Welcome, everyone. I’m Jess Jones, and I’m delighted to moderate today’s panel on Amplify with the people who built it. In this session, we’ll dive into how Shield is purpose building powerful AI capabilities, AI assistance, summarization, and improving detection and risk coverage. These innovations are transforming how compliance professionals tackle challenges in surveillance, risk management, and discovery. Before we start, let me introduce our expert team. Shlomit Laban, VP of data science at Shield, Alex Tulusina, director of product strategy at Shield, and Yair Fogel Dror, director of data science at Shield. We’ll explore how these capabilities are shaping the industry and how Shield’s new toolkit, Amplify, helps streamline operations and compliance processes. Shlomit, if I can ask you first of all, Shield’s AI powered assistant, Sheila, represents a significant leap in using AI for compliance as the drive as the driving force, but, behind AI at Shield. Could you start by explaining how AI assistance, specifically through Sheila, is enhancing productivity and accuracy for compliance professionals. Sure, Jess. So first, delighted to be here. And, maybe I’ll start by, looking from a broader perspective, but on the leap that occurred in the AI world in the past two years. So we for several years now, we have been training models, machine learning based models for specific tasks, trying to do detection, analysis, classification, and all kinds of, unique tasks, collecting data, training on data, etcetera. And since the emergence of the new GenAI technology that started with ChargePT and all of the competitors that raised since, and developed since, we are in a new world where suddenly human level reasoning is really possible. Suddenly, there are models that actually understand, what, we want them to what we’re asking asking them or can communicate with people. I think that most people in the modern world has experience with that and know how that feels. And we’re chilled want to utilize it, for the purpose of, the compliance team as well. And I think that the major change in terms of what AI can enable now is, the ability to understand direct instructions or requests from, the users. So if, we use to need to draft request or, for example, in search mode, curate queries or older fashion world, build lexicons, and try to figure out exactly how to present, or ask information to our systems. Today, we can do it in really natural human language, presenting our needs without necessarily needing to, pre train the system in advance to get what we need just by the ability to use very powerful models on the back. And Sheila is a great example for that. Sheila is addressing, the analysis task of surveillance teams. So, they have access to a lot of data. Data is stored, is managed, is accessed. Queries can be written to, fetch data and to curate it. But, usually, search tasks have been, I would say, expertise, required some level of expertise required tasks because you needed to know how to curate a question, what exactly to ask. And in the new world today, this is no longer, needed. Introducing Sheila, we are producing a platform that really allows you to ask whatever you want on your own data and get not only a direct answer, but all relevant, communications that are associated to it. This is yeah. Alex, go ahead. I I was just gonna take it. As one of those people, it’s you know, I was once one of those experts. Right? And and, who and it’s different. Those. Yeah. Right? So you had to be two people. Right? It’s one, you’re an expert who understands a platform really well. And a lot of times, that means the mechanic who knows how the old car works really well. Right? It’s just an interface. Right? And you know how to find things with that interface. It’s not the best way to search or find information. And then the other kind of expert you have to be is someone who understands those communications from a market perspective, but also from a linguistic perspective, how people speak. Right? And, you know, sitting where I used to sit for a long time, I got pretty good at knowing how people would communicate and that kind of thing. But it meant that I had to do a lot of these investigations, right, because I was the one person who could kinda squeeze the results out of a dataset and kind of interrogate it in all these different ways. The great thing about something like Sheila is is that you can get to the heart of the matter much faster, and you don’t have to be an expert. You could key in questions for the very things you’re looking for, and then you could get those messages back ordered in a relevant way. Right? And it’s much better. The example I think of is, we had this big thing a few years back where there was this ice storm. Right? And we had to investigate if a lot of money had been made around this ice storm. And so we know the trader traders associated. We pulled all of the comms, but there was no easy way for us to figure out what exactly they were talking about. We could type in ice storm, and that would bring up some messages, but not necessarily everything they were able you know, talking about. Ultimately, we had to look through everything. And this was a high pressure situation where we had to find those messages as quickly as possible. With something like Sheila, you can go in and go straight to