All right, I think we are ready to get started. Thanks for joining us today. Before we get started with the presentation, we do just have a few housekeeping items that we'd like to cover. First, we will be recording the session today. If you do not wish to be present for the recording, this webinar will be made available to watch on demand. A link to access the recording will be sent out after this webinar is completed. We will now start recording and begin our presentation. Recording in progress. Welcome to our discussion on Gen AI and innovating responsibly, your competitive advantage in the age of generative AI and large language models. I'm Michele Caselnova, Dun & Bradstreet's Chief Communications Officer, and I'm really happy to be your host today. I'd like to introduce my colleagues who are leading the approach to Gen AI at Dun & Bradstreet, and who I know are really incredibly excited about the opportunity to show you the power of D&B.AI Labs today. First is Gary Kotovets, our Chief Data and Analytics Officer. Gary is joined by Ilya Meyzin, who's Vice President and Head of Data Science here at Dun & Bradstreet. Gary and Ilya have a great session planned, including three demonstrations that will illustrate how we are leading the way in responsible AI, excuse me, at Dun & Bradstreet. Following that, we're going to leave a lot of time for your questions and comments. We have a lot to cover today, and I know that Gary is ready to kick things off. Over to you, Gary. Thanks, Michele, and thank you to all watching and listening today across the globe. As we all know, AI has been one of the hottest topics today, spanning nearly all industries and markets. We absolutely agree that it is a transformative time for everyone in the industry. Given how prominent these discussions have become, and also on the back of the launch of our D&B.AI Labs, we thought, and we would be incredibly excited to share with you what we are doing within this space and about D&B.AI Labs and its capabilities. As we kick this off, I thought it would be good to help you understand how D&B has framed its strategy. Our first step was to really talk to our customers and really understand what their strategy is going to be, as we see that this is, you know, a fast-moving technology and quickly taking over a lot of news and probably a lot of mind space, starting from the CEOs of their companies, all the way down to engineers and data science experts. Based on that feedback and the conversations that we've had, what we've heard are really pretty consistent three main points, right? First was the need to apply these capabilities to many use cases. Everyone is scrambling and kind of running to try to figure out where and how this is going to help them. Whether that is improving productivity and efficiency or whether that is going to generate additional incremental revenue and help them save costs. Another very important point is about IP and data rights. Understanding very quickly how impactful this technology could be to the data that you have and to the IP rights that you have of that data is very important. Putting the right regulatory compliance, ethical, as well as security guardrails, is extremely critical before you start doing anything related in this space. Then the third is how to productize and make it available, productize actually and operationalize these capabilities for your within your own organization, as well as really to any of your customers and your customer's customers. What we thought is based on that feedback, we would walk you through, you know, what we're doing in this space and how we plan to help our customers, really, you know, meet some of these concerns and achieve some of the objectives that they're looking for. We'll walk you through kind of separating hype from reality, meaning what is Gen AI and what does it really do, and what doesn't it do? To help kind of filter out where we think you can really use it. How you should be designing responsibly and securely, and what we've done in that space to help you achieve those objectives, and then walk you through some really interesting examples of what we've been able to achieve using our D&B.AI Labs. What we hope, you as our customer, will be able to come in and work with us on, where we can help you, again, realize and optimize these capabilities for your own internal workflows using D&B data and capabilities. As you know, you know, D&B has been a responsible and trusted brand for almost 200 years now. We take the trust and responsibility with our customers very seriously. I'm not sure if anyone, have you seen or not, but we were just named by Newsweek as one of the most trusted brands in U.S. We use these same principles of trust and responsibility to guide us in designing our D&B.AI Labs. Let me show you what I mean by that. If you look at. Starting kind of from the left-hand side in terms of our products and services that underpin our data. We have 520+ million businesses covering around the world, which is basically every single business that is commercially critical for the economy across 180 different markets. We have 4,000+ different data elements that we collect on each of those businesses. We have 400 million contacts that are relevant to those businesses, and we have 40 million+ global family trees with 10 years of historical data. These are very important components that are needed in order to sort of start to fuel the AI engine, so to speak. The other pieces are also important, that are the