Hello, everyone, and welcome to our webinar beyond the buzz, the rise of LMS and AI in communication surveillance. I’m David Aronson, product marketing at Shield, and I’ll be your moderator for this intriguing and timely conversation. Before we dive in, I have to share something interesting I discovered about some of our panelists and one of our recent conversations together. Turns out we have not one, but two accomplished tennis players, thus today and Stan and Cameron, so you can expect them to serve up some important insights and just like that hundreds of people around the world just drop their faces straight into their hands. Now for a little bit of housekeeping, we have fifteen minutes at the end of the webinar for Q and A. If you have any questions throughout the presentation, please feel free add them in the chat, someone from our team will be monitoring and compounding the questions for our speakers. If we run short on time to answer your question, will be sure to compile answers from the panelists after the event, and we’ll reach out to you with an answer. Allow me to introduce our experts. Will be leading today’s discussion. Cameron Funkhauser is a former Finra senior executive and an expert in financial services regulations and compliance Cameron’s experience and unique perspective on protecting the markets are invaluable. Stanyakof, our next panelist, is the head of supervision, In the Americas, as citadel, an expert in financial services, ecom’s compliance teaches law students on the subject as well. We look forward to Stan’s insights into the practical aspects of AI and e comms compliance. Last, but not least, Dr. Shlomit Levin is Vice President of Data Science at Shield, holding two PhD’s, and now has a very shiny best cloud innovator of the year award from the Cloud Awards, Congratulations on that, Lumi. An expert in AI data science and the financial compliance market, Lumi brings a wealth of knowledge and innovative thinking to our conversation. As we dive into the discussion, we hope to leave you with a new perspective and tools to understand and leverage the exciting developments in AI and compliance. So sit back, relax, maybe grab a tennis racket, and let’s jump right in. So let’s start with a softball. What makes large language models, and chat GPT, so interesting to us from a compliance perspective. I’m just gonna throw this out. Okay. So maybe gonna lead off? Go ahead. No. Just let me go ahead. I I cannot handle less silences very well. Okay. So in general, when we’re talking about large language models and the new generative AI, we’re talking about real breakthroughs in the domain of text analysis or a free language analysis. This is the main challenge in the compliance in the is being able to monitor human conversations. It has been a challenge for many years, and now we have innovative tools to actually deal with it. Okay. Thanks thanks for me. Stan, what you got? Thanks, David. So before I begin, let me just give my quick legal disclaimer that the views expressed herein are solely, those are my own, not any current or prior employer necessarily. So with that being said, what’s interesting here I think is that for one of the very rare times, we see convenience and compliance possibly hand in hand. So when you historically think about how technology has been developed, particularly in the compliance Tech Fintech, name whatever acronym you want to give here space. Oftentimes, you’re going to need to have two PhDs like Schulman. Right, in two different domains, one in psychology, one in computer science. If I got that correctly, in order to even interact with the tooling. Here, all you have to do is really just go on the website, sign up right, confirm via your email. And then you have access to very powerful technology. So, like for — the reason I say it’s rare is oftentimes the tech gets built, and then you have to figure out how in the world you’re going to interact with it. Here you actually have a very easy way of interacting with very powerful technology. So that’s caused I think a lot of us and our much older family members to, all of a sudden, start interacting with robots more than we ever thought we would be. Yeah, that’s true. I mean, grandmothers and grandfathers everywhere actually like going online and getting answers to all their questions except why won’t my WiFi work that still won’t help you. Karen, what do you think? What makes LMs so Yeah. I’ve been around pretty long time. I mean, in the early nineties, I was helping develop some, what was considered revolutionary technology, using artificial intelligence that looking back on is laughable now because we were doing structured word analysis and connecting the dots, I mean, the potential from there to here, you know, you can imagine in your own minds, what what the difference is between AM radio and Netflix. Right? I mean, there are lots and lots of opportunities, probably more opportunities than we could talk about today or even think about. But the opportunity is to be more effective and more cautious and reduce your regulatory risk are enormous. Thanks. So before we continue on, I just wanna pose a quick anonymous poll from the crowd So from one one to five, five being most advanced, which is like, we knew about ChatBT before was cool. So how far along is your firm in implementing AI in your compliance strategy? I know we hear a lot of questions about it, a lot of ideas get kicked around, but really curious to hear, and I think the audience really would like to hear from each other kind of like where everyone else is holding. We’ll continue on while people still answer. Stan, let me pose this question to you. How far along is your team in your AI journey? And so the one is in five. Yeah. So I generally don’t put a number on this, but I’ll talk about experience. Right? Specifically my own here. So, the way I look at it is I’ve employed certain, like algorithmic techniques that I call them. Right now, that function to essentially cluster trading activity. Which gets an idea for like what typical or normal trading activity looks like and analyzing activities that fall outside of these normal distributions over longer periods of time like thirty days, sixty days, ninety days, because generally what I see is that the longer horizons tend to help reduce noise from just an atypical market day that could be characterized by like isolated news events like more recently interest rate bumps and such. And I think this technique helps illustrates where all maybe behaving anomalously? Like for example, trading more than usual and incorporate other important dimensions like P and L into this type of analysis. As well as like whether the trading activity may have triggered trading controls that cause rejects and stuff like that, which you generally don’t want to see or if an algorithm really behaves