Early trials with generative AI show that the tangle of legacy code that hinders every aspect of the transformation of banks’ core products and services could be quickly and effectively resolved. By Alvaro Ruiz, Global Core Banking Lead, Accenture.
At the heart of most traditional banks is a mainframe computer running the software that defines and drives the organization’s core processes. This is where many of the bank’s problems start. To meet customers’ and regulators’ growing demands, to improve cost efficiency, and to keep up with the rapid pace of change – which includes taking advantage of emerging technologies – banks need the agility to innovate quickly. This is not something the industry is famous for.
The reason is that the core system behind every banking product comprises many thousands of lines of computer code. Most of it was written decades ago, in languages that can no longer meet the needs of modern banking. Over the years it has become multi-layered, convoluted and fragile. This is due to repeated changes, add-ons and connections with newer applications, and the general tendency to retain rather than decommission outdated applications. These issues, together with the fact that the code tends to be poorly documented, makes it difficult to understand and upgrade it.
Banks have long wanted to re-engineer or replace this code – convert COBOL to a modern language like Java, for example – but this would be a mammoth task requiring hundreds of engineers and taking several years. Most source code programmers have long since retired, making it difficult and expensive to find and recruit them in sufficient numbers. This, in turn, adds to the risk of the exercise. And so it remains on the back burner, an important priority but one that is hard to justify in the current circumstances.
Legacy tech: A costly roadblock to banking innovation
At last, however, we are seeing light at the end of the tunnel. We believe that generative AI, combined with new composable, interoperable and coreless architectures, could offer the solution to this decades-old problem. We are working with a handful of banking clients to test this approach, and the early results are very promising. As my colleague Michael Abbott, Accenture’s global banking lead, told American Banker: “It’s early days, but we’re seeing 80% to 85% accuracy.”
The process starts by using specialist generative AI models and a process called retrieval augmented generation (RAG) to reverse-engineer the code. This allows us to understand and document the requirements which the code is designed to meet.
The next step is forward-engineering, for which we see two main paths. The first is the automatic recoding of the software into a modern and versatile language, using an ISO-functional approach and/or specialized generative AI models, to re-imagine the functionality that is required to meet the current objectives of the technology and/or the business. We then use generative AI to automatically test every part of the new code and its performance, and to facilitate the transition from a mainframe hardware and software stack to a modern set of frameworks and compute technology.
Put simply, we replace the old code with new programs that are simpler and more flexible, and that support the bank’s modernization strategy.
“It’s early days, but we’re seeing 80% to 85% accuracy.”
Having composable, interoperable and coreless architectures is key to most banks’ modernization, as it allows parts of their legacy applications to co-exist frictionlessly with modernized parts and even with third-party products from different sources. It also allows banks to employ different hosting models at the same time, which is essential to meeting new business requirements and generating timely business outcomes while legacy modernization is under way.
The benefits of this approach are potentially game-changing. A core modernization project that in some reported cases cost more than US$700 million and caused years-long business bottlenecks during their execution could now be completed in a fraction of the time with no negative impact on the business. The cost and risk advantages are obvious. In recent work with a global bank we converted 25,000 lines of legacy code, cutting the reverse-engineering effort by 50% and boosting testing efficiency by 30%. This saved more than 50% of the original budget.
More importantly, our analysis indicates that banks that modernize their core could potentially increase their return on equity by 8.3 percentage points by improving their manufacturing, distribution and servicing. They would also dramatically facilitate risk management and regulatory compliance.
It is of course vital that we guard against the tendency of generative AI to plagiarize and, if it doesn’t know the answer, to fabricate. RAG helps in this regard by drawing on the bank’s knowledge base, including its existing code repository, rather than unverified external sources. However, until generative AI matures and these failings are remedied, the new code does need to be carefully checked and tested.
From legacy to leading with a modern digital core
Every bank knows that a modern digital core is critical to its ability to compete and meet customer needs. I believe we are on the cusp of putting this elusive goal within the reach of every financial services organization – quickly and affordably, while keeping a tight control on risk.
To find out more about this topic we have a section titled The Key to the Core in our Top 10 Trends for Banking in 2024. We have also just published a new report, The Age of AI – Banking’s New Reality, which explores the potential role of generative AI throughout the bank.
If you would like to find out how our code conversion trials are progressing, please contact me directly at LinkedIn.