Transcript
♪ ♪ ♪ ♪ ♪ ♪ ♪ ♪ This is one of the most magical periods of technological change that I have personally ever seen. You need an agentic control plan. This is a mission control center to ensure that every decision and action is governed, trusted, and grounded in your business context. Turning your ideas to prediction and doing that fast. We are in an age of ubiquitous intelligence. Each of us can have the power of a data scientist, an analyst, a statistician. All of this at our fingertips. For all of you builders out there, developers, data engineers, analysts in your organization, you can build anything. ♪ ♪ ♪ Please welcome Snowflake head of analytics, Carl Parrier. Welcome, everyone, and thanks for tuning in to join us today from around the world. We're excited to share this special edition of the Summit Builder Keynote today with you. Although we weren't able to share this with you personally as originally planned, the message of this particular keynote was far too important and frankly timely to not share with you. Because it is truly an incredible time to be a builder. We're in an era where you can go from an idea to something real and running in production faster than ever before. Coding assistants are getting better day by day. AI is revolutionizing how we actually write code, analyze data, and ship products that our customers love. As Vivek Raghunathan, our SVP of engineering, likes to say, this particular moment truly feels like a new renaissance age. More people than ever are able to jump in and build things with AI dramatically lowering the barrier between that idea that we all have and execution. And the only limit being our actual imagination. And truly, everyone is leaning in these days. Your customers are leaning in. Your colleagues are leaning in. Heck, I'm even sure your CEO is leaning in. I know that mine is leaning in aggressively. Everyone's excited. We're all building more and at a pace, frankly, that we never thought would be possible. But with this acceleration, something else is also happening. The environment that our builders operate in has exploded with complexity. More tools, more data, more workflows. And across every imaginable surface area. Frankly, it used to be that expertise, time, and sheer engineering capacity were the barriers to innovate at this scale and pace. But the true barrier today, it's this fragmentation that we all live in. It's not just messy. It's expensive. It slows decisions. It creates blind spots. And frankly, the result is that our best builders have to spend most of their time cobbling together and manually connecting these systems and tools. And if you're a builder, this probably feels way too familiar to you. But what if instead of wrangling all of this, you could actually just build? What if you could focus less on switching tools and systems together and more on the outcomes that you actually want to create? A couple of weeks ago at Snowflake Summit, you heard our vision for the agentic enterprise in our day one keynote by Sridhar. And in our day two keynote with Christian, he actually walked through all of the capabilities we're launching to help organizations solve real problems with data and AI. But today, today is about something importantly and entirely different. Today is about all of you. All the builders who actually turn these capabilities that we've launched into real outcomes and impact inside of all of your organizations. And making that possible requires three fundamental shifts for builders. So the first is that you need to build across, working across all your data, wherever it lives, across clouds, across formats, tools, and workflows, without this fragmentation actually slowing you down. Second, you need to build with AI as your true force multiplier, giving you leverage to go from that idea that you have to actually building something and being in production faster than you ever thought possible before. And third, you need to actually build for impact. It's so easy to build these days, you actually need to focus on turning what you create into outcomes that the entire organization and our customers benefit from. So let's go ahead and start with the first one, building across your data, across systems, tools, and workflows, without moving any of it. Because the reality is your data isn't all in one place. I talk to customers every single day, and they have a diversity of platforms where all their data is stored. It's spread across a ton of different systems. And the challenge isn't just accessing it. Of course, you can just open a connection, run a query, get the results. It's actually being able to work with it as a single data set, seamlessly and fluidly across all of it. And this is usually where we hit one of our very first bottlenecks. Every single organization likes to say that they're data-driven. You log into LinkedIn, you see a ton of companies with that in their byline or individuals. But frankly, as we all know, being truly data-driven is always easier said than done. While we all have a ton of data about our products and how our customers are using it and whether it's working or not, actually answering simple questions almost always requires us to stitch together pieces of information that are spread across a variety of these sources. You know, when your CEO would like to know what's happening on a specific product, and in order for you to get a clear answer, you suddenly find yourself stitching together numbers from multiple systems, reconciling the definitions of those numbers, opening five browser tabs and slacking three different people, trying to piece together an answer that