Evolving Data Insights With Privacy at Mastercard

We always like to start with thank yous at Databricks. So thank you all so much for coming. I actually see some people that I know came to one of my first talks on this topic like three years ago now. So it's fun to be part of the journey with you guys and thanks for coming again. And obviously thank you to John for always being a great partner in these things and participating again in the madness. So cheers. So it's been three years. So just a little bit of help. There's been some talk about like clean rooms. Oh, shoot. I fit into MassGard. So I'm on the strategy side of this. So I get to play with Spencer for the last three years about this concept. Yep, and I am Spencer Cook. I'm a lead solution architect at Databricks on payments. So what's been fun about this is the bit of the journey and it's been really cool to see so many things happening with clean rooms and the whole concepts around private sharing of data. Just from a context to make this relevant for you because Spencer and I have done this a few times to try to get back to it. Like who here is currently using clean rooms or wants to use a clean room? Sort of help me out a little. Okay, that's awesome. The other thing is from this was if there were some pieces about this morning that Greg Ulrich, who is our head of data and AI talked about some things. We're gonna have time for questions at the end but also they're gonna kick us out. So if you have questions after you find Spencer, I about it. I've got this really hard to find jacket instead of shoes. So like you'll never see me again in the crowd. Yeah, I had one too and Greg is now wearing it. So yeah, make sure your boss has the swag is what I've learned and I have a pen. But in all seriousness, if there's been a big thing about this journey we've had about the people, the platform and the partnership with Databricks. So if you haven't had that much of a chance to work with the teams, I cannot speak highly enough about it because this was a really fluid and dynamic deliverable that we made to put privacy capabilities in the clean room. And we learned a lot about the journey which we're gonna talk a bit about where that fits in to the AI curve and everything else today. But as we ground all this, if you think about the Databricks platform, it is the basis for our ethical data use and responsible AI. If that is of interest to you, you can just Google it. You will see MasterCard has principles and blueprints for all of that already. So we also talk about taking a privacy first approach to data governance and that helps enable private insights. And that's the basis of what we have together here. Yeah, so where John and I kind of started taking yourself back three or four data and AI summits ago, the biggest problems teams were having, it wasn't necessarily getting the latest agentic system to completely replace all of your employees. That was still like literally in the realm of science fiction, right? What MasterCard and tons of other people were looking at was basically how to get from BI to predictive analytics to having any sort of self-improving automated system or intelligence. And so Databricks spent a lot of time trying to connect that chasm. And basically it came down to those three different gaps of like getting to what will happen, how should we respond and then making the best decision. And our point of view was that each one of those had kind of some unique characteristics of why teams failed at that particular chasm. And so what John and I have kind of gone through at this point is like, sorry, you heard Greg on stage, it's public, MasterCard is at the top of the curve. And they're basically setting the standard for how to do these things. And so when we were reflecting back, it's like, oh, wow, what did we really do over the last 36 months to achieve this kind of result? And a lot of it came back to things you might not necessarily expect, things like governance, things like privacy controls, et cetera, that aren't necessarily part of some flashy chatbot, but turns out it's the actual stuff that makes all this possible. So with this, I'm assuming most of you are familiar with the data maturity curve. This is the basis of what we're gonna talk about for an AI maturity curve as well. And when Spencer and I went into this, we didn't know how well this would play out, but the one, two and three, that's the chasm, that's the gap that we're trying to cross along the way. And it's okay to be at certain parts of the sort of journey, but most people wanna get to the top along the way, the view's nicer and you can have a heck of a lot more fun coming down. And that's what set us up for the next piece here. So I have another question for you. How many here like ski and or snowboard? Yes, this is gonna make sense to you. Okay, so what's really cool about this, this is, and if you've ever seen June's presentations, he uses all like AI generated images. That is not, that is also not me, because that's a skier. With that being said, when I'm not on stage doing this stuff with Spencer, I teach snowboarding. I've been teaching it since college, so it's quite a while. And the lessons that we have here is kind of like what I learned on the mountain. I talk about things, it's all about fundamentals. What you do on the trail, you