Can AI Take the Place of a Compliance Department? American Banker Podcast ยท 11 Oct 2016 Source: https://americanbanker.libsyn.com/can-ai-take-the-place-of-a-compliance-department Transcript: https://podcasts.opennash.com/shows/american-banker-podcast/can-ai-take-the-place-of-a-compliance-department-a4b132/ (via OpenNash Podcasts, machine-generated) Welcome to the American Banker podcast. I'm Penny Crosman, editor-at-large. Our guest today is David Weiss, who is senior analyst at Aite Group. David has been doing a lot of research and he's done several reports on artificial intelligence, or as he calls it, inorganic intelligence. And our topic today is, can artificial intelligence take the place of a compliance department. So in the wake of the IBM Promontory deal that came out last week, where IBM agreed to buy this very large regulatory consulting firm called Promontory that has 600 consultants in its group, these people are going to be training IBM's Watson to handle certain aspects of compliance. David, do you think that down the road we might end up having a humanless compliance department? Not at all. No. Although, you know, from certain, some besieged traders perspective, maybe they already feel they have an inhumane compliance department, although I don't think that's particularly justified either. I think the solution is, and I'll use a phrase that's a little loaded because people a certain age may think of certain things, but it is the term of art. I think the solution is bionic versus replacement, and I think that'll be the case for some time. So for one thing, we are on the edge of cognition, as I like to say. The technology actually isn't there for replacement of just about anything. And if anybody's skeptical of that, all they have to do is call their mobile phone company or bank or credit card company and see how long it takes them to get to a human being. and most of that call routing software relies on natural language processing and some inorganic intelligence technologies. That's on the one hand. And on the other hand, we have so many things where we just need a very smart assist and augmentation where human beings readily acknowledge we can't do it all. We just can't throw enough human beings at the problem. What are some examples of that smart assist? What are some of the tasks that you can see Watson or I think there's compliance is very easy. False positives. False positives are a huge problem. Huge. And everybody knows this. Now I'm not saying all the hires in compliance in the last year or two are related to false positives. No. But it's also fair to say that during that time, throwing all those compliance people at various problems didn't help with false positives. Not at all. In some cases, it was the reverse. And on a case-by-case basis, you can kind of figure out what happened where and why. But the systems in place are, shall we say, optimized to handle that. And typical statistical analysis and whatnot really doesn't cut it. And the firms that provide these technologies are just as aware of this as the tier one and tier two firms with all these false positives to deal with. They're aware of it too. So that's why NASDAQ got together with digital reasoning. very straightforward so there's actually very public partnerships or transactional activity and what not predates Promontory and Watson that indicates that there's some use cases here that we need to address and we've got to come up with a better solution or partner with somebody else so false positives are a big one certainly the case of trade surveillance on unstructured data, particularly involving email and chat communications. And, you know, there's always been chat parsing and scraping for front office activities. But you really need natural language processing and machine learning to do near real-time anomaly detection. And that's certainly one of the things that Credit Suisse and Palantir have in mind with their efforts. So AI could help catch the London whale, say, or people trying to game LIBOR raids or things of that nature? Well, there's many means of catching them, but that's certainly a way of compliance catching them. but in the case of the London whale I mean that was just poor risk management and it seems not even paying attention to the traders who were saying enough so that's a different story but in theory when you have enterprise risk management in a near real time basis the London whale shouldn't happen at all it should be enormous red flags about the disproportional impact in the market, the liquidity risk of getting out of those positions, etc., etc., etc. It's much simpler than that. In the case of the rigging, yeah, I mean, the truth is that data has always been at these firms for a decade. Chat logs have been shipped for a long time. And I mean chat logs from everybody. AOL, Bloomberg, you name it, you know, at sell-side firms with record retention, worm storage requirements, most of that stuff has been persisted. The chat logs are shipped at the end of the day, and this is the most rudimentary stuff that's been done. I mean, more advanced stuff that's been done, like Bloomberg vault compliance offerings and whatnot. But at the most simplistic basis, the logs are shipped at the end of the day, the firm gets it, and they may parse it or they may just persist the text file. and they persist into their worm storage. It's just that nobody's been looking at that stuff, really. It's really there for investigations, I suppose. There are ways of loading that stuff into a database and running reports and whatnot and doing keyword searches. But really, to look at that stuff on an ongoing, near-real-time or real-time basis, you need machine learning. You obviously need natural language processing, and those are two of the dozen or so components of inorganic intelligence technology. Someone was telling me the other day that they feel that one outcome or one thing that IBM and Promontory are trying to work on and that other companies are trying to work on too is just ingesting all of these enormous sets of regulations that come out. Like I think Dodd-Frank was more than 3 ,000 pages. Have the software read all of that and parse out, you know, what, parse these things out into, like, absolute rules or instructions. I don't believe that for an instant. Oh, really? I know. For one thing, there's a whole