accuracy, timeliness, and most importantly, validity of the sources of the data that you're taking the information from. Of course, there's the organized, clean, and contextualized information that you need in order to sort of bring it all together in a way that feeds the machine, the AI models, in a more efficient and accurate manner. On top of all that, we have dozens and dozens of D&B insights and analytics that already contextualize and really give our customers a variety of different perspectives, whether it's on risk, supply chain management, or sales and marketing. If you take all of that and you kind of think about what does that do and how do I apply all of that into the Gen AI world? What we did was, in our D&B.AI Lab, we created a full interpretability and data lineage. What does that mean? That means that when I create or use the models to provide me the answer, I can see all the way the connection from the right to the left, so to speak, where that answer came from, that I can easily verify whether the answer that the Gen AI model produced is accurate, and I can trust it, most importantly. The second very important component that we created was a security, legal, compliance, and ethics, what we call guardrails. What that means is, you know, we actually will have a session in the next few weeks that we will separately invite all of you and others too, where we have a really deep dive into these components. At this point, this was one of the most critical steps, as I mentioned, that we needed to establish, to make sure that not only do we trust what we're doing with these Gen AI models, but also, most importantly, that you trust us in terms of what we will do or can do with these, with these capabilities. We also have deep subject matter experts in-house, whether it's our data scientists or business analysts, who are really able to bring a lot of these Gen AI capabilities to life, right? We have prompt engineers who are able to create the prompts that are needed to specifically meet your particular needs. We've been experimenting and using a lot of these obvious experts to also produce outputs and using these capabilities across our own products and services. We'll walk you through some of those examples later on. What does all of that combined really produce, right? What it really does is reduce risk of hallucination and data poisoning. What that means is, it allows you to, you know, trust the output, because as you've seen, and I'm sure have heard, that a lot of these models, when taking data from publicly available sources, have a risk of creating or providing an answer that is completely inaccurate or just blatantly wrong. The second thing it does, and the capability that we created, is because you know, because the data is structured, because you trust the data, and you trust the, the security and the, the compliance, sort of guardrails, we're able to improve significantly the productization and creation of these capabilities, not just across our products, but hopefully into yours as well. With all of that said, I'll pass it now to Ilya Meyzin, who will drill into really the overview of the D&B.AI Labs and show you some really, really cool, really substantial, kind of substantive demos of what we've been able to do with this technology. Thank you. Ilya? Thank you, Gary. I'm extremely excited to be here with you today and have an opportunity to talk about all these cool things that we have been doing with Gen AI and LLMs, and we have been very busy over the last couple of years. First things first, I just wanna go very quickly through our view on what is real in the space and what is not, and how this is driving our Gen AI and LLM strategy, and why, to Gary's point, we're positioned so well to capitalize on the opportunity. Unless you've been off the grid for the last six months, there's been a huge amount of noise and hype around all these capabilities, and some of the hype is actually true. These models are very powerful tools, and they can do amazing things. They approximate human-level understanding of language, and more recently, I would say they're even starting to approximate human reasoning. For instance, some LLMs perform very well on standardized tests. At the same time, these models, as powerful as they are, it's just that, AI models, which comes with some of the old baggage and some of the new one. On the input side, all of us have heard this phrase, garbage in-... still means garbage out, which is why the point that Gary made about our data, namely that it is clean, verified, contextualized, this is why it's so critical. And in our demos, we will actually show you the huge difference between asking a business-related question of a D&B LLM versus asking exactly the same question of a leading LLM online. Spoiler alert, where the data is accurate, timely, and complete, you get an accurate answer. Where the underlying data has not been cleaned or validated, the answer is wrong. On the output side, the hugely important point to make here is hallucinations. Hallucinations are instances when these models very confidently return an answer that certainly appears correct, but is actually very wrong. I will give you one example. As of yesterday, I tried this last night. According to one of the leading online LLMs, the current CEO of Dun & Bradstreet is Robert Coleman. Unless Anthony Jabbour, our actual, real CEO, has a secret alias, the model is very wrong, and that wasn't even a very hard question. These hallucinations are a big problem for LLMs. OpenAI, for