take to the trees, and you execute in the terrain park. Those fundamentals, that the simple things that you do in some sports work. Now, we're talking about gaps and chasms. Why not put a 30 foot table top? Because sometimes that's what it's gonna feel like, this works through. But along the way, there's information to be had. And for those of you who studied physics, or at least hurled yourself in the air, that bottom picture may be familiar. But what's really fun about this is when you think about this brave new world and how it all converges, it's really interesting about how much information we take in and we use. And that information is context, metadata, and controls. So if you've ever been through a terrain park and seen it, you can imagine, and with that curve underneath, there's pieces about environment, information, and action. That environment coming through, what's happening with the weather? What's happening there at that takeoff? What do you need to do for that information? Which way are you gonna lean one way or another? And then lastly, the actions you're gonna take, what you need to do in the air to stick the landing. And when you think about the AI curve, and we're gonna talk about how do we stick that landing and get things to production? And this is what's really fun along the way. And I can totally geek out on this if you have other questions. I could talk to you all about the physics of snowboarding and not killing yourself in a tabletop. But with this, there are controls. We talk about the context and the metadata to support it. Yeah, I mean, really what it comes down to me is everyone can strap on a set of skis like a bachelor party and aim down, and it's probably gonna work out. But if we wanna repeat this stuff at scale, it turns out that there's basically some common methodologies, some approaches you can take that make it safer, more successful, and like John said, more fun because you're kind of leading your own journey. Yeah, go ahead. No, so when we think about that journey, now think about back to these gaps. Spencer, why don't you run with this a bit? Yeah, so it's basically like as Gen AI became the type of AI that these conferences talked about four or five years ago, and it kind of showed up overnight, we've really learned a lot just in the last 12 months. So really one of the first things that Databricks delivered for enterprises coming up on two years ago was RAG and the ability to dump a huge context window into a chatbot. Well, if you're listening to Tanwen's talk today or Greg's interview with Tanwen, we talked about how that's no longer sufficient for the types of use cases that most enterprises have. And so it turns out that there's a new set of chasms that we have to go through. Now, I'll go ahead and brag again, as you heard on stage, MasterCard is actually deploying agentic systems at scale and they're creating agent factories. And so they're also, when we kind of took a step back, it's like, oh, we're also at the peak already for AI. This is fun. And so it's kind of like, how did that happen? And it turns out that a lot of the same things that we had done strategically to get ready for the original data maturity curve put us in a really strategic position to accelerate our journey on the AI maturity. So that's where the data curve and the AI curve makes sense. It's the same foundational elements that we're trying to, hopefully you guys are starting to see, of governance and lake houses, prepare you for both of those curves. So to make a bit of a connection here, if you think about that context, that's the environment, right? That environment, that's your operational reality. That's the practical things that happen, whether it be from the people and the culture and the organization to make it happen. Now you think about the metadata and we talked about along the way, that's all the information that's happening within the systems, the system interactions that you need for the data and for AI along the way. And it's that flow, that feedback of the people, the process and the human in the loop. Then lastly, those controls. What are those actions you need to take to control as you work your way up the curve and the increasing complexity that they need to happen? What's the strategic implementation you need to have for these controls? The deployment for the policies, the governance, the frameworks that help ethical and responsible deployment for scaling across the enterprise. So if we think about it, those increasing levels of complexity of working your way through, hopefully you're starting to make some sense. Now, the context, the metadata and the controls are needed for the privacy insights are more complex as well, the further you go up. Yep. And so John and I, it turns out, I'll go ahead and do the drill. It turns out we have a very different experience at the slopes. So when he's talking about launching himself off the curve, it made sense, but to me, it's almost like the different levels of complexity of going from, you know, you start with greens on the bunny hill, you might do some steeper stuff. It's a learning development area, it's not a bunny hill. Sorry, we practice this. The learning development area, yes. So once you progress past that, you can safely maybe take a more challenging blue if it's something you've seen before. But