as it should be from the start, you really shouldn’t be seeing this type of stuff. So once this occurs, right? From a detection point of view, the curiosity now goes on the investigation side and you start to ask yourself like what may be the driver for the heightened activity? Right? Once you isolate out expected factors like news and contract expiries, rollovers, like again things that you expect. And what’s pretty cool is like you start to identify unusual market data movements that may resemble possible market abuse. And what’s interesting is you see this externally in the market like this isn’t stuff that you have control over internally yourself, which really means that challenges now you’re a victim in what’s going on. And it’s much more challenging to control again, because again, this is happening in the market. So, If you really think about what’s going on, you use such creative feedback loop between something that surveillance found. And really the only thing you could do is just try to be preventative from a front office a trading perspective. And to think about like, hey, how can we identify gaming ourselves and try not to interact with this type of activity once we see this pattern. Forming, right? So, this has been pretty interesting, but I still consider this to be like experimental and really a supplement to, let’s say, non AI based algorithms, as opposed to some kind of outright substitution or replacement and I’m going to talk about that aspect of like being very careful with the word substitution or replacement quite a lot here. So just given that warning. Now if we turn to like communication surveillance, right? I think this lens of supplement, which I just gave that disclaimer for. And experimental is a really important one to analyze this domain through because From a communications perspective, like where I personally see technology like GPT having the most value in the near term is where you help provide additional information rather than draw conclusions. So a good example from a technique perspective is like topic discovery or essentially summary statements to assess quick themes from otherwise very lengthy, and I’m sure, legally, cleverly worded text. We’ve all been there, right? So visually if you think about this this is like a kind of like a word cloud where certain words are emphasized in size and font, right? I think many of us have probably seen charts this and found it quite helpful because the words jump out at you right away based on essentially their importance, right? Now, topic discovery has its own dimension such as like who’s the conversation with? For example, a client is different than a counterparty, than someone else internally from X YZ team in your firm. Secondly, like what is the tone of the conversation, right? So a complaint is very different than like angst or uncertainty over choice versus like a very passionate debate on one of the most optimal lunch choices around me. Right? And when I look at how regulators have thought about this topic too, like there’s actually a speech from January twenty seventeen from SEC’s DIR Office, the division of economic and risk analysis. Where they give an example of applying another technique called LDA, latent direct allocation to SEC register disclosures and filings where they’re looking for possible emerging trends. And like the example they gave like the study essentially was, Now that you have hindsight bias, could you have predicted the rise of credit default swaps and the effect that I had in two thousand and eight. Right? And what they saw is that In nineteen ninety eight, if I recall right, there is about three references by banks and there are ten K filings. And then six years later, there’s over one hundred references. And then between two thousand and seven, two thousand and eight, there’s like a tenfold increase, right? So it’s interesting from a topic discovery point of view because now you’re finding these emerging themes and words or phrases or possibly products you’ve never seen before that themselves carry their own risks, right? So from that topic discovery point of view, I think that there’s a natural segue into, well, what’s the tone of the conversation? And applying AI, NLP to provide observant assessments of emotional context rather than just the substantive context. You know what I’m saying? And lastly, I’ll just take Yeah. Alright. Yes, I’ll just say a couple of quick words unlike the importance of proper governance and then kind of hand over the mic on this because I’m mindful that I introduced like several new applications or kind of experimental applications again as I use that word. And I also want to caveat it with really strong importance on control and governance. So all throughout, I’ve mentioned this is essentially applied on a supplemental basis, not a kind of replacement or conclusive and therefore consequential basis. So, like having the technology and techniques be aids rather than conclusive without any oversight or supervision, right? And thus being consequential, it makes all the difference. And I know like we’re going to talk later about the nuanced risks, but I wanted to kind of throw that careful balance and way of thinking about this earlier on. Yeah. For sure and I’ll hand over the mic back. Fascinating. Thanks then. I I think it’s fascinating just like what you just described being at still the experimental you know, states still like discovery of LMs and what we can do with this technology today. That’s just reflecting back just looking at the results from our poll. I mean, it looks like there’s there’s nine people who are at you know four and five and then like thirteen at three and that’s really a lot of people are still at one and three, so it looks like people are still kind of figuring it out and it makes sense kind of what you describe as being in the experimental stage people aren’t not necessarily gonna call it a five day because it’s gonna take a lot to get there. Cameron, now you’ve heard what Stan is doing. As a former Finner regulator, what do you think? Are there any coverage issues there? Well, I would start with you know, from a regulatory standpoint, there’s a baseline rules and regulations set that you have to comply Right? So I don’t think that regulars are too interested in who’s experimenting with what technology as long as they get coverage. So that’s the baseline. Right? So when when Finra or the SEC or another regulator comes to call in, they’re gonna say, here’s the rule set, how are you how are you complying with the rules? Unfortunately, in my experience and I I should also caveat it by saying, I don’t speak for Finra. I I’m Finra Jason as a being recently retired. I think the the regulars in the US generally do not give great guidance on what to do. Right? They don’t say, here’s a best practice. Stan’s doing this at