents must escalate to a person when they go outside their guardrails.
5. **Cyber risk is real but outweighed by productivity** Hoving says the AI-driven threats are serious, but that the productivity gains are 'multiple times' the extra protection work. Banks are responding by paying more attention to low and medium vulnerabilities, which are easier to exploit.

## Questions this episode answers

### Why do so many AI pilots fail to scale in banks?

Hoving says generative AI is very easy to pilot but very hard to scale out. ING avoids this by checking whether a use case can scale before investing heavily, and by relying on standardised technology and data across its markets.

### What did ING do before AI that now makes AI easier to roll out?

Over about ten years it modernised old core banking systems, standardised its products and put proper data foundations in place. A pilot that works in one market on those systems can then be rolled out to others with confidence.

### How does a global bank handle different AI rules in different countries?

Hoving says ING builds to the standards of strict, forward-looking regulators such as BaFin in Germany and the Monetary Authority of Singapore. That way its platforms meet the highest bar and so comply with the lower ones.

### Will AI agents replace people in loan approvals at ING?

Not for large or risky loans, according to Hoving. Machine learning already decides small, low-risk applications, and agents help analysts by gathering data, but they must escalate to a human when they go beyond set limits.

### Is AI making banks less safe from fraud and cyberattacks?

Hoving calls the new AI-driven threats a real risk that must be taken seriously. He still believes the productivity benefits are multiple times larger than the extra protection work, and says banks are now paying more attention to low and medium vulnerabilities.

## Facts

- **Bouke Hoving** [0:01:25]: "Well, in essence, I joined the financial industry in 2019. It took me 20 years, although I studied economics and applied physics, but spent the first 20 years of my career in the telecommunications industry."
- **Bouke Hoving** [0:01:25]: "But the last 10 years, I've been leading the technology function in the Dutch national telco KPN."
- **Bouke Hoving** [0:02:11]: "we're covering the financial needs of corporates and financial institutions across 35 markets."
- **Bouke Hoving** [0:02:11]: "over the past 10 years, we've been working on building platforms which are able to serve client needs."
- **Estelle Quek** [0:00:03]: "ING has moved over 90 % of its Gen AI pilots into production, built a governance framework that operates across 19 countries, and is now pushing into agentic AI."
- **Bouke Hoving** [0:03:55]: "we look at strict regulators. Bafin in Germany is a good example, but definitely also the monetary authority in Singapore is a good example."
- **Bouke Hoving** [0:05:53]: "we apply two filters first the input filter to the extent we are allowed to use those data sets and then on the output side"
- **Bouke Hoving** [0:06:46]: "we have built our platforms across three hubs in the world, Amsterdam, Bucharest, and Manila being our core locations."
- **Bouke Hoving** [0:06:46]: "And those are complemented by regional locations such as Singapore and London."
- **Bouke Hoving** [0:08:18]: "Well, as ING, here we are only active in Asia for corporate and financial institutions."
- **Bouke Hoving** [0:09:01]: "we had selected six areas ranging from lending, KYC, the contact centers, AI for engineering"
- **Bouke Hoving** [0:11:32]: "obviously paying more attention to low and medium vulnerabilities, which are more easy to exploit."
- **Bouke Hoving** [0:12:55]: "where we have replaced a lot of core banks which had been installed 35 years ago."
- **Bouke Hoving** [0:14:39]: "in particular in Asia, which has been at the forefront of those modernization projects, the majority of the markets is now on target platforms."
- **Estelle Quek** [0:15:36]: "I understand that ING is delivering personalized messaging to more than 4.6 million customers through its Gen.AI marketing engine."
- **Bouke Hoving** [0:17:18]: "In some campaigns, that gives a double-digit increase in success rates"
- **Bouke Hoving** [0:17:35]: "it could easily be rolled out to markets such as Belgium and Spain because they were already on the same technology foundation."
- **Bouke Hoving** [0:19:08]: "In the past, this essentially involved quite some interaction with the customer, which was typically done by a KYC agent."
- **Bouke Hoving** [0:20:22]: "Essentially, we concluded that a lot of the lending applications, in particular the lower volumes, the lower risk ones, and machine learning can essentially do the decisioning on those files."
- **Bouke Hoving** [0:22:16]: "our colleagues in the retail bank have already successfully brought this to production for mortgage applications in the Netherlands."

## Transcript

**[0:00:03] Estelle Quek:** Welcome to CIO Leadership Live ASEAN. Artificial intelligence is reshaping banking globally, but in Southeast Asia, the stakes are particularly high. Banks are navigating a region where regulatory framework vary dramatically from market to market, where talent is scarce and data localisation requirement complicate deployment, and where a significant portion of the population remains underserved by traditional financial services. Getting AI right is harder than anywhere else, and the cost of getting it wrong is real. ING has moved over 90 % of its Gen AI pilots into production, built a governance framework that operates across 19 countries, and is now pushing into agentic AI.