eventually we all want to get to the point where we're at the top of the mountain and we can kind of pick what we want to do. So if you think about it visually, you start at the bottom and you're using things like business intelligence tools to solve certain use cases. We've been talking about for years, the challenges and taking those to production. And that's led us into advanced analytics, which is awesome, but it can be a little bit of a trough. I'll pick on myself here. I am stuck on the blues and don't have a ton of interest in going to the black diamonds. Maybe John will help me, right? And a lot of AI companies or companies that want to be AI, they end up in the same spot where you kind of are doing advanced analytics, you have a cool story, but you're not able to really take control. And where we all want to get is to the summit of AI because what's beautiful is when you're at the top of the mountain, you can actually say, oh, I want to go see the view over there. I'm going to meet you at the middle of the mountain and get barbecued. You are in control of your route and of your destination. You're really embracing this analogy, you really are. But visually, when I think about your architecture, because it's a platform, it's not built for a use case, someone coming in, they can kind of plan what they want to do in the same way I might plan a run, where it's like, I need to do this, this, and this. Do I have the right controls? Do I have the right context? And then they're basically picking this decision tree. So to that point, if you think about if the RAG's straightforward and the data's fixed, right, and the agents are more complicated, you can imagine the privacy constructs need to be completely increased as well. And the complexity gets even more when you start to look at Gentic and have that multi-dimensional controls that need to happen in that distributed environment. So more sophisticated orchestration, along with that comprehensive audit and consent along the way. Absolutely, and it lets you do, like if I saw you snowboarding, I probably would have a heart attack. But with you, it's like, I'm safe, I'm in control, and it's because you put the right systems in place. So to this, we're beating this one up, but it's really that increasing levels of complexity is the reason why MasterCard was able to get things to prod. And I'll make you bet, half this room has your data to the cloud faster than us. And I can say that because the majority of the tech teams over at another talk will be coming in soon. But the reason why we're able to get things to prod is actually having the controls between governance and value. Privacy first, privacy by design makes a fundamental difference. It's really hard to get your data to the cloud. It's even harder to get things launched into prod. All right, all right, so MasterCard. So for those of you who may or may not know us, we're grounded in trust. If you think about us, financial services powering the global economy. Trust is the core of our business. And we've expanded in many different ways to help forward a digital economy for all. So let's see, I gotta back up for this. Oh no, or this. There we go. So our expansion has happened. The expansion in payments, products and services as a trusted provider of insights. And part of these growth areas now supports our data strategy with evolving data uses. And part of this is our data and AI innovation that you see here. We talk about it in four ways that Greg talked this morning about safer, smarter, more personalized and stronger. But the challenge is with this happening, also grows our data profile. So in the last 10 years, we've expanded beyond the card realms. We've expanded exponentially. And that expansion is in three ways. Volume, variety and velocity of the data of which it's access, process and action. And if you think about some of the things that we're trying to do here, it's the strategy that manages these data risks for that expanded data profile. So part of this as we've grown through with a privacy first approach and the data intelligence platform has helped making this possible because we don't know what's coming next. But it's that investment in those controls along time is the basis for how we've grown. But now as we've grown, there we go, it worked through. So have the challenges. I'm in front of this, it just doesn't work. So if you think about the challenges for us, they fall into three areas. This is a simple build. If you think about it from a governance, deployment, but then also the sharing that needs to happen along the way. So those data combinations. And that's why the context metadating controls are needed. And that's why governance is first. Yeah, go ahead and do the Databricks build. So essentially when John and I first started working on these things, we kind of figured, these are the major tent poles of problems that we needed to address. And it mapped to three parts of the Databricks. So it's now the data intelligence platform, what was a Lakehouse platform at the time. And basically those areas were the unity catalog governance layer, the mosaic AI layer, and then clean rooms as sort of the tip of the spear for collaboration. Obviously clean rooms has a ton of stuff going on. It's like access to Delta sharing, et cetera. And so basically that family of