Generative AI is changing debate on explainability, says Deutsche

Banks are already deploying artificial intelligence and machine learning in a variety of roles, and see many other potentially valuable use cases fast appearing on the horizon. In fact, Deutsche Bank’s chief innovation officer Gil Perez refers to 2023 as “the year of AI and ML”.

“We’ve been working on AI for multiple years, establishing partnerships and building cloud infrastructure. Now, with generative AI in play, we believe we’ve reached an inflection point,” Perez says.

One crucial brake on its adoption, however, has been the discomfort of regulators who want adequate assurances that banks can risk-manage the technology appropriately. In particular, the explainability of AI models has become a vexed question for banks. But Perez thinks the emergence of generative AI could move that debate forward.

We need to transition to observability of models and work with regulators to address their concerns

Gil Perez, Deutsche Bank

Over the past few years, regulators have agreed that AI models deemed ‘black boxes’—those lacking fully explainability—cannot be deployed in production at all. But generative AI “broke the seal” on that, says Perez.

Today, he notes, discussions are shifting towards “observability”, where regulators expect banks to understand the models’ input and outputs, but not necessarily their whole inner workings—an almost impossible task for large language models (LLM) due to their extensive parameters.

“We need to transition to observability of models and work with regulators to address their concerns,” says Perez. “Otherwise, generative AI and LLM will have limited adoption within the financial industry.”

While there are ongoing challenges, Perez believes the industry will devise different approaches to tackle the issue of explainability in the next few years.

“Here’s an interesting thought: in the future, we could develop a model to explain other generative models,” says Perez. “Regulators and the financial industry will need to experiment and to adopt the new ideas that work.”

As part of a three-year initiative launched at the end of 2022, Perez aims to move promising experimental use cases into production next year. In 2025, he hopes to further scale up the use cases and maximize efficiency gains.

To achieve these goals, Perez says the bank needs to concurrently address data risks, model risks and regulatory challenges. While Deutsche is now prioritizing use cases with fewer regulatory constraints, Perez acknowledges that for the technology to be efficiently scaled up and ultimately deployed on the client side, they need to closely monitor and adapt to the swiftly evolving regulatory landscape, which is still in “uncharted territory” at the moment.

For a global bank, however, he says the most significant challenge is dealing with multiple regulators, particularly if there is a lack of consistency in AI regulations across jurisdictions.

“If you are just trying to match up rules from two different regulators, you could say it is doable,” says Perez. “But when you work with over 40 regulators, and especially when some of their requirements do not overlap, how should you deal with it?”

Emerging use cases

Against the background of changing regulatory attitudes, Deutsche made a strategic pivot in its latest three-year AI initiative. The bank is transitioning the innovation focus from classical AI and machine learning development to generative AI technology, and has been experimenting with various use cases.

While remaining cautious around front-office applications, Perez says his team has made progress in leveraging generative AI to enhance risk management.

In one of the applications under exploration, which Perez refers to as the “adverse media use case”, Deutsche harnesses generative AI tools to collect data from public sources such as court documents and media to identify individuals and entities facing criminal charges across jurisdictions. The team then matches the information with its customer lists and assesses whether to take preventative actions.

“Monitoring global media and extracting from millions of pages across multiple languages to maintain an up-to-date customer list are time-consuming tasks,” says Perez.

Generative AI’s ability to discern, extract individuals or companies’ names, and understand the context is extremely helpful to us due to the sheer amount of data out there

Gil Perez, Deutsche Bank

That’s a general problem facing the whole industry, but it is especially pertinent for Deutsche. In July this year, the bank was fined $186 million by the US Federal Reserve for failing to improve its anti-money laundering controls and enhance customer due diligence, as it had promised to do as far back as 2015. This was the third such fine since 2015.

Now, with the advancement of generative AI granting banks a way to monitor customers and maintain regulatory compliance more efficiently, there is an opportunity for the German lender to close the gap.

“Generative AI’s ability to discern, extract individuals or companies’ names, and understand the context is extremely helpful to us due to the sheer amount of data out there,” says Perez.

