Generative AI Has Made Banking Cheaper - But Has It Made It Better?

Since its origins in the 1950s, artificial intelligence (AI) has been marked by periods of intense optimism, interlaced with eras of rank pessimism.

The past decade, which saw the rise of computers capable of understanding and generating natural language, has perhaps been the most consequential boom period thus far. But the staying power of generative AI is now a hot question, with many suggesting that the advances have been far less impactful than initially promised.

Tony Hughes

The new tools have been a major focus of the banking industry. Just how big, though, are the gains in our sector, and will the worm soon turn in the direction of AI skepticism?

The main concern for ChatGPT and other generative AI systems is that they can sometimes  provide disappointing and confusing results – known as hallucinations. Despite ongoing research on the matter, the hallucinations dilemma was recently described as the “3% problem that no-one can fix.” 

On the other hand, there are also indications that generative AI has great potential to have a powerful effect on the productivity of bankers. For example,  a recent report from Citi Global Perspectives and Solutions (Citi GPS), a Citi thought-leadership unit, estimated that 54% of banking jobs could be displaced by AI – but perhaps they are simply leaning into the hype regarding the perceived benefits of the new technology?

I suspect that most of the jobs at risk will be low-level analyst positions and roles geared toward regulatory compliance rather than earning profits. In the old days, teams of bankers would pour over corporate financial statements and public filings looking for signs of trouble. But today it is conceivable that a single analyst aided by ChatGPT could replace 10 erstwhile colleagues engaged in such a menial role.

Similarly, banks are required to fulfill many regulatory tasks that generate a cost but yield little upside to the bottom line. If generative AI helps banks comply with the rules more economically, the payoff to shareholders will be significant.

Last year, I mused about the possibility that AI could help to overcome data deficiencies in certain products and perhaps even expand access to banking services to underserved populations. There is, moreover, real potential to aid fraud detection, the nature of which does not lend itself to structural modeling.

However, the extent to which all this potential is being realized is quite difficult to discern in the data. You would expect banks reaping significant improvements in their lending operations to be enjoying increased interest margins, but the aggregate U.S. net interest margin (NIM) has been falling since the mid-1990s and continued to decline through the end of 2023 at most big banks.

In short, if generative AI is to truly transform banking, it needs to do more than just displace a lot of lower-value jobs, automate tedious regulatory tasks and save banks a bit of money.

Generative AI: Cross-Industry Evolution

In my own initial foray into using ChatGPT for retail and corporate credit risk assessment, I reported that the technology was impressive but equivocal. We can also learn about the impact of generative AI by exploring commercial use cases in other industries.

In a Substack article published in March, Gary Marcus, one of the early AI industry skeptics, pointed out that commercial use cases for the new model have been very limited. Apart from its ability to generate somewhat flaky text, the tool is a boon to coders – which I recently discovered when I needed to rekindle my long dormant programming skills for a “big data” statistical exploration I wanted to pursue. Based on my experience, you must be very vigilant during the debugging phase, but there’s no doubt that generative AI is a major time saver, even an enabler for those who have let their coding skills slip in their dotage.

In its early days, generative AI also made it mark in writing and academics. There were reports of articles written (in-part) by ChatGPT appearing in peer-reviewed journals. If they are so inclined, undergraduates, it seems, now have a fantastic new way to cheat – and in a manner that is almost undetectable to their overworked lecturers.

But more efficient coding and helping undergrads cheat is not enough to justify the massive valuation of stocks related to AI. With nagging doubts about the accuracy of the output and lacking a “killer app” to transform our lives, there are real fears that the massive rush of funding will fail to yield the necessary returns.

As Marcus points out, unless ChatGPT version 5 comes out soon and is capable of matching the hype in the industry, the bubble may burst and a new trough of disappointment will be, in his view, inevitable.

Parting Thoughts

So, is the impact of generative AI being overhyped? Probably. It is certainly true that the promised payoffs to banks are now hard to perceive in real-world data. Banks are not immune from the disease that Marcus observed for the economy as a whole. 

I suspect that AI will soon either peter out or go boom! Like Marcus says, it probably depends on the success (or otherwise) of the next big breakthrough.

If the AI bubble does burst, the ironic marker of the era will be the failure of Silicon Valley Bank.  At the height of an unprecedented wave of AI optimism, the bank closest to the epicenter was that which failed.

I don’t want to spoil the hype, but might that be an omen? Or will the AI boom continue in the banking industry? Only time will tell.

Tony Hughes is an expert risk modeler. He has more than 20 years of experience as a senior risk professional in North America, Europe and Australia, specializing in model risk management, model build/validation and quantitative climate risk solutions. He writes regularly on climate-related risk management issues at UnpackingClimateRisk.com.