Skip to main content
Industry Expert Narratives

PayPal’s Guru Bhat shares on Data Management, Artificial Intelligence and Machine Learning

By July 25, 2022No Comments

Guru shares

Speaking as a professional in the field of talent acquisition, data storage is one of the things that excites people. They want to join us to solve these mountainous problems; when I look at it, it’s a triple-digit petabytes number. We are talking about close to 380 petabytes of data that is on our system across the years since we have been in existence. To put things in perspective for your listeners, if you were to burn this data onto DVD drives and maybe there are listeners who will ask what is a DVD drive. So assuming that they can google it and find it; if we were to burn all this data onto DVD drives and stack them up, it would be the height of roughly 8 Mt.Everests. That’s the kind of data we store. This data helps us in a variety of different ways. 

Traditionally, what we have done with this data is to help our business become even better at what it already is, which is security! It’s the non-compromisable characteristic of our business. You have to be secure. You do not have the luxury of being secure in a slow way. On the contrary, you have to be secure in an extremely fast way. What I mean by that is, when you are paying on a website, you don’t want that payment to take more than a few milliseconds, maybe one or two seconds at the best. If it drags on for five minutes, you are going to assume that this is not working and you’re going to close your browser window and move on, which means it has to be really fast. 

If you have to be fast and secure, you cannot do it in a manual way, you have to automate it. You have to use the data in an extremely smart way to make sure that you are making the right decision. The question is when to allow a transaction and when to block it? The safest mechanism will be to block all transactions. There will be no fraud because you are blocking everything. But that is the lousiest experience for a customer because nothing is getting through. So you have to strike the right balance between figuring out when to allow and when to block. 

For us, in the financial services industry, any company that moves high volumes of money is prone to high levels of fraud. Last year we dealt with $578 billion in payment volumes; any system that moves that kind of money can be siphoned off by those who rely on such frauds for their livelihood. Paypal has its share of people who go to work in the morning and come home in the evening while they try to poke holes and get some money off of Paypal on a daily basis. For us, our current fraud loss rates are less than one-third of 1% which is an extremely small number compared to any other payments company in the world. That accounts for why we are so profitable as a business because we lose very little to fraud. We use this mountain of data, triangulating all different data points about the customer to see which anomalies occur in payment and pointing out what doesn’t look right, for instance – this person usually shops with these kinds of things from these locations, from these IP addresses, etc… but how come today something different is happening? Can we find out more about whether this is an anomaly or if it’s okay? Keep in mind that all of this has to happen in roughly about 200 to 300 milliseconds as you have to say yes or no to the transaction in question. 

This is one of the areas where we use AI and ML heavily; then there are more straightforward frauds like somebody taking over an account, somebody spoofs somebody, that is relatively simple there is also complex fraud where some buyers and sellers collude with each other to defraud Paypal because Paypal has a unique value proposition that not too many other companies have. In fact, I don’t know if any other company does which is – we protect both our sellers as well as buyers. 

We have a policy of buyer and seller protection. For whatever reason, if a customer is not happy with your purchase, Paypal will refund the money. For whatever reason if you find, as a seller, you don’t get your money from the buyer, Paypal will intervene and will give you the money as long as you can prove that you did the right thing in sending the product. So this is a powerful value. Now, sometimes buyers and sellers collude to start defrauding Paypal. Finding that fraud is extremely complex and it cannot be done if you don’t have the right deep learning algorithms to detect which set of buyers and sellers are likely to collude with each other. This fraud and security risk has been prevalent for a long time. 

We are now beginning to also look at the other areas where it can help our business. For example, in customer support, we get roughly about 16 million customer calls every year and a significant chunk of those is actually to find out about – “Hey can I get some details about a transaction that I don’t understand or I don’t recognize?” this accounts for 11% of such calls. So now if we are able to equip our customer support executives or use other channels like chat to prevent this, I would have people not call us right? Because these are things that they can solve on their own easily, that would make it a better experience than having to pick up the phone, call and enquire. So we are trying to use AI and ML to figure out what is the best next action for a customer. How can we get chatbots to make sure that we can give them a remedy even before they ask the question – “Hey I’m stuck!” we can say, “I think you’re stuck here and would you like to do this?” People will be like wow even before I asked, they knew what I wanted. That’s a great experience. The question is, is this even possible? Yes, 300,000 servers! that’s what we have in our fleet, right now to deliver Paypal.com.

We have 300,000 servers that are serving traffic from around the world for all the transactions that happen. In these 300,000 servers, any problems could be occurring every second, and some of our consumers could be experiencing a problem. If we were to manually detect and remedy this, it would take an eternity, it would slow things down and we would always be in a fire-fighting mode. So we have moved to use data through instrumentation, we instrumented all these systems, we detect signals that indicate that some problems are probably occurring at someplace right now and we solve them even before it becomes a huge problem. So, there are an innumerable number of ways in which AI and ML could be used and we are only beginning to scratch the surface of how transformative it’s going to be to our business. 

Watch Guru’s full interview here:

Leave a Reply