- JPMorgan hired Manuela Veloso in 2018 to explore the potential of AI in finance.
- Veloso is on leave from her role as head of Carnegie Mellon’s machine learning department.
- She told Insider about the seven principles that drive her team’s research.
Incumbent financial institutions often have notoriously slow, clunky technology — but that’s exactly what lured Manuela Veloso, Carnegie Mellon University’s head of machine learning, to JPMorgan Chase three years ago.
“JPMorgan has been around for a long time. Unlike some technology companies like Amazon or Google, JPMorgan wasn’t born digital and that was one of the challenges that attracted me: How do you bring AI to these non-digital services?” Veloso told Insider.
Veloso is the head of AI research at JPMorgan, currently on leave from her role at the university. Her work, unlike many tech teams on Wall Street, is not driven by business needs. Instead, Veloso mulls the infinite what-ifs.
What if Chase gained 2 million online-only users tomorrow? What if the
aggressively hiked the interest rate this week? And what are the ripple effects of those what-ifs?
Exploration is key to Veloso’s research. It’s one of her goals as JPM’s lead AI researcher to “make people embrace, grow their hearts, for the use of AI,” even if it doesn’t necessarily lead to tech deployment or a business tool.
Veloso’s at the helm of a team of nearly 60 academics whose expertise span mathematics, cryptography, electrical engineering, and machine learning, to name a few.
She outlined the seven main challenges her team is trying to solve with AI.
JPMorgan is researching how AI can help eradicate financial crime, which is one of the most frustrating problems Veloso said she has encountered since joining the bank.
“I didn’t anticipate financial crime would be such a major focus area for my group,” she said.
Identity theft will cost US firms $721 billion in 2021, up 42% from 2019’s $502 billion loss, according to a March report from Aite Group, a financial research and advisory firm. In that timeframe, nearly half (47%) of American respondents reported being victims of some kind of identity theft, like account takeover or application fraud.
One way to fight financial crime is exploring how data and AI can improve fraud and anomaly detection, she said. The bank is working to expand its detection beyond single events and instead train AI to identify behavioral patterns, or sequences of events, that are indicative of financial crime. Graphs are used to model these behavioral networks, she added.
Large economic systems
Veloso’s team uses AI to predict the impact of interactions between the various players within large economic systems, often referred to as multi-agent systems. Examples include supply chains or when multiple parties are involved in a trade or exchange.
A key part of this — and an underplayed application that people have yet to take advantage of — is the use of AI-powered simulations to understand and test the ripple effects of those interactions, Veloso said. But to understand that, researchers have to go beyond historical data, she added.
Real data, while valuable, doesn’t necessarily allow the user to expand into different possibilities. Veloso likened deploying AI to historical data to a copy machine. Whatever happens can be reproduced, “but it doesn’t let you go beyond the cycle. Simulations enable you to explore a much larger space than what the real data tells you,” she said.
Despite the limitations historical data can have when deploying AI, data is still embedded into JPMorgan’s framework. Veloso’s team is exploring how to use AI for data management, like sharing data safely between multiple divisions within a bank or externally, she said.
The bank is using AI to fetch public information and help with the controls of who has access and sharing permissions, Veloso said. There’s a lot of emphasis on the infrastructure, specifically about having that data available on the cloud, she added.
The team is also deepening its understanding of how cryptography and synthetic data can help with data privacy and protection, she added.
Another main goal of the research team is figuring out how AI can improve how employees work.
That includes using internal chatbots to help call center agents siphon through bank resources, using AI methods to classify emails, and automating the generation of reports and PowerPoints, Veloso added.
Veloso’s work is also focused on the customer experience. That includes how quickly employees access data for customers, and using data to streamline the customer experience.
For example, information required for know-your-customer regulations often already exists on JPMorgan documents or publicly-available ones. However, mapping the existing information to a given data field is a manual process, Veloso said.
Now, JPMorgan is automating that with natural-language processing and presenting the customer with a form pre-filled by AI that they just have to check is correct, she said.
Regulation and compliance
As a global bank, JPMorgan has thousands of obligations and rules it must understand, and know which regulations apply to which jurisdictions and businesses.
If a banker talks on the phone with a client from Singapore, or they write a contract with a client in Italy, understanding what laws govern what they do is critical.
It’s a complex system to navigate, Veloso said. The bank is analyzing how natural-language processing can help and is spending time to understand how human-AI interactions are evolving, she said.
Corporate and industry values
All of this work must be done in a way that strengthens the bank’s and industry’s values of explainability, trust, fairness and social good, while keeping a reign on the powerful tech.
Much of this work is centralized in the bank’s Explainable AI Center of Excellence. Created in July 2020 and led by the AI research team, the group shares techniques and tools that support understanding how the tech works and ensuring it is used fairly.