The gravest risk firms face when opting to use generative artificial intelligence tools is the governance (or lack thereof) in overseeing their usage, according to FINRA staff.
During a panel on AI at FINRA’s annual conference in Washington, D.C., this week, Andrew McElduff, a vice president and head of retail risk monitoring with FINRA, said firms are in danger if they poorly plan who will be responsible for AI projects, what their acceptable use policy is and where firms will set prohibitions on where it will be outlawed.
“What’s our appetite?” he said. “Do we need to get to a level of perfection on output, or where will we be happy to use this? And when do we have a human in the loop to verify? And then that kind of goes downstream into everything, down to your (written supervisory procedures) and your day-to-day operational flows.”
Additionally, McElduff said FINRA was still waiting for the Securities and Exchange Commission’s guidance on when recordkeeping requirements kick in for firms using generative AI (in other words, when does what the tool generates become a record that must be stored?).
McElduff acknowledged it was the “hands down, No. 1 frequently asked question” FINRA receives about gen AI-related risks. He said FINRA will try to get firms “one wholesome response” on the myriad recordkeeping-related questions when the SEC responds.
In the meantime, McElduff affirmed that the agency’s questions for firms about gen AI are “truly educational to understand the landscape,” with no “active exam environment” for the subject.
FINRA also unveiled the results of questions the organization asked registered firms about their use of gen AI last year, in which they received 3,107 total responses, a 93% response rate.
According to FINRA, firms used gen AI (or large language models) primarily for summarization and answering questions (with drafting responses and coding slightly behind). Microsoft was by far the most frequent vendor engaged by firms (though this included Azure, OpenAI, ChatGPT, Microsoft Co-Pilot, GitHub and Phi-2).
According to Joe Cordeira, UBS Wealth Management Americas’ chief data and analytics officer, when UBS began working with gen AI, it opted for “picking low-hanging fruit,” which in their case mirrored what FINRA firms are using it for: question answering and summarization.
UBS then rolled out Microsoft Co-Pilot to most of its advisors as its first large-scale gen AI program. According to Cordeira, advisors embraced it quickly, and having largely incorporated LLMs, UBS is now “starting the journey” to using agents (defined as autonomous AI tools or systems that can perform tasks without any human interaction).
“We have a few agents that are stood up now to do basic client tasks, agile change and changing preferences. Nothing crazy yet, but we have a clear vision about where we’re going with that,” he said. “And then the next frontier we think is in the client conversation between our financial advisors and our clients.”