capabilities kind of put us in a really good position to try and address these problems holistically. So when we addressed a data governance problem, it was in support of mosaic AI application and development. We were lifting all shifts. When we would implement a privacy enhancing control in the clean room, all of those worked consistently with both ML and data applications. So we were making sure that everything we're doing for the end-to-end AI platform worked with the secure data collaboration. And so basically you were able to get the data and AI governance, the secure delivery, the different combinations that you needed. And really that having all three of those addressed was where the greater than the sum of its parts piece came in, because that was really the problem. Totally, and as Greg mentioned earlier today, and I didn't know he was gonna be talking about this, but it actually worked out really well. So you guys have some context. He talked about one of the middle ones we talked about, and we have to rename it from mosaic to their agent. No, just mosaic, that's the platform. Okay, so we're gonna call it that. So when we think about the MasterCard assistant platform, it does three real things, as you talked about. It's got that privacy-first architecture, because you think about it's gonna do the federated data governance and quality controls at scale. The second thing is that trust-by-design embedding of the sensitive information while the compliance tagging and management. And the third is that enterprise consent capabilities at scale so that the usage aligns with the customer protections. The last circle there, we're gonna talk about the cleanrooms later on, but hopefully you see the way the AI piece follows the work that we did on the AI governance. All right, journey falls into three Cs, consumer, creator, culture. I'm gonna walk through this pretty fast because some context is set. So for MasterCard, we are a consumer of Databricks. Okay, and let's assume that all of you are. That is the first thing. How the lake house is the hub of our hub-and-spoke efforts, right? So if you think about it, we've taken the strategy and data and governance goals, matched it to the right infrastructure that you see here from a cloud perspective with that privacy-first approach. And this is where we think about how do you break down those silos of centralization? Because even though the data may be centralized, it still needs efforts to be governed, available, and consumable. Next falls into the creator. What are you all building with Databricks? Are you building for internal or external? So we have some products and services that we're offering to internal teams and clients. One of them is that privacy-preserving data cleanroom. And it follows a few things, the Delta Lake for the organization and unification, excuse me, the Unity Catalog for its governance, but then also the sharing for data access. In addition, you can throw in some AIBI for the reports, and you wanna have some more fun, you could even use Mosaic for some chatbot experiences we have, and it understands that a pet is not a cat or a dog, but a pet is a privacy-enhancing technology along the way. So part of this tool provides some of the basics that we move beyond into now the culture. At Mastercard, we like to say we innovate by default. So what does that really mean? Focusing on use cases where we have a distinct advantage with our first party data, and the insights to make these frameworks available. So that gets back down to what I talked earlier about why we tend to work so well together is there's a match on the culture, but it's also the people, the platform, and the partnership that makes this happen. I think you had an experience as well that you wanted to share on this. Yeah, so I get to spend a lot of time in Mastercard's innovation centers, and they are all very nice, which is lovely. There is one in Manhattan that always strikes me because it's dedicated to one of your ex-CTOs. And it basically talks about the way that Mastercard innovated in the 90s. Like back when you were seeing the priceless commercials and everything else, we didn't have stuff like Databricks, we didn't have AI. So like, how were you guys able to build that empire when you couldn't do a fast chat GPT, what to do, right? And one of the things that he talks about in this little tribute is doing well by doing good, right? And so that really struck me in the way that John approached designing this platform is that if we do the right things, if we do good and we deliver things that work well for our customers, our clients, our partners, it tends to make us do well as a company. And so they brought that culturally to the cloud when we were first trying to design some of these things. And again, like the outcomes are undeniable, even for use cases that we wouldn't like put in a privacy enhancing technology camp, they've been so positively influenced in a fundamental way by these principles. And I give you guys so much credit for it being a part of your culture. So to that, as well as a takeaway for you all, there's other three things we do well in the culture. We have a set of data and AI responsibility principles and they're available on the web. Second thing is we have an AI governance framework that's continuing to evolve and modernize as we