How do you automate risk 8 billion times a year? In this episode of TechFirst with John Koetsier we chat with Anjali Dewan, American Express’ VP of Risk Management.
Credit card companies have some serious challenges … with trillions of dollars in transaction volume, they’re target #1 for fraud. But customers expect everything to work perfectly every time.
So American Express started managing every single risk decision on risk using AI in 2015, which makes them much faster. They can now make billions of decisions in nanoseconds, using what might be the largest commercial machine learning system on the planet (probably excluding Google and Facebook).
Get the full audio, video, and transcript of our conversation below …
Subscribe to the TechFirst podcast
Apple Podcasts
Spotify
Google Podcasts
Overcast
Other podcasting platforms
Watch: How Amex uses AI to automate 8 billion decisions
Subscribe to my YouTube channel so you’ll get notified when I go live with future guests, or see the videos later.
Read: How Amex uses AI to automate 8 billion decisions
(This transcript has been lightly edited for clarity).
John Koetsier: How do you automate risk 8 billion times a year? Welcome to TechFirst with John Koetsier.
So, as everybody knows, credit card companies have some pretty serious challenges, right? They’ve got trillions of dollars in transaction volumes, so they’re target #1 for fraud. But, of course, customers expect everything to work perfectly every single time. Well, American Express manages over a trillion dollars in annual transactions and every single decision on risk uses AI, so American Express can make those billions of decisions quickly. The system might actually be the largest commercial machine learning system on the planet.
To find out a little bit more about how it works, we’re chatting with Anjali Dewan, who is American Express’ VP of Risk Management. Anjali, welcome!
Anjali Dewan: Hey, John. Thanks for having me.
John Koetsier: It is a real pleasure to have you on the show. Thank you for coming on. Let’s kick it off right here … I mean, 8 billion decisions. Wow. What kind of decisions are you talking about?
Anjali Dewan, Vice President, Risk Management at American Express
Anjali Dewan: If you think about the scale at which American Express operates, we’re present in 106 countries. We have 140 million cards in force, and through our network over $1.2 trillion flows. So it’s very interesting. When you think about the volume, it’s really important for us as American Express to have our customers’ back. And one of the ways we do that is to protect them from fraud, as an example.
So when the trillions of dollars are going through our system, we have AI-powered credit and fraud risk models. And these models are in real time monitoring 100% of these transactions and returning 8 billion credit and fraud risk decisions in real time, we’re talking milliseconds over here. You know, we believe at American Express that we have the world’s largest and most advanced machine learning system in the financial services industry.
100% of our models are AI-powered and it cuts through the customer life cycle. So starting from new account origination, limit assignment, customer management, and fraud detection.
John Koetsier: It’s pretty impressive. I mean, to automate all of that, and we’re going to get into questions around how you started, when you started, and how you’ve learned to trust the artificial intelligence, which must’ve been a bit of a challenge. We’ll get into that.
What kind of hardware are you talking about here? What kind of server farm, cloud-enabled system are you talking about to enable all this?
Anjali Dewan: Sure. So as you can imagine, we have a large number of use cases and with that comes a mix of hardware and software, and a lot comes down to the use cases. For example, we were playing around with deep learning which really looks at GPU, GDX. So, whereas we have the AmEx private cloud, and with that we have computational servers which are configured to supporting machine learning R & D.
So we have a bunch of CPU and GPUs out there. And then finally we have a production cluster, which is really fine-tuned to make sure that our production use cases run really well. So I would say it’s a mixed bag.
John Koetsier: So it’s big, but you’re not going to tell me how big.
Anjali Dewan: Yeah. It’s big.
John Koetsier: It’s big. Okay.
Anjali Dewan: 8 billion decisions so … real time, $1.2 trillion. It’s big.
John Koetsier: Okay. So, you mentioned that 100% of your credit risk models are powered by machine learning, and actually that’s happened since 2015. I mean, we’re talking five years ago. So AI is really hot right now, we all know about that. But 2015, I mean, that’s pretty early. How did you start making that decision back then to lean on AI, and who made that?