those, you know, show me conversations about a nice storm. Show me conversations about project x. Right? And you can get to the heart of the matter much quicker and then start to interrogate that information on top of it even more. Are there signs of frustration here? Is someone, behaving in an inappropriate way with regards to x and that kind of thing and really start to pull together a vision for what your dataset is saying that’s, much more detailed and that you can get to much more quickly. Yeah. I think that, from a technique, technological perspective, the magic here is the fact that, actually, the system all of a sudden understands you. You do not need to, learn the system and its language. It understands human language, and it’s a huge difference. And it will create a huge effect on the way that people work and what can be achieved as I mentioned. Okay. Those are those are some really fascinating insights already. So can you just expand a little bit more in in terms of how that’s going to streamline processes? I mean, I think it’s great the way you explaining how, Sheila allows you allows, compliance professionals to as if I have heard you say describe it before, Shoma, and have a conversation directly with the with the data. But in terms of streamlining, workflows, how how does that fit into, you know, Amplify the new product offerings? Alex, maybe you have, some thoughts on the the process. Well, you know, we have a few tools that help with that. Right? Not just Sheila. But so with Sheila, I think we we you know, the the the idea is really kinda what I just outlined. Right? So, basically, you can you can, you know, another way to think of it actually is instead of relying on a specialist for the system, you can ask the specialist for that information to just ask the questions. Right? You can kind of bypass that. I think that’s one thing because with a lot of investigations, which is Sheila’s pretty purpose built for investigations at this point. Right? It sits over cases. So we you pull together a set of communications, and then it sits on top of it, and it can kinda tell you what’s in there. Now you could bring someone directly from legal or HR or fraud or InfoSlack or wherever it is we’re worried about, and they can kinda have at it with that information. And I think that’s really rich because often those people sit at a distance from the communications. They don’t get them until they’ve been kind of boiled down in their alley tube. So they don’t really have much sense of what’s going in there. From an efficiency perspective and a coverage perspective, not just regulatory risk, but any kind of risk that might be in a dataset, I think that adds a lot. Right? So that’s the other one is with regards to risk reasoning. Right? So risk reasoning, we could talk about a little bit more too. But, you know, one of the things we’ve seen Gen AI is really good at is summarize. And that’s all fine and good. We’ve all had this experience of getting Gen AI to summarize it for us already. Right? But what Shalit and her team, all credit go to them, right, have developed as a way to specifically, tease out what the risk is specific in an email. Right? And then summarize that for reviewers or anyone accessing those alerts. That’s another huge efficiency game. A lot of times we don’t know why we’re looking at something, especially for not necessarily more junior staff, but a lot of these issues are really complex. Right? So we’re not always gonna understand what it is, but we have a a risk reason in it as well, which is another, a way that we see ourselves add an efficiency to Teams because it tells, reviewers and anyone access alerts why they should look, you know, or this, I should say. So, Edwin, did you have Yeah. Go ahead. Yes. Eventually, the again, the system understands you, but what to continue what Alex is mentioning, it is also tailored, to look at everything from a perspective of finance, compliance, risk, financial risk, and to, give the angle of, this kind of information. Risk reasoning is a great example. Instead of, when, we’re reviewing an alert or even, when, asking a question in Sheila, instead of just, receiving an answer or getting a summarization, everything is focused from, the information generated by the system or by the model. Everything is focused to look at it from the eye of a compliance officer. And it creates a huge difference in terms of the efficiency of the results and the information that is given, to the user himself, or herself, of course. And, I think that this is the uniqueness of the entire Amplify offering. Again, it’s a leap to the really, breathtaking new technology that, AI technology that was released, but also tailoring it exactly to the use case and focusing it to look at the world from, the, perspective of a compliance officer. And, this, is something that, really can be achieved right now, and we’re very excited about it. Great. Thanks for that. And, Yaeje, can I bring you into the conversation as well? Could you give us your thoughts specifically around accuracy and scalability of the new toolkit, particularly over sub substantial datasets? And how do the technologies maintain responsiveness while analyzing these large datasets? Sure. So, I think, actually, the the main focus is not necessarily accuracy, but how to build trust in this, solution to be