Customer Behavior and Experience: Machine Learning Means Knowing Customers Better

Machine learning is helping payments companies and financial institutions by enabling a 360-degree view of the customer using data from a variety of sources both online and offline. The beauty of leveraging AI technologies such as machine learning is that the programs learn from use cases over time so your programs always get smarter. Learn how to tap into this powerful gamechanger.

Transcription:

Desiree Wolf: (00:07)

Hello everyone. We're welcome to the last track of today for the customer behavior and experience track, which is machine learning means knowing customers better. Today's panel will be moderated by John Adams, Executive Editor of Payment Source, which is being folded into American Banker. John Adams is the executive editor and has covered FinTech Banking, Financial Advertising and Capital Markets. His work also appears in American Banker magazine. He has a degree in journalism from Temple University.

John Adams: (00:44)

Our panel is customer behavior and experience machine learning, means knowing customers better. Our panelists are Donna Bailey, Principal at DB Innovators and William Tangalos, Senior Marketing and Analytics Strategists at Tredence. I'd thought you start off, if you could explain your job, your positions, what you're responsible for and the work that you're doing at at companies.

Donna Bailey: (01:16)

I've got about 20 years of financial services experience and maybe I'm part of the great resignation, but, I took a step back a couple of months ago, and decided to go out on my own as a consultant. My experience stems from American Express to Visa, to Wells Fargo, to a Processor I2C and even a money movement startup Neem. So I've really touched a lot of the different roles, that a lot of folks here at the payments conference probably are in. I thought that my, experience and just innovative spirit could really help some larger clients on the backend. I'm really passionate about artificial intelligence. I wrote a couple of articles about it and when I was in communication with American banker about what to present, I said, could we please talk about this because it's so important for the payments industry. So here I am.

William Tangalos: (02:21)

It's good to be here. My name's Will, I've worked in machine learning old schools called Predictive Analytics about 25 years, Variety of FIS, Wells Fargo, Visa, Household Credit, some retailers. I've just been, always excited about the mixture of what can feed decisioning engines and now called machine learning and AI to come up with recommendations that you can use to monetize. So whether it's kind of managing the kind of channel the person wants, or if they're likely to have fraud or maybe not pay well or in other areas I've been in a lot like customer acquisition, retention. For about 20 years, I've worked on kind of leading or even in some cases, building predictive models in that space.

William Tangalos: (03:17)

So, currently I work for a company called Tredence, probably never heard of them, but they're a pretty big growing company. They specialize in a holistic machine learning and artificial intelligence solution. They don't kind of just throw the requirements over the wall to engineers and let them go where. Sometimes maybe some of you have been experienced some disconnects. They really take a holistic approach from the minute with the client all the way through to really use the AI and ML to drive the business value. So it's kind of exciting and good to be here.

John Adams: (03:48)

Thank you both for attending. Machine learning and AI are pretty broadly used terms, particularly outside of science. How do you define them, particularly in the context of use for financial services or payments?

Donna Bailey: (04:03)

Machine learning is all about statistics and algorithms and in my own experience. I've used it, I've partnered with my risk teams in, previous, roles where it's really helped us target customers, with the right offer and in the right way. As Will said before, machine learning is definitely targeting, but it's also used behind the scenes for other artificial intelligence, focus areas like natural language processing and chatbots.

William Tangalos: (04:41)

Machine learning's are a kind of a subset of AI and there's human beings supervised learning. So giving inputs on demographics and psychographics and maybe payment behavior and letting the math be done through the algorithm to then come out with an output and it kind of can learn, but it can't learn as much as unsupervised learning, which is more like things that can use neural network and types of things that don't need human involvement. There's no real necessary initial training invalidation, like in the supervised, when in the unsupervised it's more like what we've all been hearing over the years. You let the machine go and it really learns, but there's a lot of work that still has to go in and setting those up. I've been exposed to banking, financial services and healthcare projects using both of those techniques. Like Donna said the financial services identifying fraud, identifying late payers and then maybe on the channel preference side are which consumers have data that suggests they're much more likely to come to the site or use an app, or like to use a card for payment and in the marketing world. A whole bunch of channel preference, who is more likely to buy or be cross sold a product at any point in time, whether it's healthcare or financial services.

John Adams: (06:02)

What's important in payments now, is getting money to people quickly. Getting money to people organization in real time is as fast as possible. How does this technology help fill gaps, that it was difficult for people? In other words, what for real time payments, for example, how can AI in theory is being used there or how could it be used to for real time payments?

Donna Bailey: (06:30)

Actually, machine learning for real time payments, and that's the role that I just left, I built real time payments for the startup. It's a great tool for identifying anti-money laundering and even fraud. Lots of people are nervous about if you send money globally, is the recipient getting the money, or are you sending it, is the recipient nervous that they don't know who the sender is and so machine learning's able to tag transactions, that look suspicious. it really helps institutions, feel better that there's some sort of security and monitoring. Again, you're not letting the bad players send money to terrorist organizations, so I feel like that's been my experience with machine learning. Well, you might have a different one.

