At this point,
AmEx's approach is representative of financial services companies that are entering the world of generative AI but are hesitant to incorporate customer data into their usage.
Bhavi Mehta, Global Head of Financial Services AI at Bain, found that the majority of financial services companies use OpenAI's closed-source generative AI models.
Amex has been leveraging AI and machine learning since 2010, starting with credit risk analysis and fraud detection. But Anre Williams, CEO of American Express National Bank and group president of American Express' Enterprise Services, is using generative AI to deliver better customer experiences. He sees “huge potential” in delving into specific recommendations and strengthening customer support.
“I've always been fascinated by the amount of information that AI can take in and give back answers in a concise, conversational format,” Williams says. “The speed with which we can accomplish that is groundbreaking.”
To prepare for this new wave of technology, Amex convened a Generative AI Council a year ago to prioritize use cases, manage risk, and ensure responsible implementation. The company sourced senior leaders from model building, technology, services, legal, and other departments to guide its approach to generative AI.
500 use cases surfaced.
“There were also learning opportunities,” Williams said. “Some products have great potential and we've seen them scale.”
One pilot investigated the potential of generative AI co-pilots to improve recommendations for Platinum and Centurion cardholders at Amex travel agencies. For example, if a customer calls and asks for a pet-friendly hotel in a desired area of a particular city, “it's not always easy to find,” Williams says. Creating an accurate list from the Internet can be time consuming and may require agents to call customers back later or email offers.
Amex agents handle such queries
Another use case involved software engineers at Amex. Amex tested a developer-only CoPilot for developers seeking support with code generation. This tool includes knowledge of how code is written in general and the context of what software engineers have open on their desktops.
“We believe this will enhance their work and allow them to spend more time on complex tasks,” Williams said. The results so far bear that out. The engineer reported that using this co-pilot saved him 10% in total work time. Early results also suggest improved satisfaction. In its 2023 Developer Sentiment Survey, Amex found that 60% of the developers who participated in the survey found Amex's ecosystem of technologies, the tools, applications, and platforms they have access to, to be intuitive and helpful. I discovered that. A recent study measuring satisfaction with this co-pilot found that 85% gave it a high rating for satisfaction.
All 6,000 software engineers at Amex will have access by the end of June.
Using CoPilot to help write code is “huge,” Mehta said, in terms of unlocking productivity for organizations with large technical teams.For example, Shadman Zafar of Citi
So far, Amex's early experiments are using commercial large-scale language models, open-source models, and some models embedded directly into other commercial products. The company is open to switching providers if a preferred model emerges in the coming months. There are no plans to build your own LLM at this time.
Closed-source models tend to perform better and require less infrastructure to maintain, making them the quickest way to power generative AI, Mehta said. Open source gives organizations more control over model architecture and data flow.
“But ultimately I don't think that will be the case.” [approach]“Organizations have both, depending on what they are working on and the capabilities they need,” Mehta said.
AmEx also continues to keep its employees informed with AI-driven generative recommendations, rather than providing them directly to customers.
“We are not using generative AI to make credit or approval decisions,” Williams said.
In the travel agency pilot, the counselor verified the information returned by the co-pilot.
“At this point, we're not necessarily validating the answers, because we've done enough validation in the early answers to feel confident that the answers are accurate,” Williams said.
He also pointed out that AmEx's software engineers review each line of code and adopt the co-pilot's suggestions only one-quarter of the time, and three-quarters of the time reject code recommendations. did.
AmEx declined to disclose the amount of its investments in Generative AI, but said those investments include technology infrastructure, commercial software and license purchases, and human resources. Thousands of tech organizations' 10,000 employees are working on some form of generative AI initiatives. AmEx has hired hundreds more to help expand its business.
Despite the excitement about large-scale language models, Mehta believes there is still a lot of untapped potential in more traditional forms of AI and machine learning for financial services companies.
“It's important to remember that much of the data is structured data, and not everyone is using classical machine learning models to the extent possible,” he said. “Some banks are saying, 'We need to get our act together on classic machine learning first. And until other banks figure that out, I'm going to follow suit quickly on generative AI.' ”