accurate at scale. Because we, the customers, I mean, the regulator, everyone need to trust the system Yeah. To be to be accurate enough. And this has been our, goal even with the previous technologies. But, with Gen AI, things look a bit different. When I think about trust and accuracy, I think, you can drive trust by, one of two things. You you can either, understand exactly how the system operates. And with previous technologies, this has been easier. I mean, rule based solutions, even classic machine learning, everything seemed a bit more transparent, explainable in a way. And, with Gen AI, especially with the most advanced, and most accurate solutions, we usually, see huge models. We don’t have complete information about how how they exactly they would train, which data used to train them, and so on. So this seems a bit more challenging. But on the other hand, as, Shlomit and Alex already mentioned, you can leverage the ability of the model to to communicate with you in plain language to not only, direct it and instruct it, into what exactly you want the model to do, but also to share with you its own, set its own thinking process. So the model can actually explain to you what drive, the specific prediction or specific outcome outcome of, of an analysis. So, for example, even if I don’t entirely understand as a user what exactly, how exactly the model operates, I can still, read a very simple explanation of what the model thought I asked it to do and, why given a specific input, for example, it raised the an alert or or not. So this brings some explainability and some transparency into the thinking process of the the model itself. And the other way to gain trust is, basically by experience. I mean, you need to to engage with that technology enough times to see it, works, with large enough numbers. So even though this, Gen AI technology is not that data driven, from our perspective, I mean, you you don’t need large labeled datasets to train it to your needs. You you can basically instruct it and tell it what you want it to do. We still, we actually focus even more than previously on getting a very large benchmark dataset of, labeled, dataset that we can experiment with and evaluate our solutions against. And this benchmark, is now growing even faster to, to contain data from different customers, different domains, fabricated data, real data, all these hard edge cases of hard negatives, hard positives, very subtle cases, that we find previous technologies to be struggled with. And now we can offer an accurate solution. So we put a lot of focus on, benchmarking every, version, every solution we, we deliver, and and get numbers to to support, the this accuracy. So this is about the accuracy and trust. The the other aspect other challenge is how to operate these technologies, on at scale because these technologies the and the more advanced that technology is, usually, it is slower. It costs more money. You cannot use Gen AI just simply as, as it is. It is on all the data you have. It will take too much time. So one of the challenges we are facing now when we basically shifted the way our system is built in a way, is, building not a specific model with a specific predefined task that we, want to make sure it’s accurate enough, but we build an entire pipeline or workflow of analysis, let’s say. Mhmm. So it can help us decide very wisely, where and when we want to apply which solution. So we can benefit from the smartest, models even though they are slower and cost more money. So we cannot use them on all the data, but if we apply them smartly enough, then we can benefit from them on the all the relevant data, we want to. So we can benefit from them and run them at scale. K. I mean, some fascinating points there, Yaya. I mean, the just going back to your earlier point about how the the model can explain its own thinking process, that’s a game changer, isn’t it, for anyone, engaging or asking the questions and then understanding the thinking behind it. That that is transformational. I agree. I think it actually it was found out, out by, by almost, not mistake, but, originally, they were not designed to do that, but eventually, people find out research found that, if you ask the model to explain itself and think out loud, you benefit two things. First, you can basically understand what the model have done, has done it, but, also, you get even better accuracy. So the more inter interactive, and, in flow of information we have between the user and the model, we get better results and better explainability. So we benefit from two aspects. So that that engagement is driving it. It’s I mean, really, really interesting stuff. And then secondly, can I just pick you up on the other point of benchmarking? Because, obviously, that is so pivotal for people working in compliance, isn’t it, to make sure they really have that that trust in the in the benchmarking. Are there any kind of other insights you can give us around that specifically? Yeah. So benchmarking in in our, domain is, actually very challenging because what we want is to, have a big enough benchmark so we can trust the model to be accurate on different circumstances, use cases, different data sources, different types of customers, and so on on one hand, one end. But, we’re still looking for, in many cases, for a needle in the haystack. So it’s kind of