William Tangalos: (07:28)

That's a good example. I mean, some of the things I think when we were like at Visa too, we've seen the severity level. So a lot of the automated machine learning can spit out in real time, really severe transactions for fraud and maybe moderate and low, and then route those operationally. That's been going on for a while. That's kind of a typical example of machine learning, particularly in the fraud management side of things.

John Adams: (07:53)

After the fraudsters themselves have machine learning or reformers machine learning and, even if they do, how can that be countered? Is it the same? If cat and mouse game that exist they willl swear with protecting transactions from theft.

Donna Bailey: (08:10)

I am gonna pass that one to Will. I would say the exciting thing about machine learning is that the computer programs continue to learn. It's not static, you have a computer program, machine learning model and you catch the fraudsters and you are done. It will continue to learn like the suspicious transactions where they come from, maybe what quarters they come from. It's like an iterative process. I don't know how you maybe combat the fraudsters if they have machine learning models, but they can't sort or interpret yours.

John Adams: (08:46)

It could make behavioral analytics.

William Tangalos: (08:49)

Whose unsupervised machine learning is winning. So yeah, that's actually a good example of unsupervised. Maybe these words don't mean a lot to people supervise. So the unsupervised is what Donna is talking about the way, you do not need a human involvement to say. One of the thing is when the transactions coming from these companies get the radars up and the machine can learn that on its own with no need for a training data set or a validation data set. Same thing for other characteristics, perhaps time of day weekend, certain industry codes, amount, GE dollar amount, certainly all these are different characteristics that the computer can learn on its own. It doesn't need a human to train it and then validate it. So that is a pretty cool sounding, but also very effective capability of unsupervised machine learning of that kind of capability.

John Adams: (09:41)

We have been talking about fraud mostly, but it's also a lot of potential uses and customer service. What are some of the ways that customer service can be improved by using machine learning?

Donna Bailey: (09:53)

In my experience, you wanna present the right offer to the right customer at the right time. I am sure everyone has heard of that before, but with machine learning, you really can do that because it will take in the unsupervised data from customer service calls. Maybe this customer has called in the past, so it'll take in the data from that or social media. It is able to take in data, synthesize it, and then present the right offer to a customer. In my experience at Wells Fargo, we targeted new to credit customers who we should upgrade and who we should not. We use machine learning models to target the right customers because not everybody is ready for a secured credit card or an unsecured credit card. They do not understand APRs and how to manage their money, and so by using machine learning model, we were able to target the customers who were ready for an unsecured card and it saved the bank, just a lot of money because you're presenting the right offer to the right customer and you're enabling them to build their credit as a result.

John Adams: (11:08)

You mentioned the ties to the social media use. How does it work with social media and how does it inform, let us say more traditional decision making or customer engagement channels?

William Tangalos: (11:19)

Yeah, so like I was gonna talk about a little bit an actual example, get a little media on a, on a next best product offer it was not only payments, but it involved payments, but some of the inputs came from supervised demographic account behavior, psychographics, geographic, product behavior, and then unsupervised like certain social media behavior whether it is tied to a region, the person lives in, or maybe other characteristics of the person where it is not necessarily the exact person, coz you are probably not gonna get social media feeds daily or weekly on, 14 million of your customers. So that's been an aspect I've seen, in a project I've worked on where the supervised was used to kind of come up with segments that were predefined. And then the, unsupervised came in in real time, whether it was time of day, a week, changes in social media changes in other characteristics time of or I already said time of day where that data is used together to say, Hey, out of this segment, now that we know what it, who the person is based on this real time, learning now through decisioning engine, including social media data here is the next best specific product offer and the dollar amount to offer to the consumer and I have been involved in some projects like that we have had very good success when we have done like an AB test business as usual way without the ML and AI, and then utilizing the ML and AI for formulating the next best offer. In real time, this example was in a call center

John Adams: (12:47)

On the subject of call centers, is it what kind of improvements can be derived saved in an IVR type environment? Is there ways to improve that and to the degree that it's being used, would some examples be in terms of, people calling in with a question, a complaint, some sort of problem, is there a way AI or machine learning can and is being used in those areas?

