JPMorgan Chase Launches AI Assistant Based on ChatGPT Maker OpenAI

JPMorgan Chase has rolled out a generative AI-powered assistant to tens of thousands of its employees in recent weeks, the first part of a broader plan to spread the technology across the sprawling financial giant.

The program, called LLM Suite, is already available to more than 60,000 employees, helping them with tasks like writing emails and reports. The software is expected to eventually become as ubiquitous within the bank as the videoconferencing program Zoom, people familiar with the plans told Vscek.

Rather than developing its own AI models, JPMorgan designed the LLM Suite as a portal that lets users tap into large external language models — the complex programs that power generative AI tools — and launched it with LLM from OpenAI, the maker of ChatGPT, the people said.

“Ultimately, we want to be able to move pretty seamlessly between models based on use cases,” Teresa Heitsenrether, JPMorgan’s head of data and analytics, said in an interview. “The plan is to not be locked into a single model provider.”

The move by JPMorgan, the largest U.S. bank by assets, shows how quickly generative AI has taken American businesses by storm since ChatGPT arrived in late 2022. Rival Bank Morgan Stanley has already released a couple of OpenAI-based tools for its financial advisors. And the consumer tech giant Apple said in June that it plans to integrate OpenAI models into the operating systems of hundreds of millions of consumer devices, significantly expanding its reach.

The technology, hailed by some as the “cognitive revolution” in which tasks previously performed by knowledge workers will be automated, could be as important as the advent of electricity, the printing press and the Internet, JPMorgan CEO Jamie Dimon said in April.

It will likely “grow virtually every job” at the bank, Dimon said. JPMorgan had about 313,000 employees as of June.

The bank is giving employees what is essentially OpenAI’s ChatGPT in a JPMorgan-approved wrapper more than a year after it blocked employees from using ChatGPT. That’s because JPMorgan didn’t want to expose its data to outside providers, Heitsenrether said.

“Since our data is a key differentiator, we don’t want it to be used to train the model,” he said. “We implemented it so we can leverage the model while still keeping our data secure.”

The bank has rolled out the LLM Suite across the company, with groups using it in JPMorgan’s consumer division, investment bank and asset and wealth management businesses, the people said. It can help employees write, summarize long documents, solve problems using Excel and generate ideas.

But bringing it to employee desktops is just the first step, according to Heitsenrether, who was promoted in 2023 to lead the bank’s adoption of this disruptive technology.

“You have to teach people how to do prompt engineering that’s relevant to their domain to show them what it can actually do,” Heitsenrether said. “The more people dig into it and figure out what it’s good at and what it’s not good at, the more we start to see ideas really blossom.”

The bank’s engineers can also use LLM Suite to embed functions from external AI models directly into their programs, he said.

JPMorgan has been working on traditional AI and machine learning for over a decade, but the arrival of ChatGPT has forced it to change course.

Traditional, or narrow, AI performs specific tasks that involve pattern recognition, such as making predictions based on historical data. Generative AI is more advanced, however, and trains models on large data sets with the goal of creating patterns, which is how human-sounding text or realistic images are formed.

Heitsenrether said the number of uses for generative AI is “exponentially greater” than previous technology, thanks to the flexibility of LLMs.

The bank is testing many cases for both forms of AI and has already put some into production.

JPMorgan is using generative AI to create marketing content for social media channels, map itineraries for clients of the travel agency it acquired in 2022 and summarize meetings for financial advisers, it said.

The consumer bank is using artificial intelligence to determine where to place new branches and ATMs, capturing satellite imagery and in call centers to help support staff find answers quickly, Heitsenrether said.

In the company’s global payments business, which moves more than $8 trillion worldwide every day, AI helps prevent hundreds of millions of dollars in fraud, he said.

But the bank is being more cautious with generative AI that interacts directly with individual customers, because of the risk of a chatbot providing incorrect information, Heitsenrether said.

Ultimately, the field of generative AI could develop into “five or six big foundational models” that dominate the market, he said.

The bank is testing LLMs from US tech giants and open-source models to integrate into its portal, the sources, who spoke on condition of anonymity, said about the bank’s artificial intelligence strategy.

Heitsenrether outlined three phases for the evolution of generative AI at JPMorgan.

The first is simply making models available to workers; the second involves adding proprietary JPMorgan data to help increase employee productivity, which is just beginning at the company.

The third is a bigger leap that would unlock much greater productivity gains, when generative AI is powerful enough to act as autonomous agents performing complex multi-step tasks. This would make rank-and-file workers more like managers with AI assistants at their command.

Technology will likely empower some workers while displacing others, changing the composition of the industry in ways that are difficult to predict.

Banking jobs are the most prone to automation of all industries, including technology, healthcare and retail, according to consulting firm Accenture. AI could boost the industry’s profits by $170 billion in just four years, Citizen Group analysts said.

People should think of generative AI “as an assistant that does the most mundane things that we all wish we didn’t do, and can just tell you the answer without having to scroll through spreadsheets,” Heitsenrether said.

“You can focus on the higher-value work,” he said.

— Vscek’s Leslie Picker contributed to this report.