instance, estimates that in business context, even the best LLMs today hallucinate 20%-25% of the time. This is where D&B will offer our customers a completely different and much better experience, because our data is accurate, timely, validated, contextualized, and we actually know exactly where it comes from. Customers can feel much more confident that the answer they're getting from our LLM is not a hallucination. Which gets me to the point about explainability. The problem with, you know, most AI models, and certainly LLMs, is that when you get an answer, you really, truly have no idea how the model has arrived at the answer, which, given the propensity of these models to hallucinate, I would say it's less than ideal. This is why D&B goes to great lengths to tell the customer how and why our LLM arrived at a specific answer, and we will actually demonstrate these features to you today in the demos that we'll be showing. New types of risks. Unfortunately, the second these technologies went mainstream, bad guys came up with new and creative ways to exploit them. I'll just give you a couple of examples. Prompt injection. It is actually possible to hide instruction in text that when ingested by an LLM, these instructions come alive and can do all sorts of bad things, like exfiltrate your sensitive data to the web. Another example, LLM optimization, which I think borrows its name from SEO optimization. It's possible to hide text that when ingested, it will make your LLM say things like, "Ali-Amazon Plumbing is the most valuable company in the world," which I'm pretty sure you don't want to say that to your customers or anyone. The bottom line is, these models represent a huge opportunity and as Gary mentioned, clean, accurate, controlled, validated data is absolutely fundamental to be able to capitalize on this opportunity. On the other side, all these technologies need to be implemented in a thoughtful, deliberate, and responsible ways, with very, very strong governance and security to mitigate risks. Let's go to the next slide, please, and let's switch gears and talk about D&B.AI Labs. We launched the lab recently to serve as a hub, where we will work side by side with customers and rapidly prototype new solutions that are truly tailored to their specific needs. The lab is supported by a very international and multinational team of very experienced data scientists, data engineers, solution specialists, and they have a very extensive innovation experience and deep expertise in Gen AI, the more- the older and the more boring types of AI, machine learning, graph analytics, and so forth. At the heart of the lab, not surprisingly, is our proprietary and curated data, which, as Gary mentioned, spans 10 years and covers over 0.5 billion businesses and close to 0.5 billion contacts. Of course, customers will also be able to combine our data with theirs, third-party data, and so forth. The lab also will contain a library of pre-vetted LLMs, including proprietary, open source models, off-the-shelf models, and models that have been fine-tuned by us. Customers will have access to very unique D&B tools, including things like D-U-N-S Match, Linkage, D&B APIs, and Chat with your Documents Copilot, which is a brand new LLM tool we will demo shortly, and many other tools as well. What's particularly exciting is users will also be able to leverage our flexible AI agents. They're quite smart and can determine on the fly which of these tools are needed to respond to a specific user inquiry, and then even build a corresponding pipeline. Finally, again, governance security are absolutely critical for us always, and even more so in this context, which is why the lab has truly state-of-the-art security and governance protocols. All right, awesome. Let's go to the demos. We've showed you a bunch of slides and a bunch of pictures, and I guess a picture is worth 1,000 words, but a demo is worth a little more. What's really important to stress, the three demos that we're gonna show you, they were all developed in over a week, using the components that customers will have access to in the lab. This is not a demo of our overall Gen AI and LLM capabilities, where we already have many things in production that customers are already benefiting from. This really is a demo of our capability to engage with customers in the lab and prototype things very quickly. The first demo, which we call Chat with Your Documents Copilot, it creates the same experience that you're probably already used to from online LLMs. The only difference, and it's a huge one, the underlying data is D&B data, and in the demo, we will show you exactly why we keep stressing this point. Without further ado, let's please go to the demo. Hello, everyone, and thank you for joining us today. In this first demo, I'll be highlighting one of the key features of our lab, which revolves around the incredible ability for users to extract data elements through interactive Q&A on a vast collection of documents. This demo will also emphasize the importance of using trusted and verified documents together with data lineage, to ensure reliability and interpretability of the results. Imagine a scenario where you have an extensive set of documents, be it financial documents, legal contracts, or CSR reports. Traditionally, extracting specific information from such a large corpus would require an enormous amount of time and effort. But with our lab's cutting-edge technology, this task becomes effortless, accurate, and efficient, as you