citadel. You might wanna give him a call. Right? They say, here’s the rule set, you figure out how to to comply. We’re gonna come in and kick the tires and tell us, and you can explain to us what you’re doing. And you can explain to us how it’s compliant, and we’re really interested in doing that, and we’re gonna hold a conference in, you know, once or twice a year, we’re gonna have a roundtable about what everybody else is doing, but they don’t give you the, here’s the best practice that sort of It’s, you know, it’s sort of like they wait for the plane to crash and then say, okay, you’re pretty screwed up. That’s that’s the way Regulent story world to work. So that’s my reaction to someone like Stan who who well, when we go in, he’ll we’ll interview him, and he’ll tell us everything they’re doing, and we’ll match it up against the the rule set. And, you know, that that would be on a routine exam. You know, a lot of these and I come from a dark place because I really ran most of my career doing fraud investigations is you’ve gotta have your coverage, but you’ve also gotta have some system in place or techniques to identify holes in your coverage. So in your lexicon, for example, using the AI or other types of technology is what what word was just introduced into a conversation in your investment banking department? So if somebody starts throws out a word like potato chips or, you know, being canned, that could be an interesting one off. So, you know, it’s complicated for registrants and others to follow the rules because they’re ten thousand foot level. And I and I will say this, the regulators are not expecting perfection. Absolutely not. So I think reasonable systems, techniques with the types of business you’re in or what they’re looking for. And you should be able to explain that in very sort of common fifth grade language, if I was giving advice. Because you’re gonna overwhelming the regulators with technology and tech talk is gonna probably go over their heads. That’s not to say that Finra and others aren’t tech savvy, they are. But the people who are actually enforcing the regulations are gonna need to have explained to them in compliance oriented language. That makes sense. I mean, people are people and I’ve heard you loud and clear when you say that regulators just one assurance that you’re working on risk, fashion, perfection is not necessary. So from your perspective, what opportunities do you see for AI in e coms to ultimately reduce market risk? Well, obviously the more data you have and and the better techniques and are gonna give you better results, right? If you had to hire someone to watch the market all day, think about it that way, which I did early career, we would literally watch the market with a very rudimentary computer. It’s impossible. So, you know, the more technology you can throw over it and the better data you have and with governance, I think you’re it’s unquestionably gonna add value. It’s just a question of, How do you tune that engine to produce meaningful results? And I think Stan spoke to that. A lot of surveillance systems they’re very sensitive, and they produce a lot of noise. So you end up eating up a lot of resources trying to tell yourself that’s not important. And more importantly, from an audit standpoint, regulators are gonna wanna come in and see the results and see what you’re doing with your anomalies. Right? So that’s another area of concern for regulators is what are you producing? And how did you resolve it? What’s your audit trail to resolve it? I mean, that’s not that’s not uncommon for anybody on this call, I’m sure. But you know, the better results you have, the less time you spend resolving things that are not meaningful in documenting them and producing them, and explaining them to regulators. Thanks, Karen. Shlomit, you might be the most like smartest person in the room here, a lot of PhDs lying around. So I mentioned you have a lot to say about what’s being said out there about lens and their uses. So, in your perspective, is AI, chat uPT, large language models, are these just buzzwords how is AI truly reshaping ecom’s compliance? Okay, so definitely not just buzzed words I think that the main effect that occurred in the world since tragedy was revealed is that now the every person can feel for themselves the capabilities of AI. We’re talking about large language models. They have evolved over the past few years and they exist already for, I think, three to four years in the market But now, since the emergence of Chachi pitits with the general public, everybody knows and feels how good AI solution can be. And it was shocking, I think, all around the world. You all hear the news. You all hear the overall communications in the newspapers talking about the effects of tragedy on the world with different kind of predictions of what will happen, with it, but I think that the effect is true and meaningful. The technology is legit. Having said that and referring to the poll that you just made, we can see that it is not easy for organizations to adapt to this technology, you would say it already exists for six months, and still Most organizations are somewhere in the middle in adopting new and advanced AI. And some of it is related to what Cameron was explaining about the heavily regulated industry and the way that the system are being evaluated and needs to be explained to the regulators on the one hand. And it’s how to make those transitions. But on the other hand, and I truly believe in it, since people are getting used to this technology, it will be something that will be a must have in any kind of a future solution. Because if you can do it in your own home, if your grandmother can ask something, tragedy, and get an answer if you’re a kindergarten child and they do already ask something on cell phone, and get the answers for why can’t you do it on your communication data? So from my point of view, I think that definitely this is something that will get a mandatory part on any future solution. I can also refer to everything that Stan was mentioning regarding the challenges that this industry has. And we are saying that, first of all, of course, the volume of electronic communication is just growing with the growth of technology, and the abilities to solve it with the older methodologies is just not available anymore. But again, it was referring to a different challenge. We now, with the ability of new technologies, want to solve more challenging problems that we have. Why do we have to satisfy with detection of all kinds of compliance issues. We want to also be able to predict them. This is a more sophisticated and more saddle issue to solve, and you need more robust tools to be able to do that. So for solving such challenges, you will definitely need to adopt more sophisticated technologies. Thanks, Lemit. Reminding everyone, if