William Tangalos: (13:18)

Not in this exact project I worked on in the first generation, but in other projects, we have used some like NLP to help transcribe actual calls from the call center. One of the companies I worked at about a year ago, they actually do a million calls a day, pretty world class. And in terms of who to call and why, and when definitely not world class, how to build a customer centric database, but what we did do there is we started using NLP to actually transcribe the actual conversation between either the SDR or the actual agent and classify those based on the actual words being, heard through the AI, into categories that we then use for downstream actions. And so that's one example from a call center, from in this case, an inbound call center, where that aspect was used, using an unsupervised, type of, non-defined kind of capability. It's not like hard data saying you're this age or, this income, but unsupervised where you're not really sure where the call goes. And so the AI needs to kind of help figure out how to group that in an unsupervised way to then make sense of it, to use it, to differentiate, the call for subsequent behavior.

Donna Bailey: (14:35)

Yeah, I was gonna add when I was at Wells Fargo as part of the credit card product organization we would meet and we would use the machine learning models that used NLP, from our customer service centers. So you could look at the tone of voice on the call and things were categorized. So tone of voice, subject time of day customer segment. And we came up with some, real areas of opportunities. Balance transfer is always one, right? People are just confused about how to do it, or it doesn't get done properly. I mean, we identified really key strategic areas that as an organization we worked on to improve, and we would monitor, this number of balance transfers has gone through without complaints or, we've increased the number of balance transfers, from, this certain card and really using the NLP as a result of the customer service calls that were coming in really helped us to improve the customer experience. And also, just change the way we were communicating with the customers.

John Adams: (15:44)

William, you mentioned having done some work with healthcare, the payments with healthcare are very different than normal consumer payments. It involves third parties, multiple moving parts in the explanation of benefits. What are some of the ways that AI or machine learning can be used there, particular when we are talking about multiple parties?

William Tangalos: (16:07)

Yeah, in the last healthcare company I worked with, they liked to market to the health conscious kind of in an older age group. they sold the Medicare, customers in that older age group and they also sold, sold other products to the health conscious. So one of the things that the company did is it designed a lot of quizzes to both help create engagement. And also though help classify the health conscious level using AI by the answers of the quizzes that they got down to, starting with hundreds of quizzes down to just a handful that through AI projects were able to come up with a very accurate way to, to indicate if you take these, this quiz and you answer it this way, you are this level of health conscious or that level of health conscious.

William Tangalos: (16:52)

So what that led to kind of get back to your question was on that the premium level and the kind of carrier that would maybe be associated with the premium level. The next phase of that was gonna get into a little bit of payment, but the bigger issue for the healthcare company at that point was to use AI based on, the conversation and the data they captured from the consumer to kind of tier them at a different premium level, and then align that with a carrier that fits that, premium level

John Adams: (17:24)

Now in terms of risks, what are some of the risks to using a AI and how do you manage them? In other words if something is being turned over where there's less human involvement, how do you balance that?

Donna Bailey: (17:44)

Well, I think there's an overall concern about AI and the biases that AI has and even in machine learning models, I mean the inputs, specialize especially the unsupervised learnings that a, human's not really involved. Like who's really checking if the data that's going into those models is up to date is clean. I would say those are concerns, who's kind of checking that the data built in the models that you're gonna use to target maybe customers or, mitigate risk. Who's checking it, I guess that's just it, like, who's overseeing that the data isn't biased and that we're not maybe targeting all boomers for retirement packages when maybe they're not retiring anymore. I mean, that would be my.

William Tangalos: (18:40)

It's almost like using the old school way before they call it ABC testing, you have unsupervised and it's just machine learning, but it's not all that fancy. How is that doing and now if you are bringing in unsupervised characteristics where the machine is learning on more on its own, you don't just let that the first time out of the gate run rampant. I mean, you look at a lot of iterations of it and then even when you do feel you have a substantially better solution than the business is usual, challenge or champion prior champion of using supervised human input characteristics to train and validate, you would then design a nice AB test that scientifically in empirically shows how much better the unsupervised did than the supervised in a way, and then incrementally, look at rolling that out. So you minimize your risk in this, less easy to explain, unsupervised, type of, technique.

John Adams: (19:37)

We recently read some research or a story that had research in it from KPMG that said about half of banks are abusing AI or machine learning for, in some form. Is that a higher level of adoption than you would expect or, is there room a lot more room to run there?

Donna Bailey: (19:59)

I think it's 82% or so, I forget the stat, but, I introduced or I convinced my risk team to test out machine learning models in 2017. And that was early days to be honest, machine learning and AI has been around since the fifties, but it's only gotten much more sort of exposure and visibility because of the amounts of data and the speed of technology nowadays. So, I'm not surprised if 82% of banks are using artificial intelligence, I would think competitively you need to, because, it's just been shown that it helps, mitigate risk, can help you save money on collections, target the right offers. So to me be crazy if you didn't look into using artificial intelligence, and machine learning models for those areas.

