Fighting Fraud with Machine Learning at American Express
Global financial institutions spend billions of dollars on credit card fraud. American Express is fighting this fraud through investment in machine learning.
Days before embarking on a recent trip, I received an email from American Express stating “Based on recent transactions, it appears that you may be traveling soon. We use industry-leading fraud detection capabilities and do not need to be notified of travel plans in advance to recognize when our Card Members are traveling.” While other card issuers, such as Chase  and Wells Fargo , require cardholders to notify them in advance of travel, American Express does not provide any way for customers to set up travel alerts, instead relying completely on internal fraud detection technology.
Machine Learning at American Express
American Express has long been at the forefront of transaction monitoring and fraud detection. In 1988—at a time when merchants had to call American Express for large credit authorizations over the phone—an American Express employee had to make a quick judgement call to determine if payment could be authorized. American Express leveraged expert systems, a then-cutting-edge technology that used conditional statements and data from up to 13 databases, to determine if a purchase was in line with the customer’s transaction history.
Three decades later, financial institutions around the world continue to combat credit card fraud, collectively spending $27.7 billion on card fraud costs in 2017, a number that expected to grow to $31.7 billion by 2020. For its part, American Express has invested heavily in machine learning to more accurately identify card fraud. In an interview shortly before the end of his 40+ year tenure at American Express, Ash Gupta, President of Global Credit Risk & Information Management, identified two main benefits from this investment. First, by training fraud detection models on the over $1 trillion of annual transactions, American Express has developed tools that have proven far more accurate than the manual if-then rules developed in the 1980s and 1990s. Second, training and retraining models on an ever-growing amount of data is a faster process than developing the older format of fraud rules. Whereas expert systems were manually tuned, machine learning models can be rapidly retrained on this new data.
American Express has chosen to develop its machine learning technology in-house, employing 1,500 data scientists globally as of March 2017. The company views its investment in machine learning as an evolution of its earlier investment in expert systems, and believes that they will be able to continue developing technology more efficiently internally.
Though American Express and its peer institutions have made significant progress through the use of machine learning, there is still more to be gained by improving the accuracy of their fraud detection models. One area for improvement is in reducing false positive rates, transactions that are mistakenly blocked because they are thought to be fraudulent. Card issuers are estimated to lose 13 times as much from false positives than actual fraudulent transactions. A study from MIT found that this rate of false positives can be reduced by as much as 54% using Deep Feature Synthesis, an automated method for identifying complex features from input data. For example, a “deep feature” might be “How much money was spent in a coffee shop on a Friday morning?” A change from $5 to $15 in a given week might then be flagged as potential fraud. Importantly, this technique retains a human-understandable reason for why a transaction might be declined. Gupta has spoken about the importance of identifying why an algorithm makes a decision, stating “for us, it’s very important that the model has sufficient transparency and that we can describe for the customer exactly the reason we are taking the action [that we are taking].”
Questions remain about the future of machine learning at American Express and the effect it will have on the organization. First, to what extent can machine learning be used to augment other areas in the business? American Express has begun to use this technology to provide more targeted ads and deals based on customer spending habits. Are there other areas where American Express can leverage user information to provide a better experience?
Second, to what extent will machine learning replace jobs at American Express. The company has stated that its investment in machine learning is in the pursuit of augmentation, not automation. However, as seen in the case of fraud detection, a job that was once completely manual was first augmented with expert systems and is now completely automated through machine learning models. Will the use of machine learning stretch into other areas in the company and replace other roles?
(Word Count: 754)
 American Express <AmericanExpress@welcome.aexp.com>, “Helpful Information for Your Upcoming Trip,” email message to Andrew Harris, April 10, 2018.
 Chase, “Travel Notification,” https://www.chase.com/personal/credit-cards/travel-notification, accessed November 2018.
 Wells Fargo, “Travel Tips and Tools,” https://www.wellsfargo.com/goals-banking-made-easy/travel-tips/, accessed November 2018.
 Dorothy Leonard-Barton and John Sviokla. March 1988. “Putting Expert Systems to Work”. Harvard Business Review. https://hbr.org/1988/03/putting-expert-systems-to-work, accessed November 2018.
 Sarah Kocianski and Dan Van Dyke, “AI in Banking and Payments: How artificial intelligence is cutting costs, building loyalty, and enhancing security across financial services,” Business Insider Intelligence, February 6 2018, https://intelligence.businessinsider.com/post/ai-in-payments-and-banking-2017-12, accessed November 2018.
 LendIt Conference, “Innovation in Credit Granting With Big Data – American Express’s Ash Gupta,” YouTube, published March 6, 2017, https://www.youtube.com/watch?v=4obUWkBuzs4, accessed November 2018.
 Thomas H. Davenport and Randy Bean. March 31, 2017. “How P&G and American Express Are Approaching AI”. Harvard Business Review. https://hbr.org/2017/03/how-pg-and-american-express-are-approaching-ai, accessed November 2018.
 Kocianski and Van Dyke, “AI in Banking and Payments.”
 Rob Matheson, “Reducing false positives in credit card fraud detection,” MIT News Office, September 20, 2018, http://news.mit.edu/2018/machine-learning-financial-credit-card-fraud-0920, accessed November 2018.
 LendIt Conference, “Innovation in Credit Granting With Big Data – American Express’s Ash Gupta.”
 Davenport and Bean, “How P&G and American Express Are Approaching AI.”
Student comments on Fighting Fraud with Machine Learning at American Express
Thank you for your great insight into machine learning at American Express! While I found American Express’ efforts to reduce fraud impressive, I do wonder if this is a go-forward requirement to stay competitive in the market. Capital One, for instance, is also leveraging machine learning to reduce fraud (https://partners.wsj.com/aws/capital-one-rethinking-fraud-protection-machine-learning/ ). Amex also seems to be exploring broader applications of machine learning via its investments arm. For instance, Amex Ventures has made investments in Enigma, which leverages data and machine learning to connect internal and external data. In Financial Services, this data can help with compliance programs and underwriting. Their investment in RetailNext, a startup focused on using deep learning and AI to support physical store retailers, showcases a dedication to provide broader support to struggling retail clients.
Thanks for sharing a great perspective on AmEx and machine learning! I do agree with KS’ point that using machine learning to improve fraud detection is becoming table stakes for banks, but in my opinion that doesn’t make it any less exciting. This article (https://thefinancialbrand.com/72653/artificial-intelligence-trends-banking-industry/) summarizes the perspectives of the big consulting firms, which estimate that artificial intelligence will add around $1 trillion of value to the banking industry by changing the way banks do the majority of their work. For an industry with thin margins (ROAs of less than 2% and ROEs less than 12% https://fred.stlouisfed.org/series/USROA; https://fred.stlouisfed.org/series/USROE), this is a tremendous opportunity!
Thanks for sharing this, Andy! Really interesting to see how much this has saved AmEx. Regarding your questions on jobs, I think this presents and interesting question for automation more broadly; how do we think about this trade-off in the context of our economy? Overall, I think it’s important to continue to pursue technological advances even if it does replace roles. If we think to other areas of automation, generally they have had a positive impact society even if the short term effect is fewer jobs.
Thanks, again, for sharing!