hard to to collect a lot of data and also label it just to find enough, for example, enough suspicious content, that you can then test your model against. Mhmm. Because we want the coverage to be, as high as possible with, as few noise, as possible. So, we want this huge dataset labeled. We want to find these, positive alerts, and this is very challenging. So, I mean, fortunately, we used to work on data driven solution before. So we, created all kinds of advanced, techniques and methods to to collect data, to label it fast, to to create more and more vary, various data, and add it to our benchmark. And now instead of use this data to just train solutions, we can use a lot more data, much more data to to evaluate our solutions. So, basically, we have, we the fact that we don’t need to train the model Yeah. Gives us the opportunity to test it even more, if that makes sense. Yeah. Perfect. Alright, Shlomit. Can I move on to something slightly different and come back to you? It’d be great if you could give us some thoughts around how the Amplify toolkit helps surface key insights from con from communications and also aids in identifying potential risks. Are there further thoughts you can give us around those areas specifically? Yes. So, first, we also embedded the Gen AI in in the Amplify solution in our surveillance stack. I think that, Yair also mentioned the fact that, this is a, on on one hand, it is a with a high cost and, with a, maybe you can say limited, mainly high cost and performance issues. So you need to, choose wisely when to ask, the questions. So our system is, as mentioned, built in a multilayered way when a lot of our investment is, being done in how to widen the net that we’re spreading and find the things that we actually think may be relevant to, catch. And now we have the, GenAI to actually inspect it and do a second iteration of thought human level, thinking and, map all of these that are not really important and really pinpoint, the real risks. Now we can, look at risks from a much wider perspective reading the entire conversation. Maybe next phase looking at multiple conversations, in order to understand the, growing risk. No need to pinpoint, only sentences as a classical machine learning models used to do. Additional, is the and, again, you have explained it very well. The, chain of thoughts and the risk reasoning that we are performing, serves the same, three times. It improves the model because, like, performing a conversation again, you get better results. It proves, efficiency because now the user and risk better. He understand what was going on, and, it in improves trust in the model because I can now, trust it to be accurate. So, all of this is now embedded in our new solution and makes it very robust. I want to add even yet another perspective of, in general, surfacing risks. And the, the way that we have to actually build or expand the model, again, developed first for, the shield on Amplify solution, but eventually, it is also released or released to the customers, for their own usage of, building or expanding the models or covering more specific risks or customizing to their own specific policies, on their own. So customers used to have the ability to edit lexicons or to expand lexicons in a more traditional manner. And then, some of them or a significant amount of them also shifted to some data science efforts in training the classifiers and the machine learning based models. All of it, again, same as a a performing investigation. All of it requires a significant and heavy lifting, a a significant experience and the expertise. And the same and Alex maybe can talk to that as well as he was one of the model builders from, previously. And now today, we we are also enabling again the capabilities of these models and the way that we have built our solution in a multilayered fashion, just learning from specific examples and guided by additional explanations of what needs to be found, we can actually expand the ability of the model to detect further risks that may have not been considered in the out of the box model by Shield. This is the way that we do it internally, but we also releasing it to customers. So everything today is very interactive and human level, built. It is a real, new way of thinking of, things And even, risk identification, is not all only more thorough and compound. It’s also being, the ability to, modify it and expand it is being axed, given to the user in a much more, user friendly way. Alex, maybe you wanna You’re up to you, Alex. Can I just comment on Yeah? Yeah. I just think, Shlomo, what you’re describing there just sounds like the the sort of flexibility, the customization, and so on. I think that’s where compliance professionals will be, you know, that’s exactly what they need, isn’t it? And broadening the datasets to to better to better pinpoint these. Sorry, Alex. I interrupted you too. No. Like, I I I find it really interesting and and exciting too. I I think, you know, I think the the one of the top of my concerns for, risk functions or surveillance and compliance functions is risk coverage. Right? It’s just this view that we’re not quite capturing everything. Right? We don’t quite have tools suited to how we speak or our markets or how they change and that kind of thing. Yeah. And what is describing is a capability we’ve built that allows