EXCLUSIVE: “Proceed with Care” – Bahadir Yilmaz, ING Bank in ‘The Paytech Magazine’

ING Bank has taken a ‘prudent and cautious’ approach to the introduction of GenAI, following the principles of safe design, as Chief Analytics Officer Bahadir Yilmaz explains

“Applying AI and generative AI techniques to a business problem is only five per cent of the job. The other 95 per cent starts after that, building all the systems around it to make it safe.”

That’s Bahadir Yilmaz speaking, chief analytics officer in charge of a team of 400 data scientists, data engineers and product managers at European bank ING. While they are at the coalface, weighing the risks and benefits of implementing AI, including large language models where appropriate, the other 60,000 or so of the bank’s employees are not left in the dark about the technology.

ING is building what Yilmaz has described as “an advanced data science workforce” and part of that involves awareness training that “highlights the opportunities but also the (largely unknown as yet) risks” associated with AI. ING has taken a systematic approach to assessing those dangers.

“We have a 20-step process, which evaluates any AI system for 140 risks,” explains Yilmaz. “And only after we say OK to all of these 140 risks do we go into production.”

That applies to systems built in-house  and – perhaps even more so – to those foundational models that, increasingly, the bank is sourcing from big tech and other vendors. Here, says Yilmaz, it works on the principle of “believe nothing you read on the declaration but test it yourself against banking standards.”

Developing those new methods and measures to really kick the tyres is one of the biggest challenges for the bank right now.

“These types of controls are necessary for organisations, but also for society to make sure that AI applications are safe and secure to use,” says Yilmaz.

From chatbots to reconciliations, lending to trading, it’s creating these protective systems around AI models that takes the most effort. But the bank’s plan is to position analytics and AI as “the most critical enablers in transforming ING’s operations”, so it has to get them right. There is also, of course, a substantial cost associated with such computing power, which makes prioritising domains that present a key business risk – like KYC, detection of financial crime, fraud and sustainability – a prudent focus of time and money when it comes to developing an AI tool for the process.

GenAI is certainly more tricky to evaluate. It’s still in unchartered territory, so applications within the bank have been limited to those where perhaps it can do least damage and deliver the most benefit. So far, this “prudent and responsible approach” has seen ING focus on five key areas: chatbots, personalised marketing, KYC solutions, coding/ software engineering and ING Wholesale Banking (WB) lending.

“We’ve been building chatbots since 2017 and have used several different technologies,” says Yilmaz. “All these solutions were super rudimentary. They weren’t really getting the answers that customers were looking for.

“But with the launch of large language models, and especially with ChatGPT, this changed. People saw how good they could be in getting the answer they were looking for, quickly and accurately.

“The key is finding the most suitable technique and applying it to your problem to generate real value”

“We noticed this at ING, too, and shifted our technology from a more statistical natural language program (NLP)-based approach to a large language model-based approach.”

The bank worked closely with McKinsey over seven weeks to build, test, and launch a bespoke customer facing chatbot using GenAI technology. Before any responses were sent to customers, a series of guardrails was applied (to avoid the chatbot giving advice on mortgages and investment products, for example), built using out-of-the-box tools from McKinsey’s AI and machine learning innovation hub, QuantumBlack Labs.

Something new

The new chatbot launched in September 2023, making it the first-of-its-kind, real-life customer-facing pilot conducted in Europe. ING saw an immediate improvement: the more nuanced and contextual answers that GenAI could provide instantly translating to a 20 per cent uplift in customers not having to wait in queues to speak to an agent. The chatbot is projected to eventually impact 37 million customers when it’s scaled to all of ING’s 10 retail markets.

“Clients want instant answers to their questions and they want instant solutions to their problems. Chatbots are highly capable of doing that,” believes Yilmaz.

That makes it sound easy; it isn’t.

“By getting service from one of the big tech companies and plugging it into your website, you can have a chatbot that is able to be your client services channel in a day,” says Yilmaz. But the job is not complete until you have ensured there are no biases in all the systems around it, so you’re not discriminating against anyone, and that your system is safe and the information is secure.”

While “any problem can be optimised and made better with these techniques,“ he’s not advocating that banks apply them to solve all of their problems.

“You should put them in a priority order and start with the ones that really make sense for you. Key is finding the most suitable technique and applying it to your problem to generate real value,” he says. “Lending is, historically, one of the first areas where banks have started using AI models, including advanced models. We see more institutions using the power of their transactional data and coming up with better lending propositions. They are able to provide loans in a matter of minutes instead of days.”

But his concern is that, as use cases multiply, the industry will lack the skills needed to build these systems and put appropriate guardrails in place and at scale.

“We need more and more people who know how to do it. We cannot just keep building these solutions with small teams of experts. So, that’s the first transition that we have to go through,” says Yilmaz. “The second transition needs to happen in certain parts of our workforce where their job is changing. It’s not about extracting information from documents anymore but verifying information that is extracted by AI.

That’s necessitates a different expertise and that transition also needs to be supported.

“Think about it: you’ve worked in anti-money laundering operations all your life, reviewing customers’ annual reports, then AI tools come along to answer all the questions. But they won’t be perfect. So your job is going to be reviewing whether that information is correct.

“That’s a slightly different skill that requires a different way of thinking. Your concentration has to be on different things. And that is a change that, as a bank, we also have to facilitate.”

This article was published in The Paytech Magazine Issue 15, Page 08-09