you have any questions for Stan, Cameron, Schlit, or just in general, just hit the QA button below and shoot that question over. Cameron, if it’s more than just a buzzword, I’m gonna assume that we’re gonna be hearing more about regulators responding to merging industry uses of AI and monitoring. And I understand it’s not necessarily gonna be and this is what you should do. But like one of the things we hear a lot about is explainability in the models. We wanna make sure we can talk to regulars about what we’re doing. What are your thoughts? Well, I think I choose think explainability is you know, an important aspect of communicating with regulators. You know, like I said, you’re gonna have you’re gonna need to explain, you you know, what you’re doing, why you’re doing it, if you’re if, you you know and how are you what what data governance you have, what data sources you have, and these are all out there. I don’t think I think you will find that regulators will be skeptical. Right? You’re you’re born skeptic when you’re a regulator. So when you come in and you start asking questions as a regulator, I think they’re gonna wanna actually see results. And let me just mention something about predictability of compliance issues. That would be amazing if you showed examples of that, because I think what people tend to do as regulators is they’re responsive to a crisis. If you can get ahead of a crisis and report that to a regulator, amazing. Because I always operate that, you know, a problem that exists particularly with a product, particularly a product that may impact retail investors, is a problem that exists across the industry. So if you can get ahead of it somehow and protect other firms from that problem and particularly protect investors from that problem and and give the regulator the heads up. Absolutely amazing. And I would encourage you to do to use your techniques. If you can get there to identify things quickly, eventually, these problems always emerge because you know, it’s a financial industry. Right? Money’s gonna be lost, and ultimately somebody’s gonna be harmed. So, again, ahead of it would be would be fantastic. So I do I do think Let me raise a good point if if the technology gets to the point where you have some predictability, but you know, I always operate with is you don’t know what you don’t know. Right? And so that’s why I say, expectations from regulators are not perfection ship. But if you improve from not being perfect to better, fantastic. Yeah, I like what you said there about that shift from playing defense to playing offense if we can take a new approach that allows people to know what’s going on in their own organization versus finding out about it from someone external. That’s a major shift. Shlomit, implementation concerns are certainly something that we hear from our customers. Can you speak to why implementation is so difficult with these models? Yes. So one of the first challenges is that we have to several of the newest solutions achIPity included, but also the Google models is a security issue. Can we send our data over to open APIs? This is a major concern to some of the financial industries. But I think that these issues will be solved, the new models that do not require such APIs will emerge eventually, or some of them are already out there in the market and are being experimented with. The second main concern, and this is something that Cameron as they refer to is, again, the explainability And it is important not only for the regulators, but again the phrase you don’t know what you don’t know is very is up to the point here. So I ask, for example, just because the users probably are familiar with it. But I ask Chachi Preeti and I get an excellent answer. Is it correct? Isn’t it correct? What do I know and what don’t I know? Models are including a lot of unknown in in each side of them, and some of the new ones are also are able to actually invent information, so this is something that we should be very careful in using small story. I have, San Who is working toward his final exams. In high school, he was requested to produce final work paper in history, and his teacher told him, you know, you can use Chachapiti if you want to, but please beware because if you get any big biography references, you should double check them in Google, they might be fake. So we are in the world of actually having a great technology, real abilities of doing things that we couldn’t do before, but we really need to have in place all of rakes and balances being said to make sure that we know what we have, we know what we miss, and we know that what we are producing is actually correct and valid. Sounds like Chad GPT also picked up on that human trade fake it till you make it, that that is a very interesting example. And and exactly, I mean, with all new technologies, some of any technology in general, you know, and I think all of you have kind of like hinted to this at one point, the the human part of the use of technology, it is tool to use. It is a supplementary tool, and at some point what is it doing for you? And how can you prove that value and defend that? Thanks for that for that storage. That’s a very good point. Stan, we’ve all agreed that there are opportunities for AI and surveillance and Karen mentioned the skepticism necessary as we move forward at breakneck speed. What controls do you see as a key for this forward motion? Yes, David, since you kicked off of like a tennis analogy there, I’ve kind of been thinking how we circle back to that in the spirit of your nicely timed joke there. And going to the question just asked for me and I’m like where’s the pain coming from. I have a colleague who likes to change his racket every single week. He’s been doing that for about a decade. His game has not gotten any better. At some point you got to face the reality that like the answer is it’s not so much the problem with the racket or the tooling, it’s probably more the player at that point. And I think a lot of the pain points you’re hearing is a problem manifest in the firm itself as opposed to a vendor technology. And I think this is really important because like the vendor can’t magically fix the problems that you have in internal to your firm, oftentimes doesn’t even see that until kind of the initial engagement or later on towards the point of integration and then teams themselves are kind of clustered in realizing like Oh, wow, how do we fix this when our firms have been around for fifty plus years, right, or even a decade alone in the last, you know, relative to the advancement of technology in the last decade itself is going to be equivalent to fifty years, right? And so the control aspect really gets to the heart of change management, which I know is not pleasant to hear because it arises in essentially every facet of organizational