**[0:00:53] Estelle Quek:** Today, I'm joined by Bouke Hoving, Global CIO of Wholesale Banking, to understand what the execution looks like from the inside and what it means for the CIOs in this region. Boker, welcome and it's lovely hosting you today.

**[0:01:10] Bouke Hoving:** Thank you very much, Justel.

**[0:01:13] Estelle Quek:** Could you please start by sharing with our CIO audience your career trajectory? How are you so successful and how did you become the local CIO of Wholesale Banking at ING?

**[0:01:25] Bouke Hoving:** Yeah, it's a pleasure. Well, in essence, I joined the financial industry in 2019. It took me 20 years, although I studied economics and applied physics, but spent the first 20 years of my career in the telecommunications industry. Also a very fascinating industry. Had various roles there, commercial roles. But the last 10 years, I've been leading the technology function in the Dutch national telco KPN. And in 2019, switched to the banking industry. So it was time for a change.

**[0:02:11] Bouke Hoving:** I was very much looking for a role where I could also get exposed to the challenges of operating the technology function on a more international scale. And I'm doing this role now for three years. So, in essence, we're covering the financial needs of corporates and financial institutions across 35 markets. And do this by means of a global operating model. So over the past 10 years, we've been working on building platforms which are able to serve client needs.

**[0:02:56] Bouke Hoving:** Of course, that's a fairly diverse footprint. To the point you made in your introduction, obviously a very diverse footprint in Asia, but also Europe and the Americas. so customer needs but also regulations vary quite a bit and these global platforms need to cater for that variation but it's about the challenge i was looking for and i absolutely found so still like my job each and every day that's fantastic thank you much that's a very

**[0:03:31] Estelle Quek:** impressive career that you have now service isia does not have a single ai regulatory framework Singapore, for example, is ahead. Others are still developing deaths and some have no clear position yet. For a global bank operating across these markets, how do you build a consistent AI governance posture when the regulatory ground beneath you is uneven?

**[0:03:55] Bouke Hoving:** Yeah, that's a great question and not a new one to us because obviously we face very similar questions when we need to deal with cybersecurity, change management, resiliency frameworks, that there's quite a bit of variation across the footprint. The very simple approach we tend to take is that we look at strict regulators. Bafin in Germany is a good example, but definitely also the monetary authority in Singapore is a good example. of forward-looking regulators who tend to have a high bar.

**[0:04:41] Bouke Hoving:** And by looking at those forward-looking regulators and ensuring that our platforms, our approach, our engineering practices do meet those standards, we have been able to essentially comply to all regulatory standards across that footprint. So that is how we take this approach. Obviously, in AI, it's shaping up, but we can already look at regulations in place for machine learning, consultations such as the Singaporean consultation, which is ongoing, and regulations which have been published in Europe. And based on those, we navigate across a challenging regulatory framework.

**[0:05:37] Estelle Quek:** Now, several Southeast Asia markets impose strict data localization requirements. When your AI models depend on large cross-border data sets to perform well, how much does that fragment your ability to deploy at scale within the region?

**[0:05:53] Bouke Hoving:** well in essence we apply two filters first the input filter to the extent we are allowed to use those data sets and then on the output side to another extent the reports or the applications are allowed in the particular jurisdictions And based on those two variables we have, we are very confident we can comply to each and every regulation across the sector.

**[0:06:28] Estelle Quek:** Thank you, Berke. The shortage of AI literate talent in financial services is acute across Southeast Asia. Beyond hiring, what does a bank like yours have to build internally to close that gap? And how long does that realistically take?

**[0:06:46] Bouke Hoving:** Yeah, AI talent indeed is very scarce. It's also very hard to hire experienced people as an industry. And we are essentially in the face of building up that experience. But we do have a very well-articulated approach to talent management in general beyond AI, where in essence, we have built our platforms across three hubs in the world, Amsterdam, Bucharest, and Manila being our core locations. And those are complemented by regional locations such as Singapore and London.

**[0:07:31] Bouke Hoving:** And that footprint allows us to tap into talent pools per location. And from the perspective, we have access to pretty diverse talent pool can also benefit from 24-7 proximity to important markets. So that's in general our approach to talent management in technology and AI.

**[0:08:01] Estelle Quek:** Southeast Asia still has a significant underbanked population. AR has the potential to extend credit, and financial services to people who were previously too costly to serve. In your view, is that a real opportunity which banks are actively pursuing or is it largely aspirational?

**[0:08:18] Bouke Hoving:** Well, as ING, here we are only active in Asia for corporate and financial institutions. But what I can say in general about the industry, that, to your point, AI gives us a unique opportunity to come up with different service models, lower price points potentially. So I think in general, that is definitely a threat, which banks active in those segments can capture.