While Deutsche has not yet set a specific date to roll out this specific know-your-client technology, Perez expects it to be among the early use cases adopted, because it does not rely on proprietary client information that could throw up data privacy obstacles.

In addition, Perez said his team is actively testing tens of other use cases. Most focus on areas where Deutsche can rapidly enhance its employees’ productivity with minimal regulatory challenges. These range from facilitating briefing reports for investment bankers before client meetings to assisting with coding, to improve software developer efficiency.

“Our AI transformation is not driven by technology for its own sake, it is driven by our business needs and objectives, with technology enabling it,” says Perez.

Journey to the cloud

Deutsche’s venture into AI can be traced back to the end of 2019, when Bernd Leukert, a member of the bank’s management board, initiated a technology transformation across the organization. During this time, Leukert created a new unit called TDI (Technology, Data, and Innovation) and placed Perez at its helm in early 2020.

“A few years ago, Deutsche coalesced all of its technology, data and innovation functions into a single team,” Perez says. The goal is to move the company forward into a unified, modern infrastructure, the majority of which is in a public cloud—the key element for AI and ML success.

“The explosion of generative AI nowadays is tightly coupled with the capabilities of cloud technology. A robust cloud infrastructure provides us with the computational power needed for training large language models,” Perez says.

In December 2020, Deutsche entered a strategic partnership with Google Cloud. Perez says his team chose Google for two key reasons: first, beyond Google’s AI and ML pedigree and focus on data, the tech giant is willing to help the bank address regulatory requirements across jurisdictions; second, Deutsche values the open environment Google provides, which allows firms to test a wide range of large language models.

When banks transfer data to the cloud, they are required to make a physical audit of any data centers operated by their cloud service provider. Moreover, for each country in which a bank operates and stores data, it needs access to not just one but at least three data centers for fail-over protocols, with each passing regulatory inspections—a process that can take at least six months.

“In this industry, regulatory approval is a licence to operate, and without it, you can’t do much with your business,” says Perez. “So, thinking about the numerous countries we operate in and the number of regulators we need to work with, it is important for us to have a close relationship with a cloud provider—in our case Google—who is willing to support those regulatory requirements and help us adapt to an unknown future.”

Open choices

Explaining the open environment Google has offered, Perez says Deutsche can access a diverse collection of more than 100 large language models through Google’s model garden, a generative AI support function introduced to its AI platform in 2023. The array of available models extends beyond Google’s proprietary ones, granting Perez’s team the flexibility to choose and fine tune open-source models or those from other third-party vendors based on their specific use cases.

“We are still in a very early market, and it is still unclear which approach with large language models will prevail—whether a single large language model will dominate or if a combination of smaller models will prove more successful,” says Perez. “Therefore, we want to try different options to see which one works out best.”

Deutsche Bank is also collaborating with multiple AI startups to find areas of opportunity and innovation that are unique to the bank but not covered by Google or other large technology companies.

For example, Perez says they have been actively working with AI21 Labs to evaluate their large language models’ capability while comparing them with those of Google and others. The Israel-based startup provides a suite of text-generating AI tools that can be customized to meet specific business needs, landing $155 million in a Series C funding round at a $1.4 billion valuation in August 2023.

The evaluation criteria include—but are not limited to—the cost, and the speed of fine tuning these models, Perez adds.

Deutsche also established an AI strategic partnership with Nvidia in December 2022, after months of joint exploration on numerous use cases. The partnership will initially focus on three specific areas: risk model development; high-performance computing; and the creation of a branded virtual avatar that was evaluated by internal employees to interact with HR systems. Deutsche declined to comment on whether it has purchased AI chips from its partner.

It is noteworthy that both AI21 Labs and Nvidia have recently partnered with Google Cloud, which Perez finds “very encouraging” for Deutsche.

“We are seeing our choices of strategic partners align closely with Google,” says Perez. “We don’t know what will happen in the future, but a shared ecosystem of partners and an open architecture is a solid foundation to build upon.”

Inevitably, cloud service providers will face more scrutiny from both risk managers and regulators as they become critical providers not only for banks directly, but also for other third-party AI vendors.

Update, November 7, 2023: Article adjusted to correct the date of Bernd Leukert's arrival at the bank