Jim Young sits down with Dallas Wells, Chief Success Officer at PrecisionLender, to discuss common points of confusion we've seen around the idea of Artificial Intelligence in Commercial Banks.
- AI is just bots
- AI is for retail, not commercial
- Data used by an AI will be visible to other banks
- AI will replace Relationship Managers and their interactions with customers
- AI is still 5-10 years out
Commercial Banking + Artificial Intelligence = "Andi" - Blog (2 minute read)
The Time is Now for AI in Banking - Blog (5 minute read)
Meet Andi - Podcast (25 minute listen)
PrecisionLender 2016 Highlights and 2017 Roadmap - Webinar (30 minute watch)
Jim Young: Hi and welcome to the Purposeful Banker Podcast, the podcast brought to you by PrecisionLender, where we discuss the big topics on the minds of today's best bankers. I'm your host, Jim Young, Director of Communications at PrecisionLender, and I'm joined by Dallas Wells, our Chief Success Officer. Thank you all for tuning in. Dallas recently gave a presentation on Artificial Intelligence and we'll call it AI from here on out, for a banking Salesforce users group, and while there was interest, there was also skepticism. Some of that skepticism was well founded, and some of it was not, but the latter group inspired us to record today's podcast, five misconceptions about AI in commercial banks.
Here are the five that we're going to address during the course of our conversation. One, that AI is just bots. Two, that AI is really for retail and not commercial. Three, that the data used by an AI will be visible to other banks. Four, that AI is eventually going to replace relationship managers, and five, that AI really is something that's technology that's still 5 to 10 years out. Let's start right from square one on this. When we're talking AI, Dallas, what exactly are we talking about?
Dallas Wells: Yeah, so AI is everywhere, right? In anything you read about banks, AI is kind of the hot topic. A lot of people are really asking the big question, "How are we actually going to use this?" And, "Is it just bots? Is that all that AI really is, is a way for a computer to basically impersonate a human?"
Jim Young: Is it Siri for banks, basically?
Dallas Wells: Right. I think that's where a lot of the skepticism comes from, is, "Do I really need a bot to do what I'm doing right now?" AI is a whole lot more than bots so that goes all the way back to the Turing Test, when we first started talking about artificial intelligence, and that test was basically can a computer converse, interact with a human in a way that it's impossible to determine which is the computer and which is the human? That was his test for it, but that has now morphed into much more than that. AI is this big overarching umbrella, and artificial intelligence covers lots of different things. Bots is one of those categories, this conversational interface that we have with Siri and with Cortana and with Alexa for Amazon, and all kinds of different formats.
It can be by voice, by text, and there are tons of those and more every day. That's one aspect of it, but there's this other category that I think is maybe much more interesting and one that we'll talk about a little bit more today, which is machine learning. Machine learning is a subcategory of AI and machine learning is really when you get right down to it, is a different way of programming a piece of software. The typical way, and I'll use the banker description of it because it's what makes sense to me and I think it'll maybe resonate a little bit, the typical way you would program a piece of software is just like you would do an Excel spreadsheet.
You would have data somewhere in that spreadsheet, and you would write a formula that would tell Microsoft Excel what to do with that data. So, take everything in column A and multiply it by the thing right beside it in column B and put the results in column C. That's a very crude form of programming, right? You're giving explicit instructions to that machine and telling it what to do. The problem is that you have to think of all potential outcomes so anybody who's ever made a very complex spreadsheet has run into those kind of things. Well, I've got something unexpected happening, and now I need to incorporate this other variable that I hadn't thought of before, and so now I have to piece that into my formula.
Machine learning is taking those things and adjusting the order a little bit. Instead of saying, "Here's the data, here's the algorithm, tell me the results," I basically say, "Here's the data, here's the kinds of results I'm looking for, now write me an algorithm." That algorithm can then be replicated and used on new datasets or a growing set of data, just like the one we started from, and we've produced an algorithm that morphs over time because it continues to learn from it, and it gives us a way to get to that answer. A real life example: We all use machine learning in ways that we probably don't realize. Siri's actually learning from us as we interact with Siri, but maybe a simpler one is the way we now store and interact with digital pictures.