users to take elements of language they’ve identified as potential risks, right, and create essentially mini models out of them, right, that they control. They’re theirs. Right? They sit alongside other detections we’ve developed. Right? But everything we’ve seen from the tools we’ve been using is the results are really strong. By empowering our users to sort of make some of these decisions, because we have these other tools we’ve described too. Right? But but to but to actually sort of create kind of mini models around risks they found or language that they may be relevant. They get to take ownership of those risks, but also have something tailor made to what it is, to what to their risk appetites, right, and to what their purview is across a particular region or a particular market or that kind of thing. We think that’s really, really powerful. I wanted to add something to what Yair was saying on the summaries earlier, which is just how strong these summaries are. Right? So a lot of times when you’re looking through the reasoning, it’ll tell the user. Right? This is, you know, essentially, something along the lines of, this is what I was able to find based on your question about blank. It doesn’t quite contain messages about that, but here’s some like messages that might be relevant. Right? And you have a lot of this sort of language to really situate yourself within a set of results, or an alert to really know why you’re looking at it, but also to know how it was found exactly and sort of what logic went into to to finding it. Right? So there isn’t sort of a a a a blank space between the prompt and the results. There’s this sort of reasoning that sits in between whether it’s in Sheila or risk reasoning, or some of the other, capabilities we’ve developed. Yeah. One other quick point, and then is just around just the targeted nature of everything we’re building. Right? Is that, you know, Yair talked about cost. Right? We’ve talked about process, and and all of that is true. Right? You know, we have to we wanna add value, but we don’t want to, break anything. We don’t want it to be so slow that it ceases to add value. But the other thing is is I think with where we are with Jenny and I, we aren’t in a place where people are gonna rip everything out wholesale and put this in. Right? And our homework, the work we have to do is to figure out what targeted ways we can add significant value to our customers, right, in ways that they can see measurable and tangible results. Mhmm. Right? And so all of the capabilities we’ve built involve that, and credit goes to to Shlomid and Yair and their larger team to the thought they put into that because these aren’t just bolted on GPTs. Right? There’s a lot more going on to satisfy these particular use cases we’ve talked through, whether it’s, you know, identifying why something was alerting, creating many models out of language, or, you know, Sheila and so on. So, just to to bring you back to the to the end users as well, in terms of you mentioned taking ownership of the models, which I think is really, really fascinating, and you also mentioned prompts briefly. Is there a growing confidence from clients in in and and users in, you know, understanding how to improve their own prompts, understanding how to better, you know, get get the results that they want from from Gen AI broadly and, obviously, specifically from from your offerings? So my view is that, you know, we we have to respect the journey people are on. Right? And everyone is on some kind of journey. Right? And we see a lot of different, appetites for these capabilities. Right? I think over time, we’ll see most of the market gravitate in this direction. For now, we see people who are super advanced and they’re trialing stuff themselves, or they’re really curious about it, and we see others who are more hesitant. Everyone wants to know about it. Right? And everyone wants to educate themselves and what have you. What we can do and what we’re trying to do and, again, I think it’s why having these targeted thoughtful solutions really helps, is give tangible ways people can start to use these. Right? That revolutionize elements of their program, but aren’t revolutionizing everything yet. Right? They’re still waiting for regulators to give firm guidance one way or another, but there’s enough guidance to say, you can start trying this stuff. And if you feel comfortable with the controls that are there and your own ability to vet them, then you can use these tools. Right? So we’ve made them available, and we’re starting to see people use them, and we’re starting to use them in POCs and in projects. And the results we’re seeing are really good. At some point, there’ll be enough momentum behind all of this, not just from what Shield has built, but I think just as all of us lean more into using GenAI across our lives. For now, we’ve built a path that we think is really valuable, right, and measurable that people can hold and use and kind of really start this journey, but in a meaningful way. Thanks, Alex. I mean, I’d like to link that slightly back to Yair’s point from earlier on talking about trust and, you know, building trust, building confidence. It’s all part of the same journey. Yair, so I don’t know if you’ve got additional thoughts around that. You know, that understanding gen AI and and the potential