will see. For example, let us interact and extract insights based on the extensive set of ESG-related documents from D&B. For this first question, I would like to know what are the Scope 1 emissions values for the company Boralex, so a random company that I selected. Note that you have the possibility in the platform to add a D-U-N-S Number in the question to narrow down the search to a specific entity, if you wish. According to our platform, this is the actual value to our question, which is, as you will see, the correct answer. Our platform indeed not only provides the answer to the question, but it also allows users to verify the exact documents used to generate this information, so through this link, as well as the exact location in that document where it was extracted from, which is a crucial ability for users to make decisions confidently. For example, if we click on this link, this will bring us to the CSR reports of that specific company, and it will actually bring us to that specific location within document, where we can see that the answer that has been returned by the platform is actually the correct one. For comparison, on the following screen is the results obtained from another state-of-the-art platform for that same question. Note that I have removed the D-U-N-S Number here, as this is a non-public proprietary D&B data. Not only data sources are often unavailable on these types of other platforms, but it can also be inaccurate, as you can see. We indeed obtain a very different value than the one we obtain in our platform, which we know is inaccurate, as we have seen in the CSR report of the other, of this company. These inaccuracies and lack of data sourcing from other platforms further reinforces the importance of using reliable and trusted datasets, such as D&B data, together with data lineage integrated on the platform. Let us try now other questions, highlighting the efficiency and accuracy of this lab feature. In this case, I'm interested to know what are the Scope 1 emissions for the year 2021 and the 2022. Again, this is the answer that we obtain from the platform and which we can verify through the data lineage. Again, if you click on the link, this will bring us to the actual, the same location, where we can see that the answer is again correct on our platform. As you can see, our platform is indicating a 95% decrease of Scope 1 emissions here, which is impressive. One might be curious to know how they achieved that. Let's just ask this on the platform. What is the reason for this 95% decrease on your Scope 1 emissions between the year 2021 and 2022? As you will see, the answer provided by the system is again, the correct one. We can again check this on the document. If you click on the link, this will bring you to this specific page. Again, here we can see that this is the correct answer from our system. Now let us reset the prompt and try another question on another very different dimension for that same company. For example, what is the current percentage of women in the workforce of the company, Boralex? Let us compare the answer from our platform with the answer from another state-of-the-art platform. Here our platform found that the answer is 33%. 33% of women in the workforce of that specific company. This is the answer obtained from another platform. Once again, the information is unfortunately inaccurate. Not only the answer is inaccurate, but the additional details provided below by this platform have either been taken from inaccurate sources or invented. We often call these hallucinations. In our case, the data lineage capability of our platform allows us to ensure trustworthiness of the collected answer. In this case, we need to see that the right answer was 33%, and this can also be found on the link through the link that we provided. Hallucinations often appear when information is missing and is just being invented or created by the system. In our platform, we ensured that we only return this information when it actually exists, when the system actually found it in one of the documents, and correctly notify users when it does not exist, such as in this last example. Has this company had any privacy breaches in 2022? As you can see, it cannot find the answer. If you check again the data lineage, you will see that he was not able to identify this information in the specific documents that were provided to him, which is the exact behavior that we would like the platform to have to avoid hallucinations. In conclusion, this lab feature truly empowers users to explore large collections of documents swiftly and efficiently, saving valuable time and resources, but also in a reliable and transparent way, thanks to our data lineage integration and our trusted and verified data. Thank you. Awesome. Fantastic. Let's, let's go to demo number two. This is a very different idea that can help our customers in the credit space. A very, very, a very important point to make here. Currently, we do not intend to use LLMs to generate credit scores. However, we absolutely can use these models to make our customer's credit decision process even more efficient by contextualizing more of our data into what we call a human-readable form. Which then demystifies these very valuable, but quite numeric insights and makes them very accessible and actionable for non-technical folks. Let's please go to the demo. Hello, everyone. As Gary and Ilya have mentioned, D&B has a vast amount of proprietary data on 5 million+ businesses globally. Among the many