Risk Lens back in View

First, my general sense is that while many large financial firms clearly are advancing their experimentation as well as collaborative efforts with those advancing AI techniques, esp. in the LLM arena fastest, these same firms are identifying new areas of “risk” and thus pausing some of their early enthusiasm around rapidly exposing their innovative efforts to clients.

This does not mean to say that the industry as a whole doesn’t remain very keen to develop AI use cases that involve launching virtual assistance, but while perhaps the initial outcomes of “co-pilot” applications were geared toward servicing client needs, and scaling accessibility through AI, more effort is now shifting toward internal initiatives that improve operational efficiency, promise enhanced productivity, and condense decision making effort.

This shift appears to be the result of three developments that were highlighted during the day. First, despite the advancement in what AI models can potentially deliver through simplified interfaces, they can’t immediately address challenges in relation to data quality, data transformation, and data accessibility controlled by older legacy technologies. Second, even when experiments can be ring fenced to mitigate some of these issues that undermine effectiveness, there are still a variety of risk factors related to explainability, result consistency, and overall performance that need to become a part of the overall application design, esp. as the regulatory environment is putting the industry on the alert around what self-governance as a framework, really means, in practical real world terms. Finally, while the “FOMO” effect still seems to be in the air when it comes to AI initiatives to both, please shareholders, as well as retain and grow profitability and operating leverage, a lot more internal stakeholders are asking for better evidence to be provided in relation to the ROI on investment, esp. if it ultimately going to create a major upheaval in the investment in the data fabric that permeates through almost all business processes.

Use Case Duality

Second, we are seeing a bit of bifurcation in use case focus emerging, with innovation leaders with an operational background tending to look toward adopting an “AI first” set of tests to further advance hyper-automation, while those that have taken on leadership from a marketing and client experience perspective, more interested in those aspects of first principles for generative AI that can tackle challenges that are faced by individuals who have to deal with consumption of large complex, mostly unstructured data sets.

While the objectives banks are looking for from these areas of focus might look different, they both underscore that many private and open source initiatives have been more successful, as data sets have increased, and machine learning techniques advanced, at tackling weaknesses of previous attempts to manage value generation from unstructured data at scale, improving both the accuracy of classification, entity identify and labelling, as well as the scope, i.e., business rules (both in absolute and conditional terms) that can be identified, transformed and processed without direct human intervention.

For Hyper automation projects, banks not only see this as improving STP across many different stages of the lifecycle value chain, but as importantly, enabling human to both become supervisors as well as exception handlers for complex, non-standard situations, more efficiently. Similarly, those with an eye on deploying AI solutions using advanced techniques involving NLP and Generative AI, are seeking ways to re-build systematic processes across different operational functions so that they can reduce the need and costs and time spent related to error handling and document reiteration.

Establishing an AI First Framework

Third, and perhaps, most tellingly, it appears that organizations are making progress, often leaning on their strategic partners own capabilities, to develop and distribute throughout their businesses “AI First” principles. When someone like myself who has had exposure to financial institutions tinkering with AI and ML since the mid 90’s hears language like this, it seems to suggest that, despite clear areas of immaturity, and obvious and serious gaps in relation to the capability of AI models to transfer inference (strong) into context (weak), that the application of AI, as a foundational technology for the development and improvement of the next generation of business process management has arrived.

While one might be tempted to think that this will quickly lead to an entirely new way of thinking through business process design, the feeling I got was more mixed than this. Certainly, some of the “futurists” firms/thought leaders talk toward a clear future where hyper personalization could be really possible in the realm of product design, fulfilment and service support (accompanied by a wide range of pricing models), but there are just as many people who truly think that in the near term, the primary benefits is going to be drawn by applying AI capabilities toward reducing both the operational drag and support risk that go with having to operate and integrate legacy systems with digital platforms. This is because middleware solutions have more often than not been blunt instruments to address deficiency, while transformation tools have often fallen short in solving performance issues which are putting the bank at risk.

Given this, being able to build smarter, and continuously optimizing middleware solutions through incorporating AI models into process transformation might prove to deliver far higher ROI to running the bank than those that are focused on enhancing the value and capabilities of the digital platforms. Many bankers don’t often like to talk in this language, because it is perceived as “less sexy” in the public domain, but when organizations can’t grow their way to improving their income/cost ratios and would rather be retooling their workforce (for all kinds of reasons), rather than reducing them, having data science and AI software engineers working with operational and product domain experts in this way starts to make a lot of sense.