frankly should be easy to find but frankly when I have what I think is the answer, I'm not even sure it's still accurate. So what's happened historically in the past? Frankly, the investigation just stops here. And it's not because the question wasn't important. Of course it was. Or that the answer won't help the business move forward meaningfully. It's just that it was too hard and time-consuming to actually validate that that's the right answer. But right now, all of it is changing in front of our eyes. The ability to work across all of your data, seamlessly, conversationally, and most important, in real time, is happening and becoming genuinely possible with Snowflake Cocoa. You know, I had one of these real wow moments myself recently where it all became very real for me. We have our quarterly business reviews where we talk about how analytics is doing, and in the conversation there was a question of like, well, how are we talking to new potential customers? How does our sales team talk to them about analytics? You know, is it forward-looking with AI? Or are we talking about reports and dashboards? So we took the action item to go follow up on this. And so after the meeting, I reached out to our sales team, and I said, hey, can you connect me with a bunch of our sales reps? I'd love to listen to some calls. I'd like to interview them and get a sense of them. And they said, yeah, yeah, we can go do that. But actually we have a number of transcripts. They're actually sitting inside of Snowflake. I was like, this is a treasure trove of data that's teeming with insight. So yeah, give me access. So they gave me access to these transcripts, and I looked at how much we had, and I was like, there is no way I'm going to actually be able to go through all of this. There's just way too much data here. I have my normal job, and I'm super busy with that to begin with. So I decided to try something different. Since all the transcripts are stored inside of Snowflake's data perimeter, and I've been given access to it, I decided to try to ask Coco. I prompted it to look through the full list of transcripts I had access to and analyze how our customers were being talked to from our sales team. Was it analytics-focused? How much was AI involved? What were the themes within those conversations? What resonated, what didn't? And I was absolutely blown away. Coco delivered. Within an hour, I had a full breakdown of what those conversations were about. Yes, in fact, they were AI-forward, talking about analytics and AI together. What the customer sentiment was of each one of the sub-themes that we were talking to them about, and then how the pitch had changed over time. You know, frankly, I would not have been able to do this myself in the past. It would have been impossible. I would have had to have reached out to our data science team. It would have gotten in their backlog. They're busy as well. Once it became a priority, they would have gone and investigated it. They would have created a model to cluster all of the feedback, created a bunch of features, had to operationalize it, get it up and running, keep it running in ongoing fashion. Super expensive. It would have taken weeks or months for them to do this. But I was able to, myself, in a few clicks and a number of well-ordered prompts, get a directionally accurate answer. Of course, I'm still working with our data science team to validate that. They're the experts and they understand the business and the data. But the speed at which I could go from that question to actually that meaningful directional answer using Cocoa was absolutely mind-boggling and, frankly, incredibly exciting. And that's really the shift for all of us. It's not just about accessing the data. It's being able to work across all of it in the moment to build something that's truly meaningful and impactful. So let's go ahead and make this real. Let's run a scenario together. Today, we're going to follow the fictional journey of a company. Let's call them Data Donuts. They have great products, the best donuts in the area. But, frankly, for some reason, they're seeing questionable retention. The CEO is actually looking at the numbers and she doesn't like what she's seeing. Customer churn is creeping up and she doesn't want to know in two weeks what happened in the previous two weeks and she doesn't want a static dashboard. She wants to actually understand in the moment what's happening and, frankly, what the company can do to save that customer. So she goes ahead and calls in the Dream Team. It is my pleasure to introduce you to James Cha Early, Senior Developer Advocate here at Snowflake, and Anastasia Stefanska, a Data Team Lead at TUI, and a data superhero who's joining us all the way from Berlin, Germany. Hi, Carl. Hey, James. Carl, have you seen Anna? No, actually, I haven't seen her. Do you know where she is? Anna, where are you? Oh, she's calling me on Zoom right now. Okay. I'm going to go ahead and take it. Okay, take that. Anna, my favorite data superhero. Where are you right now? We're supposed to be on stage. I know. Hey, James, my favorite developer advocate. We just got paged. What? No, no, no, no, no, no. We're supposed to be on stage right now for the keynote demo. I'm on stage right now. Can it wait? No, it cannot wait. Drop everything now. The Data Donuts C-Level meeting just ended, and then the message came in, and I knew it was trouble the moment I saw it because it started with, can you get me data real quick? No! So here's what we're dealing with. SEO wants a tool for the retention team so that customers writing to support who are at risk of never buying from