do. The third, which we just launched, and that's actually a racetrack, this launch we have with F1, we call them our data and AI momentum sessions. So it's actually happening around the globe. We have teams of data and AI strategy leads presenting to the product teams about what data's available, what are the controls, what are the governance, what are the use cases? And actually one of them's happening right this week in St. Louis, our technology hub. So this goes round and round and these are things that you guys could take home as well about the culture that makes all of this really sort of happen by default. Cool? Yeah. I would say very cool. So these slides, believe it or not, I kind of introduced back when we were first announcing clean rooms over three years ago now, and it was to kind of talk about organizationally. I'm actually gonna do an adjustment real quick just to make my diagram bigger so I'm not craning my neck. What we kind of talked about was this idea that Databricks gives you full dimensions of collaboration and trust. And what we've kind of heard feedback on when I've shared this with other people and we've kind of worked with it over time internally with the MasterCard is really it's less about trust and more about context, like John was mentioning. There are times where I trust you, but because of a failure of my own not delivering you the right context, our use case might still be a failure, right? And so context becomes this key piece. And so basically, if you haven't seen this before, the idea is that the Databricks data intelligence platform is able to provide the full set of capabilities across both axes. And that kind of going back to the mountain analogy means you're not headed in one direction or stuck on one part of the axes because of limited capabilities. You can essentially build use cases across a full gamut of privacy capabilities. And that's basically what the MasterCard PetSlayer was, is a way of bringing that to business users and making it simpler to build privacy enhancing applications. So if you look at this, when we first started, I want high collaboration, but the context was trust. I don't trust you. I don't know you. I don't know better, I don't know you. It just didn't fit on the slide. So when we talk about that, if we're going to walk around and go to our lawyers, all right, I'm going to share data, but we don't trust them. All right, we're going to bring the data in, we're going to spin it up, we're going to do this. I'm going to send them a banded laptop, wait six months. It's like, dude, this is not going to work. So what we did with the Cleanroom with Databricks is that we actually took some capabilities around privacy enhancing tech. One of our vendors is here today. You'll learn more about that in another slide and actually made it available. So in the build, does this require anonymization? Does this require synthetic data? Does this require differential privacy? Just freaking run it in the Cleanroom. It's put in as a wheel file, execute it, whichever ones you need for your use case, because that's the context of the use case that's needed for the controls they need to execute and the metadata to make it happen. So hopefully you can see, that's why we have the Cleanroom with Pets is all the way up in the top left corner, but there are other solutions you may need as well. I'm consuming data directly from the marketplace. The other thing I do is source third party data for MasterCard. We use a lot of it to add context to our data and solutions. The marketplace is just fine. I can do data trials and commercialization right there. I don't always need a Cleanroom and that's where these factors come in really handy. Yep, absolutely. And a key thing to keep in mind that we talked about a couple years ago is that before the Cleanroom, that entire low context column was essentially impossible. And so there were entire use cases that depended on building trust, starting with low context so you could mature into high context. But because that was missing, we were just stuck. And so that was one of the chasms. And so basically having this full gamut of capabilities, starting with low context, going to high is really remarkable. When I went to go check on how we had presented this a couple of years ago, I was really struck by how little of this has really changed on the data side. It's actually been kind of cool to see these things all go GA, get significant use, but this kind of blueprint I think is pretty sturdy for data. Now, what we wanted to talk about this year that we've realized throughout this journey together responding to Gen AI, everything else is what is really the version of this matrix for AI systems because they're very different from analytical systems, BI systems. And so one of the things that we focus on here is replacing the capability or level of collaboration with modality because AI is autonomous. It doesn't necessarily need interactivity. So what AI needs is to be multimodal where you can have an agent that isn't just retrieving relational data but can also talk to the internet or maybe instead of just generating a paragraph, it can also generate an image. And so that's what we mean by modality. The BI platform, like you heard all morning with our announcements is a really rich multimodal platform for AI. So it gives