operation, right? And like when I look at the way that a lot of these algorithms operate, particularly in the AI space and what this change management had on. We saw this case probably in the last couple of weeks where lawyers now facing sanctions for using GPT to basically write a motion in court, right? And the citation cited, similar to Scholomeit’s warning from professor there were all made up. Some people refer to this as like AI hallucination, where it just kind of made up the references. But I want to be a bit careful with like using more buzzwords, because there’s a lot of hype in this area. And I want to get down to like what this really is, which is a design problem. So when you think about control, you really should be thinking about design and transparency into that design. And essentially what’s going on behind the scenes, like in layman’s terms is that The algorithm probably has some sort of decision path regarding what do I do when I’m not sure. Right? And I need to have an answer. And that assumption of need to have an answer itself is already pretty dangerous slash important, something that you want to ask about, because what do you do when you’re not sure? One answer is good to the human. Right, which is kind of where you get reinforcement learning or self supervised machine learning, right? Or semi supervised machine learning, sorry. And there you essentially have an exception thrown that says, hey, I’m not sure, or here’s my confidence interval on this response, which allows a human to dig into. Right now, none of that really exists. The machine just continues to operate, right? And essentially, it doesn’t want to fail over. The other side of the coin is that where the machine continues to operate and you end up getting an answer, right? If you take it for granted, now you can kind of see the types of issues that arise. And from a communication surveillance perspective, none of this is actually new. It’s occurred, maybe just a matter of now you’re coming to terms with the reality and kind of these hidden risks, so to speak. But some very classic things that happen in reality and in that communication surveillance context is what happens if you get someone with a hyphen in their name? When the algorithm thinks that names are just essentially John Space Smith, right? Whereas we know people get married or just people have hyphen names as a matter of just you know, inheritance. Another aspect is what happens when you get two people with the same name, right? Does the algorithm think, wait a minute, I’ve already processed this before. Or is this an update to the first record? So, let me just overwrite it. And consequently now that second person or possibly the first person, depending on the decision path. Is no longer supervised, right? And you have a much bigger problem from there. So the really important thing is that you essentially have to study the black box and really focus on this from a change management perspective to understand what is the initial design and what are changes to that design? And look at the testing conducted or specifically unit testing, integration testing, Unfortunately, a lot of people don’t like to hear this, because ninety five percent of the work is going to be about writing tests, and the other five percent gonna be the fun optimization, but this is the reality that you’re in, especially when you’re in a very regulated business. One other thing I think you asked Cameron this question about how do regulators look at this, What’s unpack the word regulator? Because it’s no longer a Finner SEC. I don’t even care if you’re, that’s the only person you’re regulated by. The reality is that the FTC is looking at it, the Department of Labor, pretty much every single state regulator from a data privacy perspective. If your product has any retail facing arm and utilizes AI, especially like from insurance credit banking perspective, that’s going to be in scope. So now you kind of have to study the views of approximately I don’t know one hundred plus different agencies who are all focused on this. And if one agency is first or more progressive in this regard, even though they’re not your regulator, I can guarantee you your regulator is looking at them too, and then you’re gonna start to get the same question. So, it’s really paramount to be focusing on this, like day by day because that’s the pace of evolution. That’s a really good point stand. Thank you. I want to shoot the conversation to a certain, you know, different elements of the practical approaches that we can take Cameron How should firms approach potential conflicts of interests when implementing AI in their surveillance systems? Before I answer that I do have to tell one quick story that Stan made me think of or this is the danger of of building these systems is we started long, a long time ago. There was message boards out there that were talking about stocks. And the lexicon was, you know, we were trying to figure out. So we said, let’s put every symbol into the to the system and see what pops out because very concerning about these message boards that were popping up. And we ended up with about ten million yard sales because one stock was YARD. So the results so you have to be that’s a good point stand makes is you get, you know, your lexicon has to be careful because your machines will just you know, they could produce results that are so overwhelming that they’re useless, right? So the hyphen, and also with lots of similar names, you get every John Smith in the world. So That’s something everybody knows, but I’d like to reinforce it. And I know I’d found it out painful, and it’s an expensive lesson to learn. In terms of conflict of interest, there’s a lot of opportunities, you know, especially in the communications, surveillance area. Your your data has to be set up. Dependent kind of the type of firm you have. Right? So if you’re a full service, retail investment, banking firm, I mean, you’re gonna have buckets of people that are going to play different parts of the game. Some of them are gonna be regulated differently. In the sense of communications because of information barrier and the sharing of information across the barriers, obviously improper I think there’s, you know, investment banking communications. Clearly, you have a lot of material non public information that’s floating around, I’ve done probably forty thousand insider trading investigations in my life, and I can assure you that continues to be a very target rich area for for regulators. You know, on the retail side, I I’ve been reading about AI that is, you know, doing suitability analysis and targeting messaging to clients on products that might be suitable for them. You know? Sounds good, except until they’re unsuitable, right? And you find that out through an arbitration or a lawsuit. I think that’s the conflict there