us again get turned into satisfied customers, and they need it right now. Are you ready, James? I guess. Wait, wait, wait, wait. How are we going to do this right now? Don't worry. I got you covered. Here is the architecture diagram I put together so we know what we are dealing with. Let's walk through it together. We're building an app with a unified customer view and a real-time churn prediction, and we're using Cocoa to get it done fast. This is our Cocoa, so we set the rules. We must remind the architecture. Let me walk you through what we are building. Can you see my screen? Yeah. Our data is fragmented across three systems. Snowflake managed iceberg tables, GCP, and Databricks. Over 30 tables in total. We're not going to manually look through them. We are going to let automatic data discovery find the most relevant ones for us. We have support tickets, marketing, browsing, loyalty, purchases, returns, all scattered across three platforms. Now, here is the constraint that really matters. The solution needs to be interoperable, so we connect the data, but we never ever copy it back together to Snowflake. We use iceberg format and catalog link databases so the data stays exactly where it is, but we can treat it as a single source of truth. To do that, we built a dynamic iceberg table in Snowflake that links everything together. Oh, one more thing before we move on. Based on what I've seen in the data, we need to enrich it. Specifically, we're adding AI sentiment score using Cortex AI functions that give us a signal for how a customer is feeling when they write in, and that's the difference between a churn score and a churn score you can act on. And then we train. We are running several machine learning models in parallel using the AutoML skill, and we let the best one win. So no manual tuning, no guessing. The model earns its place. Every time a new support ticket comes in, we evaluate the churn probability in real time using streams and tasks. If that probability crosses the threshold, the retention team gets an alert in Slack immediately, and the underlying table refreshes hourly so the patterns stay current. And for the retention team working on the case, they need to find similar cases and spot the patterns across the data. And this is where we add Cortex Search and Semantic View for Cortex Analyst. The whole thing gets packaged as a Snowflake app runtime using the Cocoa Agent SDK and deployed with Snow app deployed. The retention team uses the app to handle individual churn cases, and the management sees anomaly trends in Slack. Everyone has what they need in real time. James, we're young and reckless. Let's run this as a one-shot prompt. I'm slacking it to you right now. All right, got it, Anna. I'm going to paste it into Cocoa, and let's go ahead and kick it off. Because this is a complex prompt, Cocoa starts by planning. It's going to use a team workflow skill which spins up a set of autonomous agents. Each one will own a single task. The dependencies between are mapped automatically. First, we need to make sure Cocoa understands every feature we've referenced. So a set of parallel agents fans out and explores them simultaneously. Once that's done, the output converges into a single planning agent that writes the building plan. Think of it as a team of superheroes. Each one has a role, and none of them are waiting on work they don't need to. Exactly, you're ready to rock. Thank you so much for helping me out. Look at all the magic we've made. You really are a data superhero. Actually, on that note, how did you become one? Oh, thank you, James. Thank you for asking. As a data analyst, my work historically came with clear scope and fixed boundaries, but I've always been a builder at heart. Snowflake brought it out in me. Cocoa removed the boundary for my analyst role and made me a builder again. So every why and how I can now start solving on my own. That feeling of empowerment is beyond words. But the empowerment effect was there all the way before AI. It started by me founding Women on Snowflake together with my friend Isabella to give examples to others who want to build. Then I became a squad member, and later I was eventually honored to join the data superhero program. And the Snowflake community is a safe and supportive space for me to try new things. There is always someone to learn and someone I can help. So the synergy is real. For example, I bring AI specialization. My fellow data superhero, Sophia, brings data engineering expertise. And another data superhero, Ruchi, is a great example of thought leadership. So together, we are stronger than every of us alone. And on this stage, I actually want to address the audience and everyone who is building on Snowflake. The moment to redefine what you do and who you are professionally is now. You can join a user group, connect with data superhero and squad member, or take Coco as your coding partner and build something. James, I've got to run now. You got this, and may the Coco be with you. Bye, Anna. Wow, you're not wrong, James. She truly is a data superhero. Thanks to you both. We have a one prompt demo. I love it. I can't wait to check in on Coco throughout the talk and see what happens. Thank you, James. Thank you. Anna and James just showed us what it looks like when builders can actually work across all their data. But this is not just about speed or ease of use. Those things are incredibly important. But it's also about trust. It's about the security, the governance, and the observability of your data. And there is zero room for error on any of these fronts as a builder walks through their journey. To share