When we first started uploading pictures to Facebook, you would put a picture in Facebook and you had to tag people in those pictures. I would take a picture of you and then I would tag it and say, "Here's Jim Young in this picture." If you've noticed when you upload those pictures now, Facebook usually knows who that is. That's machine learning, right? Machine learning is taking the software, so instead of saying, "All right I'm going to program everything about Jim's face into this software program so that every time I see it I know that that's Jim," instead I'm just going to take a whole bunch of pictures of Jim and I'm going to figure out what the algorithm is that makes up, so that next time I see a picture that matches that, I can label it for you.
It starts with some label data, basically I have to have some pictures that I already told this machine was Jim. Then, it will figure out those common things, those common elements. You can do that with any dataset, with any outcome. One more good example would be self-driving cars. If you think of trying to write a software program for a car to drive, it's hard enough to keep it like the right distance from the other cars and between the lines on the road, but then you have to introduce the literally infinite number of variables, right? What if it's raining? What if a dog runs in front of the road? What if somebody's driving the wrong way on the highway? How does the software deal with that?
You can't foresee all of those things, so you have to have a piece of software that can learn, so we don't have to write code for every possible variable out there. Like the way Tesla does it, is Tesla has essentially an autopilot turned on all the time in their cars. Every second, that car is making an assessment of what it would do if it were driving, but there's a human being with the hands on the steering wheel, foot on the pedals. All the sensors see a dog run out in front of the road, and now the human being interjects there, hits the brake and stops. There's an extra set of information. Yes, the human being braked and we saw, "Okay, when we see that, a dog in the road, we stop."
But also, the software has what it's guess was. "Well, I was just going to speed up when I saw that thing. That's the wrong answer, don't do that anymore. Instead do what the human did." You have the humans training the software as they go, so you get that good known label data. You have what the humans are telling the machine is a good decision. You get that across hundreds of thousands of drivers, driving millions of miles, that's writing code faster than we could ever pay enough engineers to actually do it. That's the basic concept that we're talking about with machine learning that goes way beyond just a conversation with a bot.
Jim Young: Okay, well let's take that, though, from the world of Elon Musk, which quite frankly a lot of us think Mars and space age and out there, to the world of banking. Because you mentioned there was some skepticism from your audience when you spoke about it. What was the foundation of their skepticism?
Dallas Wells: Well I think it was exactly that, right? We're running banks, we're not talking about self-driving cars, we don't need to actually build a robot that can function in the real world, and a lot of what they're reading out in the banking press about AI is very retail focused. It's very customer focused, and it's things like a bot who sits on top of your online banking platform and when your customer goes to buy that $4 cup of coffee every morning, the bot says, "Hey, if you would actually put this money into your retirement account, it would translate to this many dollars," and I think a lot of bankers look at that and they go, "Is this really what AI's all about? Is this all there is? Is this what all the hype is about?"
Because I think a lot of them look at that and say, "Okay, maybe there's some value to that, but it's not worth readjusting, ripping out all my existing systems and putting in new ones that will accommodate that" and there's a lot of banks, especially community banks that say, "I don't want bots talking to my retail customers." Then, what about the commercial customers? Can you really have a bot that's talking to your commercial customers about running their very complex business? They're not buying $4 cups of coffee, right? They're ordering inventory, they're managing receivables, they're paying employees, things that maybe you don't want the cutesy little cartoonish looking bot giving them suggestions on.
I think that's where there's this disconnect of, "We see why it works for Starbucks, and for some of the very retail focused kind of apps, but is this really right for banking?" I get the skepticism, but I think there's a lot more ways that AI and specifically machine learning is already impacting banks, and that there's some big options there for them today that are big values.
Jim Young: Let's dive into that. Let's go with that. We've established what AI is in general, we've established where it's being, we're hearing about it a lot and you're absolutely right, a lot of retail oriented type stuff, a lot of marketing type stuff, but are the people, commercial bankers, where is it now and where are the areas within commercial banking that they can make use of it?
Dallas Wells: If you think about again, the concept of machine learning, of taking a dataset and letting a piece of software learn from it, that's essentially the same way that humans learn the banking business, right? A blog post we wrote a while back about why AI makes sense now in banking, about my old buddy Ed, right?
Jim Young: Right.
Dallas Wells: And how he built up a pretty good nose for credit and a way to sniff out bad deals, but it was just based on personal experience. "I've seen deals like this before and here's what happened. I'm not going to let that happen again." You can take machine learning and do that only by several orders of magnitude bigger. You don't have to look at the 200 deals that Ed saw in his life, or the 1200 deals, even, if he saw a bunch. You can look at several million deals. There's datasets like that out there that you can look at. We have known outcomes for a lot of those, "This deal went bad, this one didn't." You can let the software learn what are the things that actually happened there that actually mattered?
