The generative AI boom may be in its infancy, but financial institutions are already looking into how they can implement large language models to streamline workflows and increase productivity.

LLMs — AI systems trained using massive amounts of data to produce human-sounding responses — have a staggering number of potential applications in banking. Use cases range from helping mortgage brokers evaluate customer creditworthiness to helping those same customers track spending and make major financial decisions, according to Zor Gorelov, chief executive and co-founder of fintech Kasisto.

Kasisto, which uses AI to create chatbots and virtual assistants for financial institutions, announced Wednesday in a news release that it was piloting an LLM called KAI-GPT, which it advertises as the first created specifically for banking.  

KAI-GPT “basically was designed from scratch, and built and trained on financial services data,” Gorelov told Bank Automation News, adding that the LLM is “built around use cases.”  

While AI and chatbots have been used in the finance world for years, analysts predict that applications such as KAI-GPT that use ChatGPT-style generative AI could be major disruptors in the industry.

The first institution to adopt KAI-GPT is Australian banking giant Westpac, which is in the process of implementing KAI Answers, a generative AI application powered by the LLM, according to the release.

“Despite all the discussions about the future implications of generative AI in banking, Kasisto has made it a reality today,” said Westpac Chief Technical Officer David Walker in the release.

‘Artificial financial intelligence’

While major LLMs like Chat-GPT and Google’s Bard are aiming for artificial general intelligence with the ability to complete any task a human can, Gorelov says his goal is for KAI-GPT to achieve what he calls “artificial financial intelligence.” 

“It’s like having a bank or financial adviser on your shoulder that can provide … unbiased advice, recommendations, and help you with your financial well-being,” Gorelov said. 

Because the LLM is designed for finance, its interactions with users are limited to that context, which Gorelov says makes it more accurate than a general-purpose AI application. 

“If you ask KAI-GPT about [a] haircut, you get an answer about what it means to get a haircut if you’re an investor,” Gorelov said. 

According to Gorelov, KAI-GPT has the potential to reduce banks’ labor spend by at least 10% by automating contact centers, a savings which he says would be “staggering.” He predicted that the LLM would become “mature” enough to offer to consumers within 12 to 18 months. 

However, Moutusi Sau, an analyst at consulting firm Gartner, questioned whether that 10% savings would be significant enough to justify the disruption and startup costs associated with implementing the LLM. 

“Cost saving will typically happen if they can augment the entire workforce element,” Sau told BAN. “If you look at the … lending process in itself, if you could remove the whole area and replace it with ChatGPT, that would be really cost effective.” 

Despite recent advances in generative AI technology, Gorelov does not feel the technology is ready for wholesale automation of banking workflows. Of specific concern is the relative frequency of AI “hallucinations,” or unsupported conclusions that LLMs offer with a high degree of confidence.

Gorelov was quick to acknowledge these risks. The CEO said that for now, Kasisto is taking things slow and offering KAI-GPT only as a “labor augmentation” tool to help bankers.

LLMs on Wall Street

While Kasisto prepares to hit the market, financial information provider Bloomberg is applying the same technology to investing. 

In March, Bloomberg released a paper providing details on BloombergGPT — its proprietary, finance-focused LLM. 

Gideon Mann, who leads the machine learning product and research team for Bloomberg’s CTO, previously told BAN that the LLM would help the company improve existing products like the Bloomberg terminal and enable the creation of “really, truly transformative new products.”

These could include summaries of breaking news and complex financial documents, interactive question-and-answer functionality and an AI-driven search experience, Mann said.

Rather than releasing a standalone chatbot like Kasisto, Mann said Bloomberg plans to gradually incorporate its LLM into the core functionality of its products. 

“Some of [the products] will be startling and new and exciting, and others will just kind of make things better,” Mann said. “Unlike the press release, [which] was … a ‘big bang,’ I think the next phase of the product development is going to be incremental, slower releases.”