How to get the most bang for your banking buck with AI and machine learning

Especially given the implications for business. For instance, the primary purpose of AI in banking is to automate tasks and generate predictions using machine learning (ML) models (including complex deep learning). This requires massive computing power and datasets, as well as significant investment. 

Also, financial institutions are already major AI users. They rely on cloud-based ML services (e.g., AWS, Microsoft Azure, Google ML) and (mainly) private or mainframe hybrid cloud environments. 

There seems to be a new tech development almost every day, and understandably, everyone wants to use it immediately. But there’s much to be gained by assessing what you need to do realistically and which solutions can help you achieve that. And by ensuring that you’re using what you already have to its greatest effect — getting the biggest bang for your buck, so to speak — before splashing out. 

^{Source: S&P, AI In Banking} 

Three-quarters of financial institutions already use AI and recent developments signpost an increasing desire toward bringing GenAI into the tech mix as soon as possible. GenAI’s low-code/no-code software could provide significant cost savings too, including cutting operational costs and risk linked to outdated infrastructure and labor-intensive processes. 

^{Source: Gartner Impact Radar 2023} 

While upscaling risk management, AI could incidentally affect the perception of bank risk profiles. Banks that can accurately price credit risk by employing patterns buried in the data to figure out the likelihood of customer debt repayment will enhance their workout models, reduce troublesome loans and optimize accuracy. 

It’s possible to use AI to improve risk management principles in several ways. However, that won’t stop inferior implementation from generating reputational and operational risks and a negative risk position for the bank. 

Get it right. Partner with an experienced technology provider who knows the banking ropes. 

Europe’s new AI Act has established a comprehensive regulation for AI products and services within a risk- and use-case-driven environment. The Act includes responsibilities for foundation models supporting general-purpose AI systems. Similar moves are afoot in other countries like Canada, Brazil, Chile and the Philippines. 

In the United States, the National Institute of Standards and Technology (NIST) has established an AI Safety Consortium. Its aim is to work with industry to develop new standards, including sector-specific guidance driven by risk management. As well as governmental regulatory activity, there’s an increasing acknowledgment of industry’s duty to protect society by initiating or strengthening AI governance. 

The new AI regulation application is based on what risk users will likely introduce and which activities should be regulated. The EU has considered four kinds of risk: 

AI regulation for financial services will determine access. In financial services, AI systems analyze client data for credit scoring and fraud detection. 

From an analytics perspective, we’re in an exponential growth phase of this remarkable technological age. Consequently, virtually anything built 2 years ago is already out of date. And the question is, how does a company use outmoded technology in a controlled, non-toxic way, mindful of the knock-on effects of the slightest error? 

Take ChatGPT, for example. No one knows where the app’s information originates (or its bias/copyright/integrity). Yet more and more employees are encouraged to use the app to produce corporate blogs, articles, white papers and even lines of code. No wonder financial institutions are considering building ringfenced AI services (an extension of the ongoing company debate: “Do we go to the cloud or stay on-prem?”). 

This is a critical issue because, for effective AI and sophisticated analytics, you must be sure that your data integrity is on point. That being so, it makes sense to work on-prem, insulated from the internet. Your corporate large language model (LLM) or grammar-checking app can be trained exclusively on the company’s internal documentation (it will never leave the mainframe). In fact, the most secure solution flies in the face of contemporary IT thinking, which advocates sending everything to the cloud. 

Mainframe hybrid cloud architecture allows you to adopt a best-fit strategy for application implementation. Here’s a banking application ecosystem example (across IBM Z and cloud). 

^{Source: IBM} 

The increased accuracy of deep learning models enables them to cut excessive false positive rates substantially. This means banks can pass transactions through fraud detection without impacting the customer experience or hemorrhaging money. 

By implementing fraud analysis on the mainframe using Umbrella.ai and IBM’s Telum chip, 100% of your transactions can be run across deep learning models in real-time, reducing fraud at the average tier-1 bank by $120 million a year.