William Tangalos: (20:57)

The only other thing I'd add to that, is the supervised approach of regressions likely to buy, like me to not buy. I think I managed my first model in 1988 in neural networks and things like that were just coming out then for the banking world. So that's when this great AB testing was going on. You had the older school at the time regression type of classification, people that would rank order your 10-15 million customers likelihood to do whatever and then you would try to see how well, a neural network type of approach, which is more unsupervised would do. And so that's actually what got me into this whole area. I was very excited from the business side and working with the technical side. My take would be that the maturity of the supervised is a hundred percent of banks are using that. I think on the unsupervised side, I would tend to think that's less, but I'd be very interested in seeing the statistics on the proportionality of big banks, using unsupervised versus supervised for things like fraud management claims collections, versus the marketing side on acquisition retention, upsell, churn. That would be a neat thing to talk to somebody at Forester aorund or Gardner.

Donna Bailey: (22:13)

Yeah. I would say too, machine learning is used in investments as well. So it helps to predict just market factors and I've just read some different examples of how investors are using machine learning to help gauge, maybe which stocks are gonna perform better than others. So I feel like I come from mostly credit card, but, banking overall is such a big industry that, it's not just credit card or customer service, it's investments, it's other things as well that people are using machine learning for.

John Adams: (22:48)

We have been discussing banks a lot. Are there differences in how let's say FinTech's or, digital financial services companies are using machine learning and is it the case where they're using it in some ways faster or adopting certain elements of it earlier than more traditional financial institutions.

William Tangalos: (23:09)

What I've seen in my last couple jobs, we're at startups, so no stores. So none of this giant offline-online battle, which that's a whole another subject. I think the data management in those two areas are two different worlds, which is unfortunate. So when you're in an environment where it's all online, it's, technically easier coz you've had the beginning and the end from the impression through what happened to the customer a lot more linked than when there's offline, involved. My experience is that it's easier and faster to be able to have less people involved in the digital world to go that way. But on the other hand, my take is that a lot of these companies organizationally are missing the boat on customers. That aren't always a hundred percent online, which is a whole another area.

Donna Bailey: (24:02)

Yeah, I would say, again, in kind of the research I've done a lot of traditional banks, or even a company like PayPal, they're acquiring these FinTech's because they are kind of nimbler and faster in using machine learning models. Already it's like not a big deal for them to switch on a machine learning model. Whereas sometimes, in a larger institution, you have got to get, a whole bunch of departments to agree, like we're gonna use this new model. It's been interesting to see that, as far as adoption some of them are either working with FinTech as like vendors or partners, or they're acquiring them as part of their company.

John Adams: (24:45)

We have a couple of minutes left. Do we have questions from the audience?

Speaker 5: (25:01)

How was able to the right product with a product like secured card and the tendency for going against lower economic income, how were you able to address or use machine learning and, address concerns about implicit or unintended bias?

Donna Bailey: (25:20)

So, before I introduce the could machine learning model!

John Adams: (25:25)

Could you repeat the question?

Donna Bailey: (25:26)

Please? Oh, I'm sorry. Can you repeat the question here.

Donna Bailey: (25:31)

I gave an example of the secured card and how were we able to use machine learning models, that didn't have a bias to targeting the right customers. So my answer to that, thank you for the question, was before we use machine learning models, our risk partners would just use risk models, which were based on when we acquired the secured card customer. So with secured cards at Wells Fargo, and I think they've actually sun-seted the product, but you had to pay on time for a year and you had to have a $300 deposit in order to be eligible for upgrade, but the risk models were only taking into account, your income at the time of acquisition. So a year later, two years later you've probably changed.

Donna Bailey: (26:26)

I mean, hopefully your income has gone up or not, but machine learning model again, took the data from different sources, not just the income we looked at, maybe if they had other products with Wells Fargo, and how we didn't overlook the biased was we would run testing to Will's point. I think AB testing will be around forever, right? No matter how much new technology or new sort of approaches come in, it always makes sense to test. So we tested to make sure that the machine learning model versus the risk model, made sense and that the customers were the right customers to be upgraded to an unsecured card.

William Tangalos: (27:15)

Yeah. The only other thing I've to add to that too, is that other finance companies and at Wells Fargo, besides the AB testing, we could also then evaluate if there was disparate impact on certain characteristics where it got a little dicey. So I remember being involved in some of those conversations, some of aren't always just, clearly, one way or the other, but, it's something very scientifically that very easily can be set up as long as you have those characteristics. And then as Donna said, set up the AB test properly so that you can really measure the results and believe you're comparing something that is statistically reliable to see if there is first of all, which tool, which technique worked. Then if there was any disparate impact, some projects I've seen, that's been around a long time.

John Adams: (28:04)

Yeah. well, thank you. Thank you, Donna. Thank you, William. Thanks, a great panel and, thank you everybody for attending. Everybody is invited to the reception outside. Thank you again, everybody.

William Tangalos: (28:20)

Thank you.