different things we can do with this incredibly valuable and unique data set is creating proprietary scores and reports designed to help you assess not only a business's current state, but also its future outlook. Most of these scores have numeric format, which means that as useful and insightful as they are, the user typically needs to translate from a set of numbers to their meaning in the physical world. In this demo, I'm going to showcase the capability of generating plain English, human readable summaries for two of our financial business scores, Financial Stress Score, and Delinquency Predictor Score. Of course, the same concept can be applied to all our scores or products that contain reason codes. I'm going to start the demo by asking about general information on a company that I only know its name and its country location. You can see that the name and the country of location is enough to find the company. I might be curious as if the company is public or private. Let's ask the question. It's a private company. Now, as I alluded to earlier, let's inquire about the Financial Stress Score of the company. The Financial Stress Score has multiple parts, three numeric parts, and some commentary on the scores. The numeric parts are score itself, its percentile, and simple classification. As you can see, all these are included in the generated report with an explanation of the ranges and meaning. The report also includes the reasons why this specific score was assigned to the company. As a simple example of interacting with the report, let's ask a simple question on the UCC filing, which is mentioned in the report. As you can see, it answers with the definition of UCC and UCC filing. Delinquency Predictor Score has a similar structure. We can generate a similar report here. A very straightforward experience overall that makes the score quite intuitive and easy to contextualize. This capability can work both as a chat, but also could potentially be delivered via an API. Hope this makes sense. Thank you. Back to Ilya. Thank you very much, Mohammed. This is great. Finally, demo number three showcases our approach that leverages flexible AI agents that I alluded to earlier. This is a particularly powerful idea because it combines the intelligence of these LLMs and the true power of D&B capabilities, such as D-U-N-S Match, Linkage, and the new D&B LLM tools. Let's go to the demo, please. Hello, everyone, thank you for joining us today. In the two previous demos, we've seen examples of how we can leverage LLMs to find specific information and explain it in an interactive way. This demo will focus on how we can use AI agents to solve more complex and diverse set of use cases by leveraging multiple capabilities exposed as tools. In this context, tools are existing D&B capabilities, presented to the LLM in a way they can understand. Once a tool is defined and can be used as a building block in a lot of different contexts, the magic happens when the LLM understand which tool to use and how to use them. The set of proprietary D&B tool inside the lab will help you drive efficiency and accuracy of your AI agent behavior. Like the previous one, this demo will emphasize the importance of using trusted and verified data. Even if the data is accessed autonomously by an AI agent through the usage of tools, it's crucial to be able to verify the accuracy of the information provided. Let's start with a simple example that will highlight the broadening shift that AI agents bring to the table regarding user experience. Let's say, for example, that I'm interested in this company located in Belgium. Specifically, I would like to know how many employees they have and how much they spend on salaries. As a cherry on top, I would like to know what is the average yearly salary based on the data. As you can see, the agent happily answers that this company has 538 employees and spends around EUR 48 million on salaries. It also provide the right number for the average yearly salary, a bit below 90K. How did it reach that conclusion? To understand, let's take a look under the hood. We can see that the AI agent took five steps. The first step was actually an action, which is using a tool. In this case, it called the Match tool, because it understood that for everything else, it will need a D‑U‑N‑S Number. D-U-N-S are the unique ID for each company in the world, that identify precisely companies in for D&B data. The AI agent understood my sentence and matched the content of my sentence towards the input of the tool correctly. It got back the D-U-N-S and correctly understood that the next step was to fetch data from this D-U-N-S, which it did, and it got all kinds of firmographics. If we look at the bottom of it, we find out that indeed, this company has 538 employees, and it also has a total payroll cost of around EUR 48 million. All in all, the data was correct. We can verify this, and the AI agent computed the right yearly salary, all of it in a single sentence. Now we can look at another type of use case, and this time, let's look at a list of companies instead. For example, let's say I'm interested in companies in Germany, which are active in the motor vehicle part industry and which are making more than $100 million in revenue. Once again, the AI agent happily answers with a list of companies for which it provide D-U-N-S, names, and activities. At a first glance, we can see that the list is matching the right industry, and by taking a look under the hood, we can see that it correctly decided