between the data that’s being fed and the actual customer has to be, you know, analyzed properly and and scoped properly. The other thing that I would just mention is, you know, just as a warning, and I again, I come from a dark place, and I apologize. Is one hundred percent no matter what you’re doing, there are people out there possibly within your own domain trying to defeat it. Right? And you that’s why I say perfection is not expected, but be be aware that the insider threat is real. I mean, people I know always talk about external hacking and things like that. But the but you have to make sure that you have governance around your policies, procedures, access to prevent insider threats. That could be That is a huge conflict. If I had advanced knowledge of, you know, whatever citadel’s trading big today, I’d probably not be sitting here. I’d be trading. So there are lots of conflicts, obvious ones that people know about when I say it, but they’re like, oh yeah, I didn’t think about that. So Don’t don’t overthink this. Right? Don’t overthink conflicts, dumb it down, and say, literally, what are some sort of what would the common person want to steal from us? Good answer. Thank you, Gavin. Stan, how have you dealt with AI integration? Are there any potential gotchas or challenges that firms should be aware of when they start to integrate AI into their operations? I don’t think it’s a gotcha. A lot of this is pretty anticipatory from the beginning. So I just previously mentioned like don’t blame the racket, like blame the player, focus on yourself really to improve your own game. And if you think about this internally, like Or actually, let’s go back to the SECD era speech from, like, I think, January twenty seventeen, so this is years ago, right, before GPT, where we are today. In that speech, I you would think the answer to this question is going to be like, oh, it’s very hard to kind of code these clever algorithms to perform this hypothesis and get our results. The answer is actually much simpler and more surprising. It’s if you think about registration filings, they’re still on paper. The answer is we should get this electronic. And then we should probably have a uniform schema for how people provide the registration or, you know, corporate filing information. Really simple answer. Kind of goes back to like when people look at pyramids, they go, wow, that’s such a beautiful structure. How did you create that? Ninety five percent of it’s like laying brick by brick, right, in a very careful deliberate manner. Pretty mundane work, but very important from a geometrical and otherwise a consistent point of view. And the analogy I think applies a lot to firms internally. So like you know, in my opinion, the preparation and kind of safekeeping of data in a consistent manner is just ninety five percent of the work and the remainder five percent What I call optimization is the kind of AI answer people typically expect to your question, David. And I think a lot of firms overlook this and rush to vendors looking for explanations on their own data problems. And as I said before, vendors don’t have control over that, right? So I give this example that’s like pretty simple. Like, go to HR after this and ask them, how many different titles do they have for people in the firm? Right? If you have a company with fifty people and you’re getting an answer of hundred, and that’s pretty problematic, right? And even that, even if you get an answer that you think is like relatively okay given the size of your firm, like, tell them you want to analyze the communications by or you want to find all the people that are traders, right? And analyze their communications because a regulator is asking you or you want to do this internal audit exercise. See what they come back with. Right? If they come back with like, hey, we have twenty three different titles for the same function. Again, that’s a problem. And interesting question to ask you is like who actually owns that data? Is it HR who’s maintaining like HR who’s maintaining employee records and doing maybe some database maintenance? Is it the business who’s like hiring these people? Maybe determining a role, is it the person writing the contract? Like, who owns that? So, at minimum, you probably want transparency on it because if you’re going to use that data, either now or down the line, right? And very likely you are going to use it down the line. You have to know whether it’s in the format you need because if it’s not, it’s not going to be usable. And vendor is not going to solve that for you. Right? Probably at a maximum, you may want to insert yourself into the decision making process for kind of how information like this gets us cited. But doing so in a commercial way is also pretty difficult. Right? One other thing, like, I guess, more of a story from the other world’s traveled is it’s really important to map out like dependencies or subbenders. So to speak. So right now we’ve just kind of been using the word vendor, I think for granted. When a communication surveillance vendor comes in, they’re going to be survelling messages like Bloomberg, right? That’s a whole separate vendor or so to speak a sub vendor, right? And similarly for iPhone messages, WhatsApp, like each of these carry their own respective data policies and attributes. So, this then creates multiple points of failure. One, it could be the organization, two, it could be the surveillance vendor who’s ingesting all this and and Audix three could actually be the underlying sub vendor here. So the most important thing is, or, you know, if I leave you with two words, from this entire question is data provenance, which is understanding like where’s this information coming from? How does it travel? Who’s every single person that touches it? Where did it change? And then if it changed, especially to a way that’s now incorrect, how are you worded? When are you worded? All right. I kind of hate to like give fifteen questions in response to a question, but this is the reality and welcome to having to build a pyramid. Yeah, that’s a great answer, Stan. I mean, we about Alec, I think Karen said like you don’t know what you don’t know, but you definitely don’t know what you don’t have and can’t make sense of that’s a nonstarter. So we’ll need with all we’ve covered so far, I know this is enough to keep us all busy, but I’m curious, any thoughts on AI in terms of like, where is it going to help us? What’s it going to help us do next? Okay. So maybe we can start with even keeping us up in the same level as we used to be a year ago. Two years ago, we will need to use more advanced technology. As mentioned first, the volumes of electronic communications is constantly growing. So we have to have in place more accurate systems. We cannot constantly increase the amount of people that are reviewing