Anjali Dewan: Yeah, that’s a great question. You know, the journey actually started in 2010.
Back in 2010, we started playing around with different machine learning techniques, both in the credit and the fraud risk space. And as we ran these experiments, we saw very meaningful and consistent improvements across our models.
And that really made us move to the next phase. And the next phase was really introducing these models as challengers to our non-AI models in a live environment. So in small populations, we would run these models and see what results were. And very consistently across credit, across fraud, these experiments were showing that these models were driving superior accuracy, superior model performance, and ultimately translating into very superior outcomes for our card members. So really when we saw these test results, it really powered us to fuel the AI transformation of the models. And it comes to who, such a large-scale transformation cannot happen without the powerful backing of truly visionary leaders. We had leaders back in 2012 who committed to investing in the infrastructure for big data, who committed to hiring the best talent, upgrading our talent.
So all of this came together to really build 100% deployment in 2015.
Why do we trust that things are working well? That’s a great question. We trust because of two reasons.
The first is we monitor. So when our data scientists are deploying these machine learning models, they go through a very rigorous process of tracking these models to make sure that the model performance, the accuracy, the decision outcomes, the card member experience is truly superior and in line with what we expected.
The other reason we trust as modelers and data scientists at American Express, we have access to three very unique data assets. The first is data.
One thing that’s very unique to American Express, because of our globally integrated payment model, is that we issue cards with our merchants and we’re also the network. So our data scientists have access to a truly unique data set to train their AI models and fine-tune it.
The second part is just our deep expertise in machine learning and deep learning. We started in 2010 and we’ve been learning since then.
And a third is our data scientists, so our machine learning group. We have a fantastic set of PhDs and MS folks who are never satisfied with the status quo. So they’re constantly challenging, they’re constantly upgrading themselves. And one thing that’s absolutely vital in the AI space is that it keeps evolving, so you need folks who are constantly getting better and experimenting.
John Koetsier: It’s pretty interesting. I mean, huge scale, obviously. Very, very impressive. Question is: what’s the business impact, right? What does it improve? Where’s it hitting your bottom line or even your top line? Maybe we’ll start there, like how much credit card fraud is there and what’s the universe maybe in general out there that you’re trying to keep at bay?
How much fraud is there in the system ?
Anjali Dewan: Yeah, let’s start with outcomes, and you hit the sweet spot, right. One of the direct outcomes of AI was a substantial improvement in the fraud space. We saw fraud losses improve dramatically as a result of the AI models.
Now there’s a lot of data out there on how much fraud there is, but if you go by the Nielsen Report, you’ll see that American Express for 13 years has come out as the lowest in the fraud space — and not by a little, but by half.
And what drives these superior outcomes in fraud is so vital as we protect our card members and have their back, is an AI-driven fraud prevention system. So even through the pandemic we’ve seen the fraud attacks evolve, but at American Express we’ve seen that, you know, the fraud losses have continued to remain very low. Some of the other outcomes that improve dramatically as a result of these AI models were in other spaces as well.
So I’ll give you an example, with automated resolution with our card members. So think about, let’s take an example, you have a card member, he’s flying from London to Rome for the first time, super excited, leaves the airport, sees a gelato store, goes out there, swipes his American Express card, and guess what, he gets disrupted because we have a fraud concern.
What’s going on in the background is within 15 seconds our machine learning algorithms are initiating a personalized communication with this card member … which could take the form of an email, it could take the form of a push notification, it could take the form of a text. And that allows us to real-time communicate with the card member and resolve the fraud concern.
So, the card member could say, ‘Hey, yes, that’s really me having the gelato.’ And we want the customer to have his gelato, but then we also take that response and feed it real time into a machine learning algorithm. So the next time the card member buys a gelato in Rome, we recognize it as legitimate spend.
