Join top executives in San Francisco on July 11-12, to hear how leaders are integrating and optimizing AI investments for success. Learn More
The banking and financial services sector has typically lagged behind in AI adoption, but that’s starting to change. According to Insider Intelligence’s AI in Banking report, 80% of banks are highly aware of the potential benefits of AI. The aggregate potential cost savings for banks from AI applications has been estimated at $447 billion by 2023.
During a panel at this year’s Transform, Roey Mechrez, CTO and co-founder of BeyondMinds, spoke to Francesco Delle Fave, executive director in machine learning products at Goldman Sachs, Steve Flinter, VP of AI and machine learning at Mastercard Labs R&D, and Hui Wang, VP of data science, PayPal, about the major AI and machine learning news in the financial services industry.
In the old days there was often a 10- or 20-year lag in terms of wide adoption of a new technology, Wang said, but that gap is becoming smaller and smaller.
“The latest and greatest from academic research and the tech giants is being adopted in fintechs like PayPal within only one or two years,” Wang said. “We now have some of the most sophisticated real-time fraud detection models running on the most sophisticated deep learning architecture.”
Join us in San Francisco on July 11-12, where top executives will share how they have integrated and optimized AI investments for success and avoided common pitfalls.
Major AI trends in the financial industry
“AI and machine learning is being used in a broader array of applications than ever before, including security, fraud detection, identity, personalization, and even internal efficiency applications,” Flinter said.
There’s also been a shift away from a very model-centric view of the world into a more data-centric, data-driven view he said. Organizations are beginning to rethink how data is collected and used in the enterprise, and made AI- and ML-ready.
Another trend that’s growing among FIs is AI engineering, or MLOps, Flinter says. Organizations are shifting their focus to think about the full-life cycle of AI, from the R&D side of AI and machine learning through to production: developing, testing, and iterating on new AI applications.
The fourth and final trend is ethical AI or explainable AI. It’s the focus on how bias can potentially lead to unintended outcomes for consumers when deploying these financial models, particularly in lights-off or black-box scenarios.
“At PayPal, we’ve been using AI for fraud detection for almost as long as we’ve been here,” Wang. “But in the last few years, it’s on almost all fronts of our business in terms of empowering a better experience for our customers.”
In investment banking, the main objective is to serve clients, said Delle Fave. “As data becomes more available, as services become systematized, AI is permeating everything we do.”
AI and machine learning is enabling personalization, improving data analysis, allowing companies to be more rigorous in the way they interpret signals, information, and noise to develop more sophisticated trading strategies, and helping them design and build new products.
But there are challenges to massively adopting AI across an organization with many different use cases, Mechrez said, and surveys have shown many companies fail to shift new machine learning models to production. Data readiness is often the problem.
“We have come from an era of data being managed for transactional processing purposes, and then through data analytics, and now we’re having to rethink how we manage data for ML purposes,” Flinter said. “That means rethinking our data architectures, looking at things like data streaming and so on. A lot of the back-end infrastructure needs to be rethought and redesigned to enable ML at scale.”
On the product side of things, many banking and financial services companies lack real understanding as to what AI or ML can actually do.
“In some cases, it’s been my experience that some product teams think it’s flying rockets, flying cars, and it will solve every available problem and you can squeeze anything out of a small amount of data,” he said.
The solution is more education around what is truly possible and managing expectations as teams try to design AI and ML into product stacks.
“But even if we have the best model running live and generating great predictions, so what?” Wang said. “What do we do with them? If we can predict perfectly that a merchant is going to churn out tomorrow as an example, what do we do with that?”
That’s called the last-mile challenge in adopting AI: a prediction that doesn’t lead to action doesn’t do a company any good. How can a company go from the machine learning model to a product or experience that improves the customer’s life or helps the company?
At Goldman Sachs, the first step is discussing the problem with their client, so the client knows exactly what problem they’re trying to solve and how they’re applying AI to solve it, Delle Fave said.
“The objective is first to demonstrate value to the client,” he said. “Then, once this is done and the client understands, you start simple and incrementally build and add layers of complexity to potentially improve effectiveness, efficiency, or speed depending on the requirements we have, always interrupting constantly with the client so that they understand the way we’re building it.”
Your model doesn’t even need to be particularly sophisticated or fancy, Flinter said.
“The idea is you need to find the most simple and explainable thing that’s going to bring value to your customer,” he explained. “There are a lot of tools already available that allow you to do that.”
Human expertise isn’t going away
Sometimes you still need to pull a human into the loop to build the most effective AI solution, Wang said. PayPal uses a story-based product methodology, which is about the interaction between the human and the machine, embracing the interdependency of human and machine in an iterative process to get the best outcomes.
“I like to say that our AI solution is like ingredients we cook into a dish,” Wang said. “The algorithm can help us by explaining the number of ingredients we need to cook the best dish, but what if we’re missing the most critical ingredients? That’s where our human experts come in with their domain expertise.”
Fraud detection, one of the most important applications for AI in financial services, also requires a human in the loop, making decisions informed by data. In very low-latency, high-frequency types of decisions, humans are there to oversee the applications and the algorithms, to see if there are drifts in the models, if the models are picking up anomalies correctly.
“It’s that kind of long-term monitoring of models that we need to think about in security applications so we can ensure that they’re performing as we expect them to perform while making automated decisions,” Flinter said.
VentureBeat's mission is to be a digital town square for technical decision-makers to gain knowledge about transformative enterprise technology and transact. Discover our Briefings.