to use another tool. This time it decided to use the company list tool. Again, it did map correctly my sentence. For example, I said Germany, and it says DE, which is the ISO alpha-2 code. It's also mapping correctly the numbers and the keywords. Now I have my company list, and one of the most frequent use case is to actually deep dive into some of those. There, I can actually leverage the fact that we're in a conversation in the chat to ask something like this: "Can you tell me if the second company is profitable?" Because it knows what we talked about earlier, it can actually directly call the right data tool using the D-U-N-S from the list before. Then it looks at all of the firmographics to answer my question. In this case, it's looking at the net profit margin to tell me that, yes, this company is profitable because the net profit margin is 9.7%. Finally, we've seen in this demo that we can leverage AI agents to ingest and understand complex data, and they can find precise answers in them and explain them, but they also have their own generic knowledge, like, that you can leverage. For example, in this case, I'm not a finance expert, and I'm not sure what this net profit margin is. I can simply ask, for example: "What does this mean?" This time, you can see that it took only one step because the agent is using its own knowledge to explain to me in a simple way how this net profit margin is computed. That's it. This ends this demo for me. Thank you for listening. Thank you, Thibault. Just to wrap up on this is a very, very powerful and a very flexible capability, and the range of use cases that something like this can enable is truly, you know, everything from driving sales and marketing campaigns and lead generation to serving as a copilot to risk-related use cases, due diligence, supplier valuation, assistance with compliance and regulatory requirements, company analysis. It can be used for sort of, you know, internal information management systems. It can be used to personalize the experience that your customers encounter. Absolutely, this is a wonderful capability that we're very excited about. This wraps up the demos, and, Gary, I will hand it over to you. Thanks, Ilya. really amazing stuff. I did wanna kinda just end it with a couple of things, we emphasized one important point. you know, we talked a lot about productization, and operationalizing these capabilities. We talked a lot about a lot of the different use cases, where this capability can apply, and really, the reality is, the use cases are kind of unlimited. right? You really can, you know, redeploy this across multiple different products and services. What we've been doing at Dun & Bradstreet is really taking the lab and the capabilities that are coming out of the lab as a result of these models and are planning to be deployed and planning to deploy them across the different products and services that a lot of our customers are already used to seeing and interacting with. You know, one of the key things that everything kind of starts with is the D-U-N-S, right? The D-U-N-S is really the entity or the number that's really embedded into everyone's workflows, very deeply into every one of our customers, whether it's from a sales and marketing point of view, or from a third-party risk and compliance or from finance. Starting from the D-U-N-S and building on top of that, all of the different products and services that we build into our customers' workflows, just overlaying the large language models and the Gen AI capabilities, will make all of them that much more effective, and most importantly, make our customers that much more efficient and productive. And really improve, you know, all of their capabilities across these products, and improve all of their workflows across the board. That's really sort of what our perspective is, and we're seeing, as you can see in some of these examples that we've shown, you know, how we can make this work. Really, the last point, if you go to the next slide, is our kind of three principles, right, that we continue to guide ourselves and hopefully will continue to guide you as well. Security and compliance, as I mentioned before, we will have a much deeper dive into that, into those sort of practices, and hopefully keep you guys informed about the different evolving regulations and laws and rules, et cetera, that relate to this and how we tend to conform to those. Trust and transparent source of truth, again, an extremely important component in any Gen AI capability. This is what we hopefully were able to demonstrate to you as we were able to achieve. The third is embedded, structured, and organized. As I mentioned, the D-U-N-S Number, which is embedded across all of our customers', infrastructures and workflows, the structured and organized data on top of that allows you to really, very effectively achieve, at a much faster rate, and productionize the LLM and Gen AI capabilities. I'll leave it at that, and thank you again, and I'll open up for, Q&A. All right. Thanks, Gary. Thanks, Ilya, and to our gentlemen who helped prepare and record those demos, thanks so much. At this time, we're going to end our recording and begin the Q&A portion. Recording stopped. I'm gonna stop our sharing. We'd love to hear from everybody that's joined us, so we invite you to ask some questions and share a comment. I will tell you that we have had a couple of questions come in, so I'm just gonna switch over to my iPad here for just a moment. Gary and Ilya, I'm gonna just start