So if you think about since 2014 when we deployed our AI solution for fraud, our resolution rate, digital resolution rate for fraud has improved by 100%.
So it’s really staggering how much customer experience can improve through AI. I think the third most significant impact it’s had is really reducing customer disruption.
John Koetsier: Hmm-hmm.
Anjali Dewan: So we deployed our machine learning solution for purchases back in 2014, and several things happened. We elevated our fraud prevention services. We reduced loss rates. But we also substantially improved customer experience.
So point of sale disruption has been reduced by 21% since 2014, which is absolutely fantastic for our customers.
John Koetsier: It’s really, really impressive. And I speak as one who traveled a lot. I mean, I think there were a couple of years where I had a hundred thousand kilometers in the air — sorry, I’m Canadian, I speak kilometers — but there’s nothing worse than arriving and you can’t use your card.
And that can be really, really challenging as well in the days maybe before Uber or Lyft or something like that as well, and you’re in a cab you need to pay for transport and other things like that. So that’s a really, really critical piece. You don’t want to lose money to the fraudsters, but you also don’t want to make your customers angry and inconvenience them.
Let’s chat a little bit about the AI technologies that you use. I mean you’ve had, it sounds like almost, you know, more than five years because you started before 2015, but you’ve had almost a decade to really perfect what you’re doing and evolve what you’re doing.
What AI technologies are you using?
Anjali Dewan: Yeah, that’s a great question. And again, I think the key is that AI continues to evolve. So it’s really important for us to be able to constantly test into the newer techniques. I’ll give you a flavor of some of the things that we are looking at right now.
So one is sequential RNN. The way you can think about it, it’s the ability to read data sequentially and establish relationship between transactions. So if we go back to your example of this card member who is sitting in Rome. Let’s assume he buys a gelato at 5:00 PM and we know it’s a legitimate transaction, and then we see at 5:15 a gas spend in London on this card. And so the sequence really doesn’t add up. It’s extremely suspicious. And then, you know, something like sequential RNN will allow you to detect this and in a truly automated way.
So it’s really interesting as we explore sequential RNN to really improve our fraud detection capabilities.
Some of the other stuff out there is self-learning models, RNN, GAN, and I think what we’re really excited about in the coming month is the 10th generation of our global fraud detection model.
If you go back to pre-2014, we used to have 150 models to manage fraud detection across the globe, across countries, across our consumer and commercial portfolios. And the model as it stands today, is a single global model which has a view to the $1.2 trillion flowing through our networks, and a new model that’ll come out next month will have just the freshest data and very powerful new features.
So I think the great thing is we’ve just scratched the surface of AI at American Express, and we are very excited to see how it can improve a customer experience.
John Koetsier: Super impressive. You’ve talked a little bit about some of the data that feeds into this and obviously location data is one piece, and probably amount of purchase is another piece, history of your purchases as well. What other kinds of data feed into your fraud models?
Anjali Dewan: Yeah, that’s a great question. I mean, you think about American Express, and again comes back to the scale. You know, we’re accepted in 160 countries. We have 140 million cards and we have a trillion plus flowing through our network.
But what’s really unique about data American Express is our integrated payment model. So we actually have transparency into the entire payment chain, and that starts with, you know, we issue cards, we acquire merchants, and we are also the network. And all of that comes together in creating a data set which is truly unique and powers enterprise-wide decisions. So really the fraud model truly benefits from this closed loop integrated model at American Express.
John Koetsier: Anjali, thank you so much for taking this time with me. I really do appreciate it.
Anjali Dewan: Of course. Thank you for having me.
John Koetsier: Absolutely. And thank you everybody else for joining us on TechFirst. My name of course is John Koetsier. I appreciate you being along for the ride. You’ll be able to get a full transcript of this podcast in about a week at JohnKoetsier.com, and of course the full story at Forbes will appear shortly thereafter.
Plus the full video will remain available on my YouTube channel. Thanks for joining. Hey, maybe share with a friend. Until next time … this is John Koetsier.