you really rich AI interfaces. And then we're using pets within MasterCard to continue bringing the privacy enhanced focus to AI. So what we see just as an example is in a situation where you have low context and low modality, I might be handing you just a bucket of data with an instance profile and say, hey, I promise there's a model here, good luck. And for most of my career, that was kind of the state of affairs for data science. What we realize is that as we deliver more modality like when I've seen teams Delta share models, they spend so much less time in the POC proof of value phase they're able to actually go and say, oh, the model works just like you said, I see the schema, I registered it, what next? They kind of are moving on innovation. And then an example of full modality is like we saw this morning with model serving. If instead of even making you responsible for infrastructure, access management, et cetera, if I just serve it to you, all of a sudden you can build complicated applications on top of my platform. And so those are some of the things that this pet platform has allowed us to do for both data and AI. Now, going back down to low modality, what we've seen is when we add context, like with RAG serving, that's what originally allowed for the product onboarding assistant that Greg talked about this morning. That went live before agents were really in vogue. And so we were able to get to that kind of amazing capability through just adding more context to the same model. And so as you go further and take that to its natural conclusion, Databricks last year introduced the data intelligence platform concepts and basically made it so that all these different pieces of context that are floating around in Unity catalog can be turned into a product and consumed through Gene. And so now all of a sudden you have high context, high control, high structure. And what we can basically see is that as you continue to build out this structure, the value increases. So when we add things like marketplace models, when we look at things like models through a clean room, they're all basically giving me more modality through preserving privacy, through preserving context. And so, yeah, go ahead. So if you think about the clean rooms again, it's not just intersections of data. You can do model training, model serving in a clean room. It works. So when everybody else talks about clean rooms intersect data, think beyond what else you can do with that trusted serverless compute. Yes, there are some limitations based on the size of the nodes, but really the art of imagination is up to you of what you can try to pull off and if you get stuck, let them know. True. That's what we've done. So I really say it's not just data. There's a lot you can do in a clean room that you're gonna be hearing more and more about, probably from many other teams as well as they start to dig in and explore. Yep, I think it's gonna become a fairly standard practice for model collaboration in the same way that clean room products are really starting to gain traction. What's awesome is because of this architecture, we don't have to change anything. Basically this pattern that we made for data work for AI. And as Databricks has added capabilities like Unity Catalog tool capabilities, you can define really context-rich interfaces that add modality to your data AI platform. And then you can have an agentic system that ties all these different pieces together, dynamically applies privacy, dynamically brings in context, but it basically gives you the complete picture for having the AI capabilities you need. And the beauty is that it mirrors what we found for data AI and analytics. So you're not on a different mountain. You're kind of at the peak and you can still do analytics. You can still do BI, but you're doing it in this enhanced way. So I wanna get into the architecture, but go ahead. No, I'm just looking at the time. Yeah, totally. That's what I mean. So in closing, I wanted to kind of talk through literally what this kind of estate looks like from the bottom up. So when John and I first started working on this, it was how to put MasterCard data and client data together in a privacy enhancing or privacy preserving way, but that also created value for both parties. Totally. And so what we did was we brought together Delta sharing, what was at the time Databricks new serverless capabilities and Unity catalog. What's cool about serverless is it allows you to kind of do compute in a trusted environment. Databricks is kind of playing the trusted party because the other two don't necessarily know each other yet. So what we basically launched, announced three years ago, I think went GA two years ago, is the collaboration ecosystem that includes things like marketplace, clean rooms, et cetera. Basically that layer made it simple to use things like serverless compute, because you aren't designing your own API to shred the data every time. It happens automatically. And then what we worked closely on with John's team, and then also with partners like RE-ARC who are super awesome in terms of turning some of these things into systems and frameworks, we were able to build out a ton of unique capabilities like the pet ecosystem, like a bunch of self-service tools to run all this. Go ahead. Yeah, so the orange is what Databricks is