with a couple of questions that came in. The first one came in when we were back on the D&B.AI Labs slide, and the question that came in is: What makes up the robust compute component of the labs? Okay, I can jump on this one. It's a great question. The name of the game in this space is GPUs. You can have brilliant people and amazing assets. If you can't run these things on GPUs, they're not going to run very well. As I'm sure everybody knows, there's been a pretty serious shortage of GPUs in the industry. Now, thankfully, we have amazing partners, and we have an absolutely wonderful technology team. They saw these things coming even before it really became a big problem end of 2022. We have what we need to do these things. Okay, thanks, Ilya. Next question. Do we, Dun & Bradstreet, also train on internet data or just datasets that are built by Dun & Bradstreet? Great question, Pradeep. Thank you very much. An incredibly important point. We talked about all the issues with just grabbing internet data. We absolutely will use data from the internet, but we will put so much effort in validating it, standardizing it, contextualizing it, and doing sanity checks against other data points that relate to the data that we bring in. At that point, it is no longer internet data. At that point, it becomes another D&B full-fledged dataset, and then, of course, we will use it in our models. Next question, and I'm going to keep rolling through these. On average, how long would it take to produce a beta version of one of the models that we saw today? My perspective is, these models, these demos that we showed, it took, I would say, like a week and a half. That's roughly what we should expect. Like, a fully-fledged beta version may take a little bit longer, but we're talking about, you know, weeks, not months, as these things normally take. Okay. Gary, we have a couple of questions in here, just related to, the service being available for customers to use and, a couple of questions about, like, specifically which customers, like, who is the ideal customer to use this. Can you talk about that? Sure. First of all, you know, the lab now is, this is sort of the formal kickoff where we're inviting customers to join us and experiment with us and help them, you know, produce these capabilities. It is available to customers. Please contact your account rep, and or also please leave your information in the Q&A, and we will be reaching out and setting up a, you know, a separate, sort of, session, to walk you through, how to get you all started. A question from Adrian and a couple of other folks in the Q&A module. The AI chat capability, when will the chat tool be available in some of the products or solutions that we offer? The chat tool is already available in the D&B.AI Labs. It will be the output that we are creating out of the D&B.AI Labs that will be produced or will be replicated in some of the products that we showed. The chat D&B bot itself will also be available in certain products and services, which we're going to be deploying. For example, D&B Hoovers will have that product available, will have that capability available in it. The next question that came in is: How is Dun & Bradstreet using AI for its own global data sourcing? That's again, a great question. We're absolutely, we're very excited that we can leverage these things for data sourcing. For example, we've used these capabilities, I apologize, I'm losing my voice. We've used these capabilities to greatly improve our ability to source URLs from the web. There are other areas where we're leveraging them. Okay. I think at the moment, those are the questions. I'm just gonna give people an opportunity to perhaps see if there are some additional questions that have come in. Let me just check the console here. Another question that's come in: Can the AI model interact with customers' workspaces directly? For example, add a report to their portfolio. We absolutely envision that customers will be able to kind of, like, come into the lab and use the tools, but API out to their own tools and data. This is something that we absolutely plan to enable. Okay, terrific. We'll give it another minute to see if some additional questions come in. I think I see a really good question from Mr. Burr: How is the data kept fresh and up-to-date for the LLMs? Excuse me. This is a very important question, and our LLMs, because they directly access the data where we keep it, so we actually, we don't really need to do almost anything. If we add new data to those locations, the LLMs, they're smart enough to change their view of the answer based on the data that we add. Which is really great. This is really one of the drivers of our ability to prototype things very quickly, because we don't really need to worry about, well, now we need to recode stuff because we added three months of data. Great. Great. There are a couple of other questions in here. Let's see. The next question comes from Himadri. Has D&B created its own embeddings, or are they using third-party embeddings? We absolutely have the capability to create our own embeddings and train our own models, and we have done a lot of that stuff. The sort of like the fundamental atom, the driving principle in our decisions, which tools and technologies to use, will always be: What does the customer need? What kind of tools does the customer need to get the answer to the questions that they're trying to get? Sometimes we'll use things off the shelf, sometimes we'll fine-tune them. We will always