what this looks like in one of the most demanding regulatory environments in the world, I'd like to welcome Durgesh Das, VP of Data and Analytics and Governance at the NYSC. Hi, guys. Hey, Durgesh. How are you? I'm great. How are you doing, man? Good. So you operate in one of the most demanding regulatory environments in the world. Can you give us a sense of what this actually looks like at the scale and complexity of the data that you're working with at the Intercontinental Exchange at NYSE? Sure, Karl. Let me start with the number, $40 trillion. That's the market cap of companies listed on the New York Stock Exchange alone. Every single day, our market processed 10 million trades across 2 million listed contracts and securities. But NYSE is just one part of ICE, Intercontinental Exchange, which runs 13 exchanges and six clearinghouses globally. The data we are working with spans equities, energy futures, environmental markets, fixed income, and mortgage. Three completely different business segments, each with their own data models, their own latency requirements, their own consumers. On the Snowflake Data Cloud, over 350 accounts, 12 petabytes of data, 20 petabytes of data are scanned every month by more than 200 million queries, three clouds and 21 regions. The challenge isn't just the volume. It's that every one of those data domains has its own rules about who can access it, how it can be used, and what has to be audited. Bringing that together in a meaningful, usable way, that's the real problem we are solving, Karl. $40 trillion, that is mind-blowing to me. So your data is spread across multiple systems, and frankly, governance isn't optional. What made this especially challenging for you all? So the good news is we have actually solved the foundational governance problem. Snowflake's role-based access control gave us exactly what we needed. We have maintained that balance where data is accessible enough to be useful and controlled enough to be safe. But then AI changed the equation. As AI capabilities matured, the demand from business for meaningful self-serve analysis from data increased. I would say multifold. Teams that were previously happy waiting for a report now want to explore data themselves, ask their own questions, get answers in real time, and they want to use AI to do it. That's where the challenge became more stringent because now you're not just controlling who can see a report. You are governing how AI interacts with sensitive data at scale across the entire organization. The bar didn't go away. It got higher. That's, I mean, I can imagine it gets much higher there with so many people wanting to access it. So what changed when you started using Snowflake and Cocoa Desktop to work across all of that data that you have? Before this, our data lived across multiple places, Carl. Different dashboard, different tools, and getting any meaningful insight meant someone had to extract it, move it, or manually piece it together. That introduced risk, introduced delay, and frankly, introduced inconsistency. Snowflake gave us a single governed environment where all of the distributed data could live and be accessed together without anyone having to break security rules to get to it. That was the first unlock. Cocoa Desktop didn't sit on top of that as a separate layer. It plugs directly into Snowflake's existing governance model. The access control, the security policies, the data lineage, all of that carries through automatically. We didn't have to rebuild governance for AI. We extended the governance we already had. So now our user can explore, query, and ask complex questions right where the data already lives. No shadow copies, no can you send me the file, no waiting on a data team. The data stays put and the work comes to the data, not the other way around. So what does that enable your teams to do today, actually? It opened up access to data for business teams. They can now get to inside themselves without creating a ticket and waiting days for a response. That might sound like a small thing, Karl. It isn't. It fundamentally changes how those teams operate. The quality of questions change too. It's not just how many accounts do we have. It show me the revenue trend for this account, which, what products they own, where the risk is, what's changed in the last quarter. That's kind of cross-domain analysis used to require a data engineer and a project. Now it happens in the flow of work. And critically, governance didn't get traded away for that speed. Security, consistency, and access controls are maintained end-to-end. In our environment, that's non-negotiable. We got speed without sacrificing control. Yeah, being able to have both is frankly game-changing. Can you give us an example of how this actually shows up day-to-day in the work that your teams do? Think about a sales rep preparing for a major account meeting. Before, that mean bouncing between many tools, chasing down the latest number, maybe emailing someone on the data team. By the time they walked into the room, the prep had taken half a morning. And some of the data was already stale. Now everything they need is in one place. Account help, revenue, product usage, all consolidated into a single interface. No more jumping between tools. No more manual reconciliation. And if they still have a question mid-prep, we build a customer 360 skill directly in Cocoa Desktop. They ask it the way they would ask a colleague, and they get an answer grounded in the actual governance snowflake data. That's the real shift. It's not just that the prep got faster. It's that the sales rep can walk into a meeting fully