**[0:08:55] Estelle Quek:** Do you think ING would come up with new products, for example, or enter a certain market to capture these opportunities?

**[0:09:01] Bouke Hoving:** Well, what we are looking for, first and foremost, is making the service on our existing products to our existing clients better. So far, we started with a fairly focused approach on thinking about use cases. So we had selected six areas ranging from lending, KYC, the contact centers, AI for engineering, and all those use cases were aimed at eliminating friction in which customers were experiencing in dealing with ING, getting back to faster execution. And those are the benefits we were predominantly looking for.

**[0:09:54] Bouke Hoving:** Now, obviously, we have moved beyond that phase where we were focused at those six use cases. We're looking at a broader set of use cases. But still very much looking forward at eliminating friction and getting to faster execution. And obviously, also looking for efficiency where we come across efficiency opportunities.

**[0:10:21] Estelle Quek:** You would agree with me that AI is making fraud and cyber threats more sophisticated and harder to detect. For banks in Southeast Asia, where digital adoption has outpaced security infrastructure in certain markets, how much of a step change in risk are we looking at?

**[0:10:38] Bouke Hoving:** Well, I think the risks triggered by the whole hyper-anti and the mythos model is definitely a real risk. we need to protect ourselves against. But overall, I would argue that the opportunities in terms of productivity AI is bringing us are multiple times the additional work which comes with protecting ourselves against those risks. So overall, I'm positive about the opportunities it will bring. But those are, to your point, and partly offset by some risks, we also need to take very serious.

**[0:11:21] Estelle Quek:** Could you share some example? I mean, not confidential data. How is your bank kind of like safeguarding against this advanced cyber threat?

**[0:11:32] Bouke Hoving:** In general, what you see the whole industry now doing is obviously paying more attention to low and medium vulnerabilities, which are more easy to exploit. Whereas the whole industry was a bit more focused focused at the high vulnerabilities as an industry-wide example of what is currently changing.

**[0:11:55] Estelle Quek:** Now, AI transformation failures, although common, they are rarely publicized. From what you have seen across global markets, what are the conditions that most reliably predict failure? And are banks in Southeast Asia particularly exposed to any of them? And in your view, why?

**[0:12:16] Bouke Hoving:** Yeah, that I would say is definitely new versus other large transformations such as digital. In AI, in particular with Gen AI, it's very easy through a pilot, but it's very, very hard to scale this out. And that's what you read a lot about. and that also explains why we initially were very much focused at our six initial use cases to really put a lot of effort in making them scale and reflecting on the learnings which came on the back of those pilots.

**[0:12:55] Bouke Hoving:** And we essentially have learned to not underestimate and to spend a lot of time in asking ourselves the question, do these use cases scale or do they have the potential to scale before we really double down on very initial proof of concepts and starting investing in those. and for an international bank like us, we have learned that the enormous time and effort and investments we have done in a global operating model over the last decade, definitely also in this region where we have replaced a lot of core banks which had been installed 35 years ago.

**[0:13:44] Bouke Hoving:** We have modernized those core banks. And on the back of that modernization, we have standardized our products. We've put in proper data foundations, but we've also harmonized the products. We deliver to our customers. That was initially, first and foremost, a modernization exercise. that was meant that standardizing our customer experience, getting a more consistent experience for our customers. But it's also given us a very scalable technology and data foundation. And we benefit a lot from that scalable data and technology foundation because once we have found a pilot which works in one of those markets which is on the target systems, And we are quite confident we can scale this out.

**[0:14:39] Bouke Hoving:** And in particular in Asia, which has been at the forefront of those modernization projects, the majority of the markets is now on target platforms. So, well, once we have a proof of concept which is successful, for example, in KYC, to accelerate the customer due diligence process or in lending to accelerate the credit approval process. Yeah, we are very confident we can scale it to our markets which are on core platforms which have been modernized. So that scalability and the importance of having a common foundation, that for us is the big success factor.

**[0:15:28] Bouke Hoving:** Also explaining our additive success in combination with our focused approach in the pilots we've scaled so far.

**[0:15:36] Estelle Quek:** Thank you. I understand that ING is delivering personalized messaging to more than 4.6 million customers through its Gen.AI marketing engine. Can you give me some real tangible results? What has that changed commercially in retention, in cost of customer engagement, for example?

**[0:15:56] Bouke Hoving:** Yeah, this is one of the six focus areas colleagues in our retail bank have been working on. And in essence, what this does is previously personalization was very much a rule-based approach. But for personalization, you had to specify the rules up front. And obviously with AI and machine learning, you could already make that a bit more intelligent and eliminate some of the rules and let the algorithm learn to personalize. And based on the customer feedback, adopt itself.