the false alerts all around. Second thing is that — and Cameron mentioned that we are dealing with people who are trying to eventually the ones that we’re trying to catch or trying to manipulate the system. So let’s say that you have in place a kind of lexicon solution you put something in place, for example, all kinds of phrases don’t text me, don’t WhatsApp me, or anything like that, It will take one month, two months, the surveilled employees will be familiar with those phrases. They will vanish from communications. Will this behavior stop occurring? No. But they will make sure that you cannot monitor it in the exact same way that you used to before. That’s why our solutions need to be growing solutions, solutions that are subtle to and this is the great advantage of large language models, settle to nuances in languages that are able to detect different phrasings of the same idea or the same notion or the same action that is being taken place. And do not allow for humans to quickly cheat the system. So I think that as a first ground rule, even to stay in the same place that we are today, being able to detect the same risks, we will need more advanced technologies. In terms of the future of AI, I think that What’s happening in the world today is really a revolution. And the the main thing that happened since the I think the manual multiple use of are activities that people are getting more and more used to being able to use free language to question complicated systems. So if I have a compliance officer, I don’t want any more to collect set of rules or set of lexicons or connect to different databases on my own or set different queries just to be able to get information. I want to present the question in a human manner, and to be able, we call it chat with my own data in a free text language. And I think that the technology is going to enable us to do that in the very near future, if not even today. This is one thing that is definitely going to change because people can feel that they can do it, and the technology allows it, and it makes life so much easier for everybody. The second thing that I think the next generation is, again, for models to evolve, and this is already happening, to be able to do more than just answering questions, to connect different tools and systems. So if Stan was saying, I need to put in place a sophisticated system to detect the vendor, the data source type, HR data and everything, instead of having manual work to do to put in place there, The different models that are doing different tasks will be able to talk one with another. So I will be able to chat in a free text, but also get something to create a sophisticated query or a sophisticated rule for me and collect the data in a manner that I need it eventually to occur to me. So to sum up, we are facing really exciting times, a real exciting future, and it’s from my point of view, great times to be in. Very cool. You’re in the front row seat tall this, so that’s really cool. I have a question for Shalomit. How how do you deal with people who are using emojis or pictures or something that’s not language? Have you have you is that a way to is there a way to surveil for that and is it meaningful? Oh, definitely. So emojis are part of our language, the fact that I would say old fashioned solutions are treating words as a trigger It’s just one manner of looking at the data. The more advanced models we look at definitely on emojis and their we are doing it ourselves as well, of course. But on all kinds of data sources, if we have pictures included, if we have voice included, have all kinds of hints and layers and types of information added to any kind of communication, they should be analyzed with a way that with a model that knows these kinds of input data. So, yes, and this is becoming more important because we are all in from one generation. But when I see the younger generation, They are not talking in words anymore. They are talking in pictures. They are talking with emojis. They are talking with a lot of other symbols that are different than what we used to do, and it definitely needs to be incorporated in our solution, in any kind of solution. Very cool, Cameron. We got a question here from the audience, how can firms ensure AI and LMs are used ethically in compliance and surveillance? Oh, that’s such a great question. That’s a really great question. The answer is I don’t know exactly how that could be formulated, but you have to think about it from a standpoint, and it would obviously be an important aspect of your employee agreement that you would act ethically and only in the business interest usage is Right? But if you think about it. Right? Let’s let’s use Stan’s example of the SEC did some lexicon search or or I saw some trend in in filings that Accredited Vault were mentioned. So let’s say, in your own world of communications, you had access to a trend in a product or a or a target, or some sort of investment. And one of your employees leak that information to somebody. Right? That would clearly be something that would be unethical Right? But I don’t think people really think about it that way because in today and the the modern or or the the less modern world is You know, ethics is, you know it when you see it. And so I would think you would want to sit in a room and you’re never gonna cover anything what to come up with with practical, ethical guidance that doesn’t you know, it’s is it, like, the regular thing? We we know it when we see it. I think you wanna think about your business, what what the vulnerabilities are in terms of opportunities, for employees to extrapolate information and use it improperly, and and I and and train them. I mean, you can’t You cannot over train people, and there’s enough real world examples of misfires cases brought by the SEC, FinRA and use those to train people and say, you know and and and I think and I do think the regulators are impressed when you have docu you have a training cycle document what you’re talking about, and it’s relevant, not just the stale sort of bring in some guest speaker who’s talking about stuff that happened three years ago, five years ago. And this stuff is very rapidly changing. So from an ethics standpoint, it’s a really good thought provoking question that everybody should be thinking about. And the other the other thing, you know, that always comes to mind for me is people who are leaving your firm, Right? What are their ethical obligations? What should you be looking at for? I mean, you have this concept of bad levers. Right? People who are leaving, for whatever reason and they’re going to cause chaos. Right? So, you know, that that clearly is unethical and you’ve gotta put things in place to monitor people who are leaving. Not everybody who’s leaving is happy. Right? Happy with you. So be careful of bad leavers. There’s