industry, there's a whole legal industry. There's lots of compliance. See, that's not just compliance. That's compliance lawyers. and there's actually a lot of outside opinions on the nuances and the edges of that stuff. So in as much as you could have digestion of that, you're still going to have just as much legal opinions rendered on that. And, you know, I mean... So this isn't going to put what you're saying. It's more than like, I mean, I would say ingestion of ISDA contracts, yeah. Ingestion of new regulations. And again, you've got in the U.S., just in the U.S., in the U.S. in particular, where we are on the leading edge of 2009 G20 agreement implementation, you've got multiple bodies with multiple time frames. And, you know, the SEC is so behind the CFTC in many areas when it comes to swaps, for example. But you've got so many regulatory bodies, and first you've got the legislation, then you've got the regulation, then you've got the self-regulatory organization's tasks. So you've got multiple levels, and they're all sequenced in succession. And the NFA can interpret something, and FINRA can interpret something else. And not everybody uses FINRA, by the way. There are some people who do their own thing. So you've got all that, plus you have legal decisions. And I don't, personally, I don't think the technology is that far along, number one. But number two, one of the things that I identify in this stuff, some of the challenges, are very low tolerance for failure. Are low tolerance for false positives, which is discussed, and low tolerance for false negatives. Tolerance for failure. So in the consumer world, I would say maybe a 10 % failure rate is acceptable. Although if you're ever trying to reach a human being at a call center, it's frustrating if borderline acceptable. In business to business, maybe it's on the order of 1 % failure rate is acceptable. But in that particular use case, as SEC Commissioner Stein said, how do you certify an algorithm to the fiduciary standard? How do you certify inorganic intelligence technology to have met a much higher legal standard in just making sure this is what the rule is and then how are we going to meet it? So it's just too much to risk. And of course the lawyers will tell you, are you crazy? You can't replace the lawyer. and I don't think anybody's I mean people are talking about you know shooting the lawyers going back to shakes I don't think anybody's talking about replacing a lawyer with hal what about uh anti-money laundering I've heard a few people talk about that that um that's going on right now that's big time yeah but that's not capital markets that's financial I mean that is that can be capital markets but the main areas of that are financial services and uh any times you try and do a wire transfer as a consumer, you may come up against this. And the false positives there are also problematic. If you're doing a wire transfer, it shouldn't take that long if it's going from a pre-existing wire transfer. I mean, it really doesn't take that long. It's a matter of seconds or minutes, really. But if you're trying to do a wire transfer, and this is a consumer I'm just talking about, if you're trying to do a wire transfer, say into an escrow account or some real estate transaction or something you know that can be is it there yet? is it there? okay you sent it did they get it? or did I? you know and the whole thing is supposed to be automated right? but a lot of what's going on is there's false positives in that area too and then something gets flagged for review and it could even be something less than $10 ,000 I've heard of these days where it gets flagged why do they think it's flagged? False positives are a problem there, too. And we're not talking about huge million-dollar transactions. And, of course, when you're looking for structuring, and this is in the retail world, machine learning is huge, as is the flip side of data mining. And to do that and to flag something for review before you commit, before the bank commits to the transaction on behalf of the customer, just for, you know, finding, of course, that's a huge use case. It's a use case for detection, and it's a use case for mitigating false positives. And just to close the loop on that, what is it that artificial intelligence can do when it comes to looking at false positives that a person can't do? Is it the ability to, like, quickly do some research on a transaction? It's speed and scale. That's the best way to put it. It's not that we can't do it. It's we can't keep up with it. So if we can teach a machine to what we know, and then it can keep up with it, and then we can monitor the machine monitoring everything else, which is bionics versus replacement, then you've got a solution. On just about all these, given enough time, somebody can figure out, oh, this is a real problem or this is a false positive, and why do we get this false positive? and then you can teach a machine to figure that out and that's that's one thing and you can do that by feeding it a whole bunch of scenarios and force feeding it to find out or in a more passive way machine learning can figure out when something is actually flagged that it was a false positive a machine can start to predict when something might be a false positive and then taking another iteration out, the machine can start flagging for itself what it thinks is a false positive versus real and then figuring out why it was determined to be a false positive. So there's multiple iterations of this. Does that answer? Yeah, sure. So there's multiple levels of this. Now if you think about what I just described, a human being can do that too, easily. But we have just totally outstripped human capacity capacity for dealing with this in terms of volume and speed. I mean, we're just, these regulations, some of which may be very reasonable, are creating scenarios where everybody is trying to mitigate the risk of shooting one's firm in the foot. And there's nothing wrong with that, right? I mean, if you, for anybody who has a credit card, personally, they understand that. but there's just so many overlapping regulations and there's such a rush and there's such a regulatory churn every time this happens to react. And many people say regulations are not happening quickly enough and that may be true but when they happen there's a rush and there's