Maybe debt service coverage isn't the holy grail of underwriting. Yeah, it matters but maybe there's other things, other factors that are linked to those deals that went well or ones that were linked in the deals that went bad that we need to be looking out for. Let the software learn from the data. That can happen in credit, it can certainly happen in pricing. It can happen when you're running your investment portfolio. One of the big areas where it's making a big impact now, which I think is probably pretty interesting to most bankers is fraud detection. Again, if we have to write code that stays one step in front of the bad guys, that's a losing battle, right?
We will be fighting that battle forever and we will always be one step behind because it's only reactionary. We see something happen, we go, "Oh, man we didn't think of that, we better write code to cover that little thing and to find that little anomaly," so now we've got that one good, and then we wait for the next bad thing to happen. Again, we have to see it, recognize it, find a solution, write code, get it in production, and you're six months down the road, and several scams down the road. Machine learning can again, look at this dataset and see those things happening, find anomalies in more real time, find patterns. Humans are pretty good at pattern matching, but we also have a lot of built in biases that lead us astray.
Machines don't have those, and they are exceptional at pattern matching, way better than we could ever be. The idea is let the machines learn from these big datasets, let them match the patterns, find the trends, and then humans can make the judgment. They can still make the final decisions about things, but they're doing so with a lot of extra context and a lot of extra information. That's the AI, that's a machine taking that human ability to learn from things that are happening. It's not a bot. It can be delivered through a bot, that's just a delivery channel, but it's more about learning in real time based on what's happening, providing that information to the human, to the banker, to then say, "That's an extra piece of information and context that I can use to make my final decision."
Jim Young: Two things here. One, the datasets you mentioned. Those are huge. Those are a lot bigger than my bank. Are you telling me that my data's going to be used by other banks?
Dallas Wells: Good question. First of all there's a lot of publicly available data there from corporate bond markets and syndicated loan deals. There's tons of stuff out there that our machines can learn from, but there's a much bigger pile of data sitting on every bank's balance sheet, collectively. Even some of the very largest banks, we've had this conversation with a few of them, where they say, "Look, part of our scale, part of our benefit of our size is that we do have this giant set of data, and in fact you all should feel lucky if we become clients and you get some access to this big data" which is maybe bigger than some of the other stuff you've seen. The important thing to remember about no matter the size of your own data, is there's some inherent bias in that too, right?
Jim Young: Right.
Dallas Wells: Even if you are the largest bank in the world, the data that you have, those deals landed on your books based on your underwriting systems, your sales process. There's some natural selection bias that happens there, the types of deals you source, the types of deals you approve, and what you don't have is that known label data for the things that you said no to. You turned down a whole bunch of deals. How many of those actually went bad and how many of them were perfectly fine and you said no for no reason? You don't have that to help train your models, so the real value is in this collective set of data, and I think that's what interesting and what banks are starting to grapple with a little bit is you don't have to go hire a data scientist to do this with your own set of data.
Good luck hiring those. Everybody wants a data scientist. What you have is vendors who can provide this, and especially SaaS type vendors. This conversation I had was at a Salesforce users group. It came up as a topic because Salesforce is now all about AI. They've introduced Einstein which is their bot, but what Einstein is really doing is machine learning. Einstein has this giant collection of data which is all businesses who use Salesforce. I've seen opportunities like this before, I've seen patterns like this before. "Here's what I expect to happen, so you can have this projected pipeline," and it's not just based on your deals, your personal experience, but it's based on what Salesforce knows, based on all those deals. They can use this anonymized data. PrecisionLender does the same thing. We have this ...
Jim Young: Anonymized, though.
Dallas Wells: Anonymized.
Jim Young: Because I was going to say, that's the part that might make me nervous here.
Dallas Wells: When we look at it, when you think about which customer it's from, we don't know which bank it's from, and we don't care. What we're looking for are characteristics of a transaction. Same thing as Salesforce. They're looking for sales opportunities that they had this size of contract value, they had this pattern of contact, they had this long between the last contact, was it by email, by phone. There's tons of variables that are consistent, that can be anonymous, and that can be telling about the business that you're working on. I think that's what you'll see, is a lot more vendors who do business with lots of other banks, lots of other businesses in general, and they can take this scrubbed dataset as a whole and get a much bigger picture than any one individual bank can ever get.
