AI agents, Beer says, will change the way one thinks about work, the tasks to complete that work, how to break those tasks down, the tasks the bank is comfortable automating, the tasks that require human reflection, and then the proper technology ecosystem with the proper security, resiliency, and controls.
“We have been focusing very early on, on simple things, like what’s the right level to create an agent; how do you give them identity and access?” says Beer.
Her approach is fairly flexible. In HR, a human has broader license to see JPMorgan employee data than an agent. “You don’t want them to go outside the bounds of the specific tasks that they can do, because they don’t have the same thinking a human does,” says Beer. But in software engineering, there’s a bit more pliability with the permissions granted to an AI agent, because there’s a validation layer to check and correct any errors those autonomous systems could generate.
That same monitoring layer will need to be applied to other parts of the business as agents are increasingly embraced, says Beer, which involves keeping the human in the loop but also monitoring the outputs that large language models are producing.
One clear certainty when it comes to JPMorgan’s agentic AI strategy is that these tools won’t run through a third-party vendor. “This is going to be critical, because it’s the underlying flow of how we do business,” she says. “We want to secure it and we want to make sure it’s organized.”
“You’re moving $12 trillion a day, and you have a lot of customers and clients,” says Beer. “So this balance between innovation and risk taking is critical for us to get it right. We spend a lot of time focusing on that.”
AI coding tools, meanwhile, have vastly improved and led Beer to work on rebuilding what she calls “the factory,” which encompasses efforts to rethink how product teams and engineers build. That means less time working in integrated development environments, or IDEs, which are the software applications that allow programmers to write, test, and debug code in one interface; and more time giving AI models the context they need to handle complex tasks.
“We’ve had great examples of some of our deep architects that are really great as specifications, but didn’t like to code so much,” says Beer. “Now, they’re able to spend the time up front creating the specification. And we need senior engineers reviewing the code output.”
Meanwhile, an internal team led by Teresa Heitsenrether, chief data and analytics officer, and Robin Leopold, head of HR, is working with other leaders to speadhead the effort to more thoroughly re-engineer workflows across JPMorgan. The bank is also working with an unnamed, large academic university for this ongoing project. This coincides with “several hundred” AI use cases already in production today, as well as future tech initiatives that the company’s operating committee—which includes Beer, Leopold, and Heitsenrether—regularly monitors to create “hard value creation.” That value has both potential to generate top-line revenue and productivity gains.
“I think the change management and how you think about the ways of working is ultimately the hardest part here, and reimagining how you can use these tools,” says Beer.
As AI models mature, JPMorgan has had to think of when to play offense versus defense, and in both cases, AI is a tool that can help. “The tools are getting better at helping you find vulnerabilities,” says Beer. “We have to make sure that we’re better at fixing them faster.”
John Kell