initially a sense of exercising too much caution these days, too much. and in the same token you can see what's going on with Wells Fargo and it's like you know people knew what was going on there that's not a question of machine learning that's not a software problem so last thing what do you think that the AI assisted compliance department might look like you know five or ten years from now but I I gather there'll be fewer people well there's going to be fewer people um and let me just preface by saying that you know the trend has been there's a 20 to 30 % headcount reduction in the last maybe two years, and it's mostly been in the front office. And in the place of those guys, and if you ever saw a view of UBS's trading floor in Stanford and what it looks like back in the day and what it looks like now, you're like, wow, there's a lot of space there. But a lot of, not all of them, but a number of seats on the trading desk have been taken up by compliance and risk management pros who now formally sit in the front office standard and literally sit on the trading desks. And in the case of compliance, that hiring has kind of plateaued off, and maybe we're going to see some reduction or some attrition and just not refilling. You're going to have to do less with more, which is the standard, I think. That's a fairly standard thing that's going on. in financial services to some degree, but definitely in capital markets. And so naturally, you're going to have people who are going to have to become more efficient. And there's a variety of points of view of what that really means and what we're really talking about. But let's just say people have to become more efficient. And not just, I mean, just overall in developed economies, it's been widely noted that efficiency has peaked. And some say that's why we have, for example, in the U.S., we have an economy in the doldrums. We just haven't seen the efficiency gains. Now, depending on your political point of view, you may believe absolutely, you may totally call BS on that, right? But, and that's not just, That's in every level of industry in this country, in the U.S. But in financial services, it's very clear that efficiency is the order of the day, and you're going to have to be more efficient. And at the same time, we're seeing these peak human conditions like the false positives, like looking at all the chat messages in an automated way, which is actually really intelligent versus just keyword scanning or something. You need more than keyword scanning. And so you're going to see this in trade surveillance in particular. You're going to see this compliance. And what you're going to see is, I believe, a lot of front office techniques and technologies and algorithms for people in the tier twos and the people who are coming in after the alpha has dried out a little bit. The people who rely on vendor solutions as opposed to proprietary principal trading firms where just as with those, you're able to drag and drop algorithms on a screen and you're able to drag and drop trade processes. This is front office stuff, technology solutions. It's like you can do this something with trading technology. They have this where you can drag and drop algorithms. You can use it with Deltics. And this used to be a province of Chicago firms, right? Mostly Chicago firms. Well, now you can do this in the front office. And I believe that you'll see compliance staff and compliance officers be able to do the same kind of drag and drop of what do I want to screen for? What do I want to look for? You may even see this in the front office and with compliance in terms of trade toxicity detection, internal and external. And when compliance starts getting involved in trade toxicity, then they actually become a value add as opposed to an overseer from the point of view of the people in the front office and desk heads. So you're going to see them being able to actually have a technology solution where they can drag and drop in near real time what they want to look at and set this algorithm out into their internal data feeds and compare that with external data feeds and say, hey, what are we doing in the firm and what's going on and I want to catch this. and that's so beyond the relatively inflexible level of iteration you have in compliance these days that it's just it's hard to fathom and I would dare say there's most probably not a Credit Suisse because of what they're partnered with Palantir which is like two or three years it's not a recent thing it's just recently public the joint venture with Sinyak but I would think for many chief compliance officers if you were to pitch them that today and I'm aware of some tools that can do this today, they would be like, I don't believe you. They'd laugh at you. And it's very much how it was viewed about voice recording and voice transcription a couple years ago. And now those same firms, like a phonetic, for example, they go in and people don't laugh at them and say, you've got to be kidding me. I would say, oh, I want to see it because now they get that it's real. If Credit Suisse, not just a tier one bank, but a tier one Swiss bank, pretty conservative. If they could have the foresight two or three years ago to retain Palantir, and then down the road say, this is pretty good, let's have a joint venture and form a startup called Sinyak, and that's real, and they have an iterative product inside the firm that they're constantly refreshing and sticking out of more problems, if they're doing it, anybody can do it, and everybody should and will be doing it. First it was them, Credit Suisse and Sinyak. Then it was NASDAQ and Digital Reasoning. And now it's IBM Watson and Promontory, and really making Promontory the centerpiece of IBM financial services. That's three different events with different angles to each of them. One is two vendors. One is a vendor in a firm. One is the exemplar of inorganic intelligence, IBM Watson, literally buying up one of the most prominent, no pun intended, one of the most prominent regulatory compliance. There's stalwart. Promontory is a stalwart. You don't need any more signposts to realize this is definitely happening. It's right there in front of you in the papers. Okay, great. Well, David, thank you so much for coming. Really appreciated your insights. We'll talk again soon. Great. Thanks for inviting me. Thank you.