of the of the products that you’re now offering, is sort of in tandem with with that trust around, you know, the data and and and what it’s what it can deliver. So maybe one thought about it. One of the, I think the methods I like more, more in how we, develop some of the solutions is basically, we let a domain expert to interact with the system, and we track the way they interact with it. So for example, let’s say I’m a domain expert. They want to, make the system, catch a new type of suspicious behavior that we haven’t alerted on, as you know. And so the first thing I will do is try to come up with more stuff. I’ll find some some noise. I’ll try to reduce the noise, mainly pinpoint the system to the specific type of, content I’m interested in. And what we do behind the scenes, we can, collect all this interaction because this interaction is basically in plain, language. And we use this in order to better learn how to instruct the system, from now on, when we try to to cover this behavior. So the fact that the interaction is, basically in plain language help us to bring some transparency into how we learned from the feedback or from the way, the user interacted with the system. When we when we design the instructions, we want to give the system from now on. And one interesting, trick we sometimes use is, as I as as we said, I mean, GenAI can understand things. It can understand this interaction as well. So we can, ask the Gen AI offline, of course, to review all this interaction from the user with the system and come up with instructions that will represent the same idea, the same stuff I’m look I and the user, I mean, is interested in finding. And from now on, we get very tailor made, instructions, that were designed using the feedback or using the experience the user had with the system. And we can show these instructions to the user. So the user actually can see, okay, this is what I was looking for. This is how I looked for it. And this is how from now on, if I choose to, the system can look for it in the future. So everything I mean, the model itself is not transparent, but the instructions, can be at some, in some points, to some extent can be transparent, and the way we, benefit from feedback can be more transparent. So this is a very, I think, cool trick that we use, but it gives some visibility into how the system, operates. And it gives, I think it adds more to the trust. And also, it gives, the user some easier ability to control how the system will operate because if the system have not detected some specific, behavior, the user can, much, much easier than it used to be, add the and and expand the coverage of the system using these these methods. Yeah. I mean, that that is incredibly powerful. Shoma, if I can come back to you, I I what I would really love is your perspective on how Shield is really staying at the forefront of implementing latest technology and driving innovation. Everything that you’ve all described so far is incredibly powerful, and it almost feels like as humans, we’re kind of catching up with with the technology. But be be fascinating to hear your thoughts around, you know, what the what future in innovation is yet to bring. And now just we’ve, been in these sessions for a while, and you know all of my stories, but I I really have to go back to your previous point and say something about really the shift that also is occurring in the industry itself. You asked, are we more are a compliance teams or financial industry more a recipient to a new AI? And I think that in general, since everybody is exposed to these capabilities, the willingness to actually adopt the new technology is is growing in a huge pace because we see it in everyday everyday’s life. So why can’t I do it at work? If I can ask Judge GPT a very simple question on anything in the world and get an answer, how can it be that I, as a compliance officer, need to curate a heavy question, need to send send something to the ediscovery department in order to get answers, need to work, tediously on the models to actually find out information. This does not make any sense. And I think, that, the willingness to receive it on one hand is extremely going. If I look at communications with customers, even a year ago, there was still a lot of suspicious regarding AI. Today, this is the first question that is being asked. Where is it? How can I see it in the product? How will it help me? And, I think that with the amplify offering, we have been, deploying AI for a very long time at Chill, but the amplify offering really gives the the leap to, benefit benefit of the latest latest latest best technology. I and, again, tailored to the use cases and looked at from the, eyes of compliance team. And, I want to, maybe mention two more things. One, risks in general. So it’s not only the compliance that need to trust it, also regulations, and we see that regulations are heavily evolving. We see the UAE act that is, emerging, was it was a few weeks ago that it only came into initial final release. It will take time until it will be enforced, but still and I think that it does represent, the fact that everything is everybody is already acknowledging that this technology is here to stay. It needs to be regulated. Every company that develops or uses this technology must follow those rules, and we achieved, are very strict in complying