pick the best and the most optimal technology that serves the customer. Okay. Another question that's come in is: Can the AI tool help deploying data solutions by programming interfaces to ERP, for example? That's a wonderful question. It's definitely something on our radar. I don't see why the answer is no. It absolutely can. I think as we're going through, Gary alluded, we have so many use cases, really great use cases to go through. I think this one is a great one. We'll focus on it going forward. I'll ask the question, though I know that this is one that will be so dependent on the use case. The question did come in around pricing. How would pricing evolve over time for customers leveraging our AI resources in the lab with products? Gary, I'll let you comment on that one. You know, pricing will evolve, obviously, as we start to deploy these capabilities to our products and services. You know, as we either create new capabilities or enhance existing capabilities, and, you know, again, from what Michele was alluding to, it really depends on the use cases. As we start to release and productionize, you know, we'll be letting customers know, obviously, you know, what the, what, if any, commercial changes and pricing changes there will be based on these new features. Thanks, Gary. The next question reads: Will the agent also use customer data or files in combination with Dun & Bradstreet data, which would then be connected to the platform? Again, wonderful question. It absolutely can, and we would be very excited if that is the case. As I mentioned, there is an option for customers to bring in their data or third-party data into the platform. Of course, if that happens, we take all the, you know, all the necessary security and governance precautions, but once it's there, we absolutely can deploy these agents on customer data and third-party. Right. Eventually, there will be another variation of the lab, where it will be on the customer's premises and in their instance of whatever cloud they have. Hopefully, they, you know, it will have all of our data and all of our models. Mm-hmm. We'll also be able to, you know, if they give us access, we can certainly help them achieve some of their goals, through sort of that, those means, if that makes sense. Gary, I know we touched on this earlier, but it's great to see these questions. We have a couple of questions regarding whether the service is already available for customers to use and/or how they could inquire for more. Maybe you would just want to repeat the information about that. Sure. Pretty simple. Please contact your account rep, representative from D&B, and/or also please put your information down into the Q&A section, and we will be reaching out to you. The next question is, just quickly, sort of a secondary to that, question is: How does this differ from, the Analytics Studio that, we offer? This is an extension of our Analytics Studio. It's basically, a variation of these existing capabilities. Which is also the reason why a lot of our customers, will be used to that kind of experience, and, because they already have the Studio and, already using it for a multitude of different things. All right. We'll give it just another moment. Let's see what other questions are here. Another question is: Do you need a specific product to access the beta? No. Right now, the capability exists in the lab. That's where, you know, you will come in, and we'll set you up and experiment. Beta versions will be available based on that. Yeah, there's a very specific question here regarding small businesses. You know, small businesses with a low digital footprint and low transactional data definitely require some analysis, as we know. How would your solution work to kind of close that gap and help better assess small businesses? Wonderful, wonderful question. Very important for us. Small businesses definitely struggle with interpretation of some of the more complex data sets, which is why that the second demo, the what we call human readable, is a great opportunity to improve the experience that small business customers have with us even more. You can very easily imagine, like, whatever products that they use with D&B, if we deploy these capabilities to make the output, like, even more clear and digestible and accessible to users who are not very technical. That's a very clear opportunity for us to improve their experience. As we've shown, these tools are extremely good at just enabling a chat with documents, right? We have a tremendous breadth and depth of educational experience for small business customers in the credit space. These capabilities, we can deploy them, and we're actually planning to do it very soon. We can deploy them to allow these small business customers just to get smarter on the space without, you know, like, without needing to code or without. It's just a chat experience. All right. Well, we're just about at the top of the hour, so I think we will conclude today's session here. On behalf of Gary, Ilya, and our entire team at Dun & Bradstreet, I certainly want to thank all of you for joining us. A recording of the webinar is going to be made available for viewing shortly. Certainly, again, if you're interested in learning more, I know a few of you have already sent your name in through the question console and your contact information. We do appreciate that, and we will be sure to follow up with you. Thanks for joining us, and we wish you a fantastic day. Thanks for joining us. Thank you. Thank you very much.