Recent advances in generative artificial intelligence have been remarkable. ChatGPT and similar software have understandably captured the public's imagination. Many consultancies think that banking will be among the sectors that will benefit most from its development. For example,
But there is a long and uncertain journey for technologies from scientific invention to commercial application. While generative AI may transform banking as the technology matures, there are three key reasons to doubt the most bullish claims that the sector is especially well-positioned to profit from generative AI quickly or easily.
The first is that, so far, deep learning has had a shallow impact on banking.
Generative AI is fundamentally a type of deep learning, the machine learning method at the heart of advances in AI over the past decade. Deep learning has been most successful when applied to 'unstructured data' such as text and images. Unsurprisingly then, the most prominent instances of generative AI have been large language models like GPT and deepfake videos.
However, the most vital information banks have is numerical data on customers and products.
Another reason deep learning models have not gained traction in banks is that the way they work is difficult to audit. Generative AI, like GPT, with billions of parameters trained on poorly vetted data, is likely to face intense scrutiny from model risk management functions inside banks, as well as external regulators. Some banks may find that the compliance risks from generative AI outweigh the potential business benefits in the near term.
The second reason to be skeptical about generative AI transforming banking is that the fundamental drivers of bank profitability haven't changed.
Generative AI is frequently described as revolutionary. Yet, however much technology changes, the ways in which banks earn their profits will remain the same. These permanent profit drivers flow from the structure of banks' income statements. Higher profits can come from higher net interest income or higher fees and trading income, or by lowering operating costs, provisions or taxes paid.
At a Senate Banking subcommittee meeting, Republican and Democratic lawmakers both promoted the mission of community development financial institutions and warned of upcoming threats to their funding and proposals to revamp the CDFI certification process.
Generative AI will not impact most of these profit drivers. The macroeconomy drives banks' net interest income, while taxes paid are determined by governments. Since banks today operate in a highly competitive globalized economy, they have limited ability to charge higher fees or earn more trading income sustainably. More precise risk models might lead to lower provisions. But as mentioned, these models are likely to use other machine learning methods besides generative AI.
The main way generative AI might boost bank profits is by lowering operational costs. The greatest opportunity to do so is probably in marketing. Generative AI could dramatically reduce the costs of creating advertising content while also improving its quality. Other applications include handling customer requests, creating documents and
Finally, most banks don't (and won't) have a comparative advantage in AI.
It is sometimes said that there is "no AI without the right IA," where IA means information architecture, but, equally, information availability. Many banks are weak on both fronts. Banks' data is often dispersed across multiple warehouses, with discontinuities in their definition over time, making the data hard to retrieve and stitch together. Information governance policies and workplace politics often reinforce these silos so that access to data in banks is highly restricted. As a result, bankers often suffer the paradox of having a poverty of data amid plenty.
Without sufficient computational power to train models or easy access to data for fine-tuning them, most banks will find they have no comparative advantage in generative AI. Instead, they will consume what Big Tech produces. This is a rational strategy for small banks given that most lack the talent or technology to produce generative AI themselves. It is also a rational strategy for many large banks, allowing them to avoid the execution risks associated with developing their own capabilities while mitigating the reputational risks of generative AI producing offensive or inaccurate content through third-party warranties. But buying instead of "building" generative AI also limits the opportunities for competitive differentiation. Instead, for most banks, generative AI means procuring a commoditized service.
So, what's a bank to do?
Generative AI is an exciting set of technologies. Large language models such as GPT already make fantastic ideational assistants that offer stimulating suggestions to their users. In the future, these AIs may mature to the point that they can inspire banks' strategic thinking around new asset classes and business models.
But for most banks, the most prudent approach will be to pace their adoption of generative AI while the technology is still evolving. There is a well-known saying among software developers that "premature optimization is the root of all evil," and nowhere does that hold truer today than with generative AI.