some really good examples and some good ideas. Thank you. Stan, I got a question for you here. What strategies can can employ? West strategies can employers use, I think, to upskill current employees to work with AI. There’s a lot of free resources like on EDX Coursera on the topic quite frankly, like literally super simple answer. Google it and look for introductory courses. To this just to understand how it’s be utilized to understand like where it came from. Like this stuff is not all groundbreaking deal in the regard of like you’re in an entirely isolated place. Just understanding the history and will help you get grounding. And then beginning to understand the context that all this technology works in will be key. Like, this isn’t one of those, like, immediately go set up for Python courses, they’re coding tomorrow. Like this is not what you should be rushing to. Like take it carefully, but there’s a lot of free resources just understanding how all this comes to fruition, how it’s architected, the risks associated with the activity Right? And I think it’s really important to, like, don’t have your head in the sand on this. Like you essentially have constructive notice right now that a lot of firms are experimenting with this. And as I said before, pretty sure everyone and their family members have been talking to this just as a chatbot itself. So the more that you essentially come to grips with evolving tech, emerging tech, and interactive it yourself, you will start to see faults, weaknesses, advantages of it, which will help give you really good context when you speak to business on it. But I really do strongly encourage taking some introductory classes, which again are free. And then kind of using that to reform your firm training. Because people find that to be a more interesting approach to training as well rather than having off the shelf material, especially developed by some other vendor, like I give this analogy of like, if you’re in the financial industry and your training presentation has a character with a construction add on, that training is probably not well suited for the activity that you’re engaged in. Thank you, Sam, yeah, good to make sure that whatever we are learning actually seems like it’s directed at us and makes sense for what we’re trying to do. Last question just for getting towards time here, Shlomit, what advice do you have for firms that are skeptical of integrating AI? This is a good question, and there are a lot of the firms that are skeptical. I tend to say from an AIP or perspective, eventually the proof is in the pudding. Results are the measurements that you need to make sure that are good for you. And a second thing and a second suggestion is to request and ask for all formats of explanatory models and model governance and the bias evaluation of the models to make you more confident in your ability to use it. But again, I think that it is already common knowledge that AI can do better for those companies who are hesitant, just ask to test it and see if the results are really better for you. And I’m sure that if the solution is implemented correctly, you will find out that this is the case. That’s a good answer. I mean, certainly being used in so many use cases. I think it’s something you’ve actually mentioned to me a number of times, LA, every time if you go on like a country with passport control, there’s an AI that’s scanning your face and they’re actually real, depending on the country and set up, but like actually relying on this fully almost. So yeah, there’s definitely what to be skeptical about, but also let’s see these cases. So I think that’s I mean Every skeptical person is looking at their phone and expecting it to happen just from recognizing himself. So Think about that the next time that you say no to AI. That’s a good point. Thank you. David, if I could also add to Shwomit’s answer there, going back to the previous question just asked the mention of the word framework is really important because like the Federal Reserve just published like updated commentary on like Model Risk Management, right? And there’s a lot of frameworks that are available that look at AI and how to think about AI and change management technology governance in general that are really, really effective training material for any firm. And again, they’re written in a way that anyone should be able to pick this up and understand and go along. So I highly encourage looking at it from that perspective. I’m pretty sure Finra is also published on this too. Yeah, thank you Stan. I just wanna wrap up with the short anecdote, so I was thinking about this earlier in the week. So imagine you’re running to get to train. And just barely missed it, so you’re like, you know, frustrated or sweating, losing your breath, you know, gosh, missed my train in my wrists hallway just to try to get it. You wait for the next one and finally it arrives, except it’s not a train. It’s a hypersonic vehicle that takes you home in half the time. Now that sounds nuts, but that’s basically what’s happening right now with AI and compliance and actually really any industry. Just what you can do changes every time you look at the news again. And I mean, if you’re really paying close attention, there’s every day, there’s just new companies popping up with the whole new use cases and and all of them are really good. So, the question isn’t just about what’s the best model or tool to use now, but more like how can we position ourselves to adapt and leverage the next wave of technological advances. This level of opportunities is a fundamental part of why our guest today and many others are focusing on AI and particularly Chad GPT and LMs. Thank you to our panelists, Stan, Cameron, Shlomit, your insights, provide us with a real, I think, just nuanced understanding of how AI and especially GPT and arch language models are reshaping the world of financial e commerce compliance and how you can effectively leverage them to control risk. So, thank you to everyone. Who joined us today, we hope that this discussion has challenged your thinking and sparked curiosity. As the conversations around AI and compliance continues to evolve, we invite you to stay engaged ask questions and continue learning. Now just a quick word about shield, the shield platform is built with modern architecture designed to deliver actionable insights about risk stakeholders throughout your organization. It’s engineered to adapt to the fast pace of technology ensuring you’re always ahead of the curve. We invite you to learn more about visiting our website, reaching out to our team of experts. In closing, the rise of AI and financial services, e comms compliance is an exciting journey, and one that we’re pretty thrilled to be a part of with all of you. We hope to see you at our next webinar. Thank you all, and have a great day.