prepared, all from one place, in the time it used to take just to find the right dashboard. So we've been talking about how AI can actually accelerate everything that we're doing. So how do you think about this acceleration in the highly regulated environment that you actually exist and operate in? We don't just think about it, Karl. We have a formal framework for it. ICE's responsible AI principle govern how AI is developed and deployed across the entire organization. The framework comes down to five things. Is the AI accurate? Is it secure? Is it fair? Is it explainable? And is it properly governed? Not aspirational values. Embedded into how we actually build and ship software. What makes Cocoa Desktop the right fit for us is that it operates within Snowflake's governance model. The access policy, the audit trails, the data protections, all of that extend directly to AI usage. We are not creating a new risk surface. We are extending a governance model we already trust. So where do you see this going next for your organization? We are expanding more teams, more use cases, more self-serve capabilities across the organization. The pattern we are following is start with governance architecture, then expand access. That's allowed us to move quickly and confidently because we know the foundation is solid. The longer-term vision is an organization where people closest to business, the ones who understand what question needs to be asked, can get to the data they need without friction and without risk. Where AI is a normal part of how we work, operating within the same principle that govern everything else we do. We are not done. We are at the beginning of what's possible when we can move fast and stay governed at the same time. And frankly, in a business like ours, where the stakes are as high as they get, that combination is everything. I would like to give a shout-out to my team at ICE and the Snowbreak team, whom we work very closely. Let's build together. Thank you so much. It's been awesome working with you. And I'm, like, blown away, frankly, about how you're operating at the scale you operate at, which there are very few, on that secure, governed foundation. It's really accelerated everything you do. So thank you so much for coming today. Thank you, Dol. It's been a pleasure. Talk to you later. Thank you. Thank you. We just saw what it looked like when builders can truly work across all their data, across systems, across tools, across formats, and across workflows, connecting, querying, and actually building without any replatforming involved. And as our guest just showed us, you can do all that without impacting the governance, security, and control. At Snowflake, we understand that your particular world is extremely complex. And that's why openness and interoperability are foundational to how we go about building our platform, from Apache Iceberg tables for open table formats to Apache Polaris for open data catalogs. And finally, the open semantic interchange that makes your lives easier so that the semantics you have flow across all the systems that you need to work on. And frankly, interoperability doesn't happen by accident. It's powered by those communities, standards, and foundations that are advancing these open technologies across the entire set of industries here. And that's why you'll continue to see Snowflake support organizations like these. Thank you for everything that you do. And that's what it really means to build across your data. So now you have everything, right? Your data connected and ready to use. But here's the thing. Because once you have your data, the problem isn't access anymore. It's actually figuring out what matters. What's gonna predict an outcome? What's signal and what's noise? And what's the pattern that's hiding inside of all this data I have? And historically, answering these questions have taken long lead times, a ton of manual effort, and huge teams. Building models, stitching together workflows, trying to connect signals across all of these systems fast enough, and that's critical. It needs to be fast enough to actually do something useful with that. But AI changes. But AI changes all of that. I have a PM on my team. His name is Rudy. And his job is to focus on helping make Snowflake faster and more cost-effective for customers. And when customers are looking at bringing on new workloads to Snowflake, they go ahead and run either that workload, a synthetic version of it, or benchmarks that are available publicly. And sometimes, customers will run these evaluations, and they're unintentionally flawed. Maybe the workload actually ended up having configuration that wasn't the best for that particular workload. Or maybe they actually run the test from a constrained network environment when it should really be running in the cloud next to the platforms that they're working with. Or maybe they copied an outdated benchmark online. And usually, when those things happen, Rudy and the team only find out days or weeks later, after the customer's already spent time testing, and frankly, in many of these cases, end up with a flawed result with Snowflake. So Rudy asked a very simple question. Can Cocoa actually help us find the pattern? Could it recognize that a customer was running one of these workloads and detect, actually, that something was wrong with the test that didn't look right? So in a couple of days, he built a prototype that could identify these benchmarking workloads in real time, detect the signals that something was off, and then we could talk to our account teams to work with the customer in the moment to help them end up with a better