And, they can provide it turnkey. If you are a Salesforce customer, you get access to Einstein. You don't have to build it, you don't have to have a data scientist, you don't have to program any of that yourself. You get access to that wealth of knowledge and the latest technology. PrecisionLender customers get it in the form of Andy. We have a bot, looks at that anonymous data, gives insight on deals like this to customers like this, competing against another deal like this, all of that extra contact that we have from sitting on giant piles of data that can be then fed back to the RM's actually working on that deal. "Here's the things to watch out for, heres what'll probably work, here's what's likely going to happen."
Jim Young: You mentioned at the beginning of this final question here, is that in some cases I don't want a bot dealing with my customers here. We should [inaudible 00:17:43] clear, we talk B2C, but B2B, we're not dealing with a bot on the other end. It's a customer there, too. There's a human being for that business, but then if I've got all of this AI machine learning stuff doing all this work, how much of a human do I need in between that and the customer?
Dallas Wells: We always talk about AI being a means of augmenting human interactions. The idea being that humans are really good at working with other humans. That's what humans are way better than machines at, and what they will probably always be better than machines at, so we are a long ways from a C-3PO kind of robot who can just walk up and help us with something. Plus, that guy was pretty obnoxious.
Jim Young: I started to laugh when you said C-3PO. I'm like, "God help us if that's what we end up."
Dallas Wells: Yeah, exactly. I think the idea is that banks need to go from saying, "I know my customers" which means you know their name when they walk in the lobby versus do you know your customers? Can you look at transaction data and say, "Hey by the way, based on our machine learning here, and all the data that we have, do you know that you pay probably 20% too much for your insurance?" That's valuable information. That's knowing your customer and their business and that is insight that can be provided from machine learning, but it's delivered by the human, at the right time, and the right place, and the right way that it makes sense.
I think the role of machine learning and AI in general is to help find that meaningful information. "Here's the things that are most important. Out of this giant set of data, here's 5 or 10 things that are going to be really useful to you." Use the human to decide how and when to actually use them, but they get sorted out for you that way, and you can actually know your customer, give them personalized service but at scale. I don't have to come to your kids' birthday parties to know everything about you. I can learn about how your business operates just by watching the data and having an AI help surface the meaningful things for me.
Jim Young: Actually, I lied. One more final question here on this.
Dallas Wells: Sure.
Jim Young: Because I think one of the questions you got was sort of, there was never a ... When we said skepticism, it wasn't necessarily a skepticism that AI was important, but rather that seems like a 5-10 year down the road type thing right now. Can you afford to wait 5-10 years on this sort of thing?
Dallas Wells: No. I mean, it's been amazing to watch how much it's changed in one year, and as we've gotten into it and started building it into our own platform, the tools that 12 months ago you had to build from scratch and now they're out there as kind of a pre-made thing, ready to use, there is an explosion happening right now in AI. It will get cheaper and better really fast and I don't think you can just sit around and wait for C-3PO to be perfected. It creeps in to our life, it's crept into our personal lives I think maybe more so than people realize. Every time you ask Siri a question, you are training Siri. You are doing some of the machine learning with a bot. Every time you tag a picture, you are helping train that piece of software to recognize that thing going forward.
That kind of stuff is creeping in and I think the same thing will happen with banking where your fraud software will just all of a sudden get a little better. You may not even know how they're doing that, but it's then instead of trying to program everything up front, they're letting the data guide them the right way to go. I think bankers will get comfortable with it sooner than they realize, but you have to dig a little deeper. It's not just bots, it's not just retail stuff. Look at the other stuff that's out there and just know that you don't have to do it all on your own. There's tons of vendors that are going to help you with this, and most of them are just going to be incorporating it into the stuff that you're already using.
Jim Young: All right. Well, that'll do it for us today. Thanks for listening. If you'd like to learn more, visit our resource page at explore.PrecisionLender.com. I suspect we will have a lot of links on this week's podcast page, because this is an area if you've been listening to our podcast that we've been discussing a lot in a lot of different ways. If you like what you've been hearing, make sure to subscribe to the feed for the podcast in iTunes, SoundCloud, Google Play, or Stitcher. We'd love to get ratings and feedback on any of those platforms. Thanks again for listening. Until next time, this has been Jim Young and Dallas Wells, and you've been listening to Purposeful Banker.