The biggest near-term opportunity of generative AI (GenAI) in retail banking is in delivering operational efficiencies for incumbents, according to a new report.
GlobalData’s Generative AI in Banking report asserts that expectations for GenAI have been tempered since the launch of ChatGPT in December 2022 thrust the technology into public consciousness.
Smarter leaders trust GlobalData
“Many banks now have beta deployments underway and have found it takes longer – and costs more – to achieve the requisite level of accuracy in output,” it says. “The set-up costs, in terms of processing power, and model training, mean ROI is in many cases further away than originally thought, which has dampened enthusiasm for an ongoing scatter gun approach …”
However, it suggests that there are opportunities for using GenAI to make “incremental improvements to existing processes, job roles and procedures”. Indeed, the report suggests such changes could deliver increases of 20-30% in productivity, which for major banks would represent multi-million-dollar cost savings.
“Retail banks are experimenting with GenAI across front, middle and back-office activities,” it says. “New digital banks, with more modern tech stacks, are able to operate more quickly and cheaply, but big incumbent banks have the scale to drive significant cost reduction and process efficiency.”
Despite some tempering of expectations for AI, GlobalData forecasts that the sector will be worth $1.04trn by 2030, having grown at a compound annual rate of 39.1% from $103bn in 2023. The global GenAI market in retail banking will be worth $8.6bn by 2028, it forecasts, up from $261m in 2023.
Trends for GenAI in retail banking
Among the most notable trends of GenAI in retail banking – and, indeed, elsewhere – is the shift in focus from large language models (LLMs) to small language models (SLMs). With the recognition that LLMs require significant training, are prone to errors and are power hungry comes the acknowledgement that SLMs – which typically have fewer than 10 billion parameters compared to the potential trillions of LLMs – are more manageable, make fewer mistakes and require less training. Fundamentally, SLMs can be just as useful to companies by being focussed on a specific relevant topic area.
The report adds: “As training techniques improve, SLMs with fewer parameters are becoming more accurate. Also, they have a faster processing time. SLMs use smaller and more focused datasets, which means that training can be done in weeks, depending on the use case, in contrast to the several months for LLMs. The use of focused datasets makes SLMs particularly well-suited to domain-specific functions and small-scale applications. SLMs are ideal for mobile applications, edge computing and environments with limited compute resources.
“Having fewer parameters than LLMs, SLMs have shorter inference times and therefore lower latency. This makes them ideal for applications that are interactive and need to perform in real time. Being designed for specific applications, SLMs can handle data variations, monitoring and modifications to the primary training more effectively. This simplifies their maintenance and model optimisation. SLMs are more resilient to cyberattacks compared to larger models because their smaller datasets represent a smaller attack surface. This also simplifies the security process, as there is less data to protect, making it easier to identify and address vulnerabilities.”
Elsewhere, it is noted that there remains regulatory uncertainty about AI, not least as a result of Donald Trump having been elected as US President for a second term. Specifically, the report cites the new administration as preferring a light-touch approach, the EU’s AI Act as being the most advanced regulatory framework for AI and China as having a more government-centric approach.
The report says: “AI regulation, especially initially, is likely to be patchy, and playing a game of catch-up with advances in GenAI. As of today, regulation is being discussed and/or approved across different jurisdictions. But the approaches are likely to diverge on key aspects, such as the use of personal data. As a result, businesses exploring generative AI will face a quite uncertain global regulatory landscape.
“This is another reason giving some banks pause, especially those with a truly global footprint, subject to various regional differences. These players are forced to take a pragmatic approach in which they favour a ‘minimum common’ approach to stay compliant in all jurisdictions but with some regional customisation where it is most necessary. This, for example, has been an approach for truly global banks like HSBC, ‘common where possible, different where it matters’.”
