Banking and finance companies hop on the generative AI bandwagon

The banking and finance industries, heavily regulated though they may be, are clambering onto the generative AI bandwagon.

Morgan Stanley said it is working with OpenAI to develop a tool that offers insights to its financial advisors; JPMorgan Chase recently submitted a trademark application for an AI-based program called IndexGPT focused on analyzing and recommending investments to customers; and Goldman Sachs is currently testing large language models (LLMs) for internal use.

They aren’t the only ones. A recent Gartner report found that while nearly half of banking execs surveyed in February said they had no plans to incorporate generative AI into their business, that portion had plummeted to 7% by April. Around 17% said in April that they were already implementing the tech, up from 7% in February.

Banking execs and experts we spoke to said the appeal is clear: Finance is a data-intensive industry where any tech that can summarize and synthesize massive amounts of information has the potential to be a game-changer. But it’s also a space where mistakes and inaccuracies can be especially high-stakes; that plethora of data is heavily regulated, presenting challenges to implementation.

Moutusi Sau, a VP analyst at Gartner who focuses on banking and AI, said that in her seven years covering AI for the research firm, this tech hype wave stands out.

“The interest level is not dropping down. Like, every single inquiry I have is on this topic,” Sau said. “This time, it’s real. It’s happening.”

Summing up documents

At Goldman Sachs, Chief Information Officer Marco Argenti said the bank is exploring generative AI for a variety of purposes that are mostly tied to internal workflows. So far, much of the focus has been on augmenting the coding capabilities of developers working at the company.

“We are at various stages of experimentation, but definitely the whole developer productivity area is a big focus,” Argenti told Tech Brew. “We have active proof of concept there, where we essentially augment the capability of our developers.”

Argenti said that the company is also exploring how AI summarization trained on the many types of documents its workforce handles can help extract information from earnings reports, company statements, contracts, and regulatory guidance.

“Imagine this input-output of documents and within that there is a mixture of humans and machines,” Argenti said. “This new generation of AI is good at, broadly, three things. One is summarizing very vast amounts of information, the other one is filling in the blanks—so when there is incomplete information, actually kind of trying to predict what is in those gaps…and then the other is connecting the dots.”

For instance, Argenti said AI might eventually be able to summarize a company’s financial statements and check whether it is in compliance with a loan covenant.

“It’s an example of an activity that is largely manual, and an AI that is properly trained could pinpoint things like, ‘OK, your cash flow is not necessarily where it needs to be,’” Argenti said.

Regulation risk

Many of these use cases, however, are contingent on the AI not hallucinating, the technical term for fabricating information, which is especially important for a business like a bank.

“Obviously, accuracy, explainability, precision of our output at our scale is really what we are about,” Argenti said. “We’re a bank; it’s all about trust, it’s all about the right information.”

“[The regulatory aspect] is one piece that they need to figure out,” Sau said. “On the AI front, the adoption level was always a little bit lower, because they had to deal with regulations. And they cannot use customer information directly in any of the machine learning models.”

At Capital One, Prem Natarajan, the bank’s chief scientist and head of enterprise data and AI, said that any implementation of generative AI should be done in a “responsible, thoughtful” way.

“There are lots of sectors where the burden of precision is higher than social chitchat,” Natarajan said. “And finance happens to be one of them.”

AI for customer service

Natarajan said Capital One is in the “design and internal-experimentation phase” with LLMs, but he wants to avoid rolling out headline-grabbing demos of the tech that aren’t able to “deliver safe, accurate information.” He said the earliest use cases for the bank will likely be around customer experiences.

“Everywhere where there are customer interactions, everywhere where you have to consume large amounts of data and get to the specific things that you want in the data quickly, this is going to be super useful,” Natarajan said.

The current LLM frenzy is not the first time that banks have explored AI for customer service issues. Capital One previously used natural language processing to automate customer service chats and better understand their inputs and concerns.

Natarajan said the complexity of customer interactions in the banking realm has made AI particularly appealing.

“What puts banking in that special place is the richness of customer interactions,” Natarajan said. “Whether they’re buying stuff with their cards, or whether they’re depositing money or pulling out money in stuff, etc. There’s just a lot of interactions that we’re doing every day with all of the financial institutions in our lives that provide a very rich interaction basis.”

While customer service interactions might be one of the first use cases for generative AI at Capital One, Natarajan eventually sees AI changing the way the company does business, though he’s light on specifics for now.

“While I said customer service, in general, and customer interactions are where I think early things will be, it’s hard to imagine an area for our operation that’s not going to be influenced,” he said.

Correction 06/21/23: Language in the penultimate paragraph has been updated to clarify that Natarajan was speaking about AI generally when it comes to the future of Capital One's business.