As ever, the focus was on payments – digital, real-time, cross-border - and the standards needed to support them. There was talk about how legacy mainframes are starting to become major inhibitors and the need to move towards cloud-native core banking systems. And the usual discussions on fraud, compliance, risk and security.

But, of course, the major topic was generative AI (GenAI). We spotted four themes:

• Risk and governance is a major hurdle: Staff working on GenAI are typically in their own disconnected silos and not connected to the business divisions. There doesn’t appear to be a single governance model for GenAI activity yet. Because of its lack of predictability and the newness of GenAI technology, banks are now exposed to greater risk. There is a space for AI in payments, but fraud is a significant hurdle. VALL-E takes 2 seconds to copy and reproduce a synthetic voice. Responsible use of AI is still a major challenge. As part of our latest GenAI research, one US banker remarked: “If you were to insert GenAI anywhere along your pipeline, we have to get regulators comfortable. On traditional tasks like Excel analysis, it's straightforward. But if you're doing something which is more customer-facing … that one time you took the wrong advice and you didn't check the backend data for it … you can lose your job.”


• GenAI is still growing: AI is predominantly around Robotic Process Automation and predictive AI, while GenAI is relatively small and exploratory by comparison, with even less use cases actually in production. In one Google workshop, the response to ‘Where are you in your GenAI adoption journey?’ was:
• We have only begun considering GenAI (39%)
• We are researching and consulting with experts to identify use cases (31%)
• We are piloting our first GenAI use case (24%)
• We have fully deployed at least one GenAI use case (7%)


• Use your own data: Banks need lots of data for their models to work. Using third party (external) data carries significant risk, especially in areas such as algorithmic bias. RBC CEO, Dave McKay, remarked that you should train your AI model to work on your own data sets. However, in order to train their Large Language Models banks often need more data than they have available in-house (and often the data is not AI-ready anyway). But external data carries its own set of risks. Using synthetic data can be helpful. But building models is expensive and banks should assess the cost carefully or it will outweigh the benefits. Banks are typically large, heavily regulated organizations with complex data structures, often due to a history of multiple acquisitions. You can’t build value if your data is inaccessible, poorly organized, low-quality, siloed, or unsecure.


• It’s still about people: Accenture suggested that GenAI could be applied to 60% of teller activities, 37% of customer service actions and 55% of credit analysis. Our latest research found that bankers see automation and efficiency as the key game-changers from generative AI. The top three use cases among board members were automation of customer onboarding (credit checks, KYC, etc.; 42%), monitoring fraud detection/auto-healing of security loopholes and automation of risk, regulation and compliance requests (both 41%). But GenAI will augment staff in their roles. Whether it is around code generation, real-time customer interaction or agent productivity, there is always a need for ‘humans in the loop’ (a common phrase at Sibos) for support and control. Microsoft’s Copilot service is designed to do just that.

I hope you had a good Sibos.

If you couldn’t make it, but would like to hear about what happened, we’d be happy to give you our perspective. If you’d like to talk further about what we’re finding out about GenAI, either through our research or delivering client projects, please get in touch.