gonna give you. The green is what you can build on top. Yes, thank you. So if you think about that as well, and this journey set us, so when Spence and I look back, we're like, dude, this actually kind of fits to the curve, because you've got the context of whether it be your first party data or your client's data, where you're taking third party data from a marketplace and you're jamming it into that use case, that's the base. And then you think about the metadata or all that information that's going to happen along the way. How are the controls I need in this next layer? Do I need to do, you could talk a little bit about some of the pets and other ones that we have there in a moment, but with this, it's really cool with the logging capabilities. You've got multiple Unity catalogs going on, and based upon the privacy risk and transparency you need, you can actually have full audit trails here. So I think there's one more build on that. Oh, yeah, yeah. Well, I had a point to make as well, where basically, well, one of the things we've talked about in designing this is, it tells the consumer creator story as well. Basically on the bottom is when MasterCard was consuming things like their data using Unity catalog, over time you started to create net new capabilities. And that's what those green boxes are. And then essentially you developed this complete self-service layer that gives you full access to modalities and context. And the current plan is to deliver that through an app, have it be consumed through an app. That's some of the things that we've demoed last year. And this is all top to bottom, integrated with MasterCard's Databricks controls for things like workflows, approval, governance, and in a more traditional way. So with this, you can think about based upon the use cases that you're having. You can do this even internally. So we'll use this with some of our acquisitions with moving data around. Use the clean room internally. It doesn't always need to be excellent. It can be between your own partners. This is a really easy way to get the insights. And you can think about, you can use those privacy tech on the way in. Do you need to synthesize or anonymize? And we have with Trata and Synthesize, one of our partners is here, Howso, as well. Synthesize on the way in, synthesize on the way out. What rules do I want to fire based upon the use case? That context makes so much, so helpful. And RE-ARC's been great with building out the different solution accelerators. You're all gonna see in months coming up based upon the flavors you have. And I've been pushing back, and you guys have been really helpful, is how do we make this so a business analyst can do this? That's that self-service layer. And then the privacy lawyers come back. You know, they're used to saying no. If you ever see a privacy lawyer say yes, take a picture. Because now they're saying yes and. Because and depends upon the use case and the context for what I need to do. And our lawyers love saying yes because they're always saying no. And this is so much easier for us. We have hundreds of use cases that are going on because the lawyers now understand that they can finally say yes and get the transparency. But really, where you guys want to push this, you really can. Call to action. Okay, wrapping this stuff up. Three things. When you think about investing the fundamentals, you better have a governance and privacy strategy. And the third one is, how do you utilize that context, metadata, and controls? I mean, the governance and controls, it's not an afterthought. It's not to get through security. It's not to pacify your lawyer. If you treat the governance and controls as a value function, you're going to get ready for AI. And I don't know what's coming next, but if you use the same enthusiasm of like AI bricks and everybody going crazy for governance, I mean, you may look a little psychotic if your lawyers start to geek out on all that. But it's really important that we do as well. Because at the end of the day, I mean, yes, I work at MassCard and I'm a snowboarder. I have no idea what's coming next. We sure know we're going to be able to tackle it because you've got your house in order and you understand enough about the data, the use cases, and whether this be even bringing data into Unity. I mean, we've got manifests running that have, you can use, if you're familiar with Unity, you can sit there and do the AI descriptions for the table. How do you have the context at the column and the row level for everything? For every data set that comes in your data usage, rights, controls, load it into Unity. And then you start to ask questions on it. Now imagine you can ask questions of what use cases can I do and then have it recommend this is better for a clean room. With these privacy controls, you're not far off. This gets really, really fun to getting ready for the future. And as I talked about here, just having those three things, use the right system at the right time. Absolutely. I think that whatever you and I end up doing next is going to involve CMC. There's no way it won't. And so invest in this now and you're going to be ready for the future. I think that's the story. Absolutely. Thank you so much, everyone. Email's there. And I have to do the plug, yes, to complete your surveys. Thank you so much, everyone.