to all regulations in that manner. This is something that is very important for users, to know and to enforce. But also, again, it means that, this technology is currently here to stay. And the last thing maybe to answer your question, how do we make sure that we are most updated in terms of technology? I think that in the new world, it used to be that, you curated your own algorithms, and then you trained your own models. But this day and age, open source models or, available models out there in the market are, winning over anything. No car company will imagine, today developing all parts of themselves. Everybody is specializing in everything. And with Shield, we are taking, an approach that is called the model agnostic approach, meaning the minute that a new model is being released to the market and is proven to be better, faster, and also applies to all of our strict and put in place testings, in terms of accuracy and performance for our use case. It will, immediately be plugged into the system. What we did is to tailor the solution around it, curate the questions, create the flow, worked on instructions, created the pipeline of spreading a net using additional technology in order to, decide when to use it, but we are ready to be most updated in most recent technology. And this is the situation that we need to be adopted by anyone in the future. That that I mean, so many points you’ve made there are really fascinating. I want to pick up on one, which is, if there’s been a sort of seismic shift that it’s gone from suspicion to now expectation. I mean, that is that is really, really fascinating. And, Alex, I wonder if you’ve got you know, that has been, you know, from the clients that you’re working with or, you know, engaging with on a regular basis, did you find that that is, you know, that sort of is across the board that it’s gone from suspicion to now people wanna see how this is gonna be really implemented in their professional lives and not just, you know, in in sort of on the side personal life, look at ChatGPT. Yeah. Yeah. Yeah. Yeah. Yeah. I think we’re moving on from, like, the wow effect to, like, okay. So how could this benefit me? Right? I think across firms, we’ve seen a lot of education happening internally, development of policies, right, a view on what they can and can’t do, with AI, right, as well as sort of regulators kind of, I think, laying down what are gonna stay the the kind of lines for what’s what’s permissible and what’s not. And and and then and then we’re all kind of figuring out how to vet it in between. Right? So we’re at a place where I think everyone feels more informed as a consumer, right, about how they can kinda now go on to bring on these capabilities. So we’ve gone from, I would say, a year ago where we saw more of the fascinated, you know, what are you guys doing? Can we talk about it? But that’s it. Right? To okay. You know, the RFPs we’re seeing, the customer questions we’re getting, a lot of our demos are spent around really digging into what our capabilities can do, how they can bring them on, how they would work if we could bring them on. And it’s made for really exciting conversations because once your customers are engaged or starting to ask those questions, the whole relationship becomes more symbiotic because it feeds how we continue to develop our product and what have you. So for me, things are just starting to get really exciting as we look to bring these capabilities to market, but also continue to innovate on them in ways that can continue to to help our customers. Great stuff. And, Shlob, just to come back to you, we’re nearly at the end of our time. But when you mentioned model agnostic, that gives you the the capability to stay nimble and to stay agile and to really just see what what’s working the best and and really help pull those things out. Is that right? Is that what it means to you? It allows us it allows us both, the flexibility in terms of deployment, cost, cloud availability, preferences of the users to choose the relevant models, for each. But also in terms of, releases, Chargept released, GPT four point o, we can immediately adapt it. Same for a cloud or any other version of model. This industry is fast growing, and models keep, getting improved really on a monthly basis. And, we want to offer, the best, to use the best solution that we can, already in our technology. We built it in a way that allows us relatively very easily to replace and to upgrade to the next version of the model. No need to upgrade the entire platform for the user. It is like a plug and play. And, really, it has a a lot of importance. And still, the the the trick of everything is how to leverage it to the best use of our customers, how to tailor it to our workflows, to our product, to our use case, to our, layering in a way that will give our customers, the best, assist in their day to day job. And, really, it’s it’s a shift from doing a lot needing a lot of expertise to run these technologies to actually being able to interact with it. Unfortunately, we have to leave the discussion there. Thank you all so much for for such interesting input. Thanks, Shlomit, Alex, and and Yair for everything you had to say today, and I really look forward to seeing how these innovative conversations around AI continue to to evolve. Thank you. Bye. Thank you. Bye bye.