**[0:16:41] Bouke Hoving:** But with Gen.ai, you can even make it much more personalized, in particular in how you essentially communicate with your end customers and do the segmentation of your client segments much more on a micro level versus what the traditional rule-based engines were allowing you to do so. And that gives an uplift, for example, in next-best actions, which is significant.

**[0:17:13] Estelle Quek:** Any value you can share with us, like percentage-wise?

**[0:17:18] Bouke Hoving:** In some campaigns, that gives a double-digit increase in success rates, so does compage.

**[0:17:25] Estelle Quek:** Okay, fantastic. Now, what did it take to get the data infrastructure right before that was possible, and how did you do it?

**[0:17:35] Bouke Hoving:** Getting the data structure right and the technology right was something we have been investing on a lot over the last decade. In wholesale, I mentioned the modernization of the core banks. But if I, for example, take the contact center use case in the retail bank, that contact center had already been scaled to all our major retail markets in Europe and in Australia. So once the contact center chat function was powered by Gen.ai and that was successfully piloted in the Netherlands, It could easily be rolled out to markets such as Belgium and Spain because they were already on the same technology foundation.

**[0:18:26] Bouke Hoving:** So again, there are commitments and our conviction that we as a bank need to build platforms and use those platforms to serve customers in multiple markets is really benefiting the retail as well as the wholesale bank in scaling out your AI and AI use cases across our footprint.

**[0:18:48] Estelle Quek:** Now, reducing manual onboarding tasks by 80 % across 19 countries is a significant operational result. Could you please walk us through what that process looked like before AI and what does it look like now? And where does the remaining 20 % of the manual work sit with and why?

**[0:19:08] Bouke Hoving:** This is in essence referring to the customer due diligence part of the onboarding process. In the past, this essentially involved quite some interaction with the customer, which was typically done by a KYC agent. Now we can leverage a lot more data and a lot of the checks are done in the background versus these checks requiring a reach out to customers. and that has significantly boosted onboarding timelines. And obviously, if there are still on an exceptional basis missing data elements of checks to be done, then there will still be a reach-out required, which is then essentially explaining the 20%.

**[0:20:04] Estelle Quek:** Moving from a chatbot that answers questions to an agent that collects documents, runs credit checks and progresses an application is a fundamentally different kind of AI deployment. Where does that agent hand off to human exactly, and how do you decide where the line sits?

**[0:20:22] Bouke Hoving:** Yeah, that is a fascinating question. And maybe it's also good to not only focus on Gen.AI and maybe even a Gen.Tech where this question pops in, but also take a couple of steps back and look at, well, the whole approval process for loans, which is one of the areas where the whole industry is focusing on. We were already, before Gen AI was even a trending name in technology, looking at machine learning. Essentially, we concluded that a lot of the lending applications, in particular the lower volumes, the lower risk ones, and machine learning can essentially do the decisioning on those files.

**[0:21:12] Bouke Hoving:** And if you then look at the files which fall outside that category, either by size or by risk, we, in particular in the wholesale domain, lending amounts are quite substantial. still have a credit analyst look at the application. But in terms of extracting all the data elements out of the annual reports, in terms of checking the credit agency, a lot of those tasks can be done by agents or can be done by a co-pilot. And that's, in essence, what we are doing for those use cases, where the current focus is very much on augmenting the analyst and pre-populating the existing approval flows by all these data elements.

**[0:22:16] Bouke Hoving:** Abbott were slowly but surely moving towards a more agentic approach, where our colleagues in the retail bank have already successfully brought this to production for mortgage applications in the Netherlands. And there, to your question, yeah, obviously also there, if the Asians start to operate outside, well, their guardrails, then they need to escalate towards a human. So it will always remain a close interaction between machine learning, ALGOS, Asians, and humans. And we are, well, just at the early phase of how this will exactly play out.

**[0:23:02] Bouke Hoving:** And there will probably be a bit of variation across the full spectrum of mortgages till large corporate syndicate loans. And on the higher end of the spectrum, humans will remain involved. But more and more technology will take, well, heavy lifting out of the process and make the whole process also significantly more faster and customer-friendly in terms of reducing the need to reach out to customers.

**[0:23:35] Estelle Quek:** Pogut, thank you so much for being very open about the progress and the hard parts of your AI transformation journey. This has been a valuable conversation.

**[0:23:45] Bouke Hoving:** Well, thank you very much for having me here as hell. It was a pleasure.

**[0:23:49] Estelle Quek:** What your team and you have built at ING gives this region's technology leaders a genuinely useful reference point. as proof that it can be done at scale, responsibly and with measurable results. Our conversation makes clear that AI transformation in banking is an execution problem, one that requires governance built in from the start, a clear view of where AI and humans work together and the discipline to move from pilots into production without losing control of the process.

**[0:24:20] Estelle Quek:** These are the challenges CIOs across Southeast Asia are grappling with right now. To our CIO audience, thank you for joining us. Until next time.
