Zest AI formally debuted its large language lending intelligence bot LuLu in late February. The conversational AI assistant, which is kept separate from underwriting models as per regulatory requirements, is first trained using roughly 15 years' worth of customer queries recorded by the fintech as well as public sources of data such as National Credit Union Administration quarterly call report data and Home Mortgage Disclosure Act filings.

From there, LuLu is tailored to each institution, including the $3.7 billion-asset ORNL, through sets of business data on loan portfolios and applications, as well as internal reports and documents that are unique to each organization. Users conversing with the bot can ask questions about their institution's loan performance compared to others in a similar asset class, in addition to questions about how they can improve automation or fair lending compliance.

Vipperman said that she hopes to use LuLu in conjunction with Zest AI's underwriting models to "increase approvals across protected classes while not taking anything away from non protected classes" and continually check in on "what would have happened if we made different decisions" while asking "could we have brought more consumers in and grown even more with lower risk," amid other questions.

"The reason that we exist as a not-for-profit cooperative, is that our intention is to serve the underserved and what better way to serve the underserved than to be able to [use] LuLu … and figure out what can we do differently to bring everybody in and then still do it in a safe and sound way," Vipperman said. The credit union's iteration of the gen AI tool is set to go live this month.

Use of gen AI tools is growing across the financial services space. A survey released last month by Arizent, which publishes American Banker, found that roughly 55% of global and national banks with more than $100 billion of assets are implementing generative AI in some capacity. Credit unions and regional banks with assets between $10 billion and $100 billion recorded 40% implementation, and community banks with less than $10 billion of assets responded with 28%.

More specific use cases involve Citi's rollout of the GitHub Copilot to developers and the $733 million-asset Grasshopper Bank in New York, which instituted an AI-based assistant for its compliance team handling tasks necessary under the Bank Secrecy Act. Credit Karma, which was acquired by Intuit in 2020, implemented its financial assistant earlier this year.

Jerry Haywood, CEO of the Sandnes, Norway-based conversational AI provider boost.ai, said customer experience, marketing and customer assessment for credit-based decisions are the three key areas where gen AI is being tested, but understanding how to apply it in individual use cases means knowing how much involvement is needed.

"While gen AI is the newest tech on the block, there are still many use cases where traditional, pre-written flows are the right tool for the job, and may even be a more practical solution. … For example, any process that needs to be 100% the same in every case, such as the transfer of funds between accounts, should be handled by a pre-written flow," Haywood said. The fintech debuted its newest iteration of AI-powered assistants earlier this week.

Not all financial institutions are keen on rushing to adopt new technologies, however.

Roughly 15% of respondents to the aforementioned Arizent research have prohibited their employees from using any form of gen AI for work-related tasks, while a further 46% either restrict its use to specific functions and roles or are considering putting limiting policies in place. Many hold back due to concerns that technology that can produce new content can have unforeseen results.

"Unlike deterministic tools, generative AI produces outputs that aren't always foreseeable," said Lei Wang, chief technology officer of Torpago, a card and spend management fintech. "This lack of control over the output becomes especially concerning when these tools are directly interfacing with end-users."

Thorough testing is important when developing and implementing these models to minimize the possibility of hallucinations — the creation of false information or results — and biases unintentionally included in the training data, said Jay Venkateswaran, business unit head of banking and financial services for the Mumbai, Maharashtra-based global WNS.

Regulatory concerns are also a worry. Following the White House's executive order on AI released last November, developers of AI models like Zest and the financial institutions they partner with have been cautiously moving ahead when implementing products such as underwriting algorithms, conversational bots, employee co-pilots and more — all to avoid any potential missteps with regulators.

Banking officials with the Federal Deposit Insurance Corp. that are exploring the risks of overreliance on AI maintain that existing laws and tools are capable of preventing any vulnerabilities from impacting consumers or the financial system at large. But others with the Consumer Financial Protection Bureau, which has continued its campaign to root out instances of bias in algorithmic-based lending and other transparency issues, remain skeptical.

Another hurdle to gen AI adoption in the  banking industry is the fear among entry-level employees that AI will take over their tasks, and thus render their roles redundant. Executives are working to assuage these doubts by including staffers who would be most affected by the addition of AI tools in the testing and rollout of any new products.

There is still work to be done where end users are concerned, as institutions "are understandably being prudent while savvy fintechs are fast at work to roll out customer-facing generative AI tools," said Dylan Lerner, senior digital banking analyst at Javelin Strategy & Research.

"The last thing financial institutions need right now is a misunderstood element embedded in their tech stack," Lerner said.