result. So instead of us reacting weeks later, we could step in during that exact moment, help the customer run better tests, end up with better results, and frankly, avoid wasting a ton of time actually evaluating misleading outcomes. And that's really what this next demo is about. Once AI can actually help us find the pattern across systems, behaviors, and signals, you can always start solving problems in a completely different way. James, over to you. Thanks, Carl. So as we build, it's important for us to catch where we're going wrong early so we can address it. An expectation of what we build today is something that lives beyond a single interaction. It's going to be something that will continue to live and inform how we work. So prospects who are likely to churn are brought to our attention immediately on a platform we use every day, Slack, so we can quickly make our moves and not wait for an escalation. So earlier, we asked Coco to build us an end-to-end churn detection and Slack alerting system for data donuts and to use a team of agents to do it. So before any code gets written, it comes back with this plan. So let's go ahead and walk through it. So phase A is the foundation. It creates four schemas, the AI, features, ML, and agent, plus everything else we need for Slack. It mounts our iceberg tables on Google Cloud Storage through Cloud-linked databases. That's a live connection to our open catalog, so no data copying there. Then we do the same for Databricks Unity catalog. Both sources are queryable instantly as they were native Snowflake tables. Next, we have phase B. And this is where AI features get pre-computed. One thing I wanted to call out here is I asked for AI sentiments, but it looks like it decided to use something else. It decided to pivot to AI complete with a JSON schema to get the floats I asked for. It's not exactly what I asked for, but it had a valid reason to switch. So let's go ahead and let it run with AI complete. Phase B indicates there was also a fix, which we'll review later. It states it will create two append-only streams on the support tickets table. Then we have phase C, which will create the dynamic iceberg table called churn features. Then in phase D, this is where our model gets built out with AutoML. It should do the data prep and any additional feature engineering, train the models, do some hyperparameter tuning, and eventually evaluate the model performance. Then it will host the best one for us to use. Phase E is the hot path work. So it will create the procedure that fires when a ticket lands. It will score it and the customer itself and see if that customer is at risk. And if it is, Google will do a research and then send an alert back to Slack. Phase F and G are the tools we will provide to our Cocoa agent. So we have Cortex search service over the ticket subject, description, comments, and semantic view for Cortex analysts. Phase I is autonomous monitoring agents. So deploy via Snow app deploy. And there's one correction I do want to call out here. It says it's going to use Cortex agent rest API for the reasoning, which is not what we asked for and not what we want. I wanted to use the Cocoa agent SDK. So we'll take care of that later. Phase J is the verification step, which I think is one of the most important parts. It will do a 12 check, including an end-to-end verification that will do a critical cancellation ticket that has to flow all the way through, all the way into Slack. And then now we have our revisor-driven revisions. Now, this is where we see the stream fix it caught earlier on. It says it will cause a race condition between the feature pipeline and the hot path, which is why we're going to do the two append-only streams. And look, it caught the SDK issue as well. So that's fantastic. Then we have the verification harness. We'll make sure that every key piece of the project is solid. And then, ooh, this is the best part. Cocoa splits this up into a team of agents. So it's going to work in parallel. So there's going to be three implementers. One will own phase A. Next one will own phase B through G. And the last one will own the native app itself. Then a tester agent runs the 12 checks. Then a reviewer agent does an independent end-to-end audit to ensure everything is followed by plan. So let's go ahead and accept this plan and let the team loose. So what we are doing here with Cocoa today is admittedly a bit ambitious compared to how it's used today. But we wanted to see what the world of possible is. So when we build with Snowflake Cocoa, we saw how to build out a whole plan and validate each step. And as part of the team, the validator caught issues on race conditions of the stream, how dynamic iceberg table would have been a full refresh, would have been very costly, and the whole Cortex API versus the Cocoa agent SDK issue. And finally, it's split it up into tests to do into independent agents to do multiple tests at once. And to remove confirmation bias, the reviewer and the tester are different agents as well. Well, thanks, James. What we just saw in that demo is that the barrier between the question that we have and a working solution has completely collapsed. You can go from those raw signals to identifying a pattern to actually building something operational around it in minutes or days, not in months like it used to take. But of course, the real question is this. What does this look like beyond the stage here at enterprise scale across thousands of users, teams, and decisions? Our next guest is Ravi Achakola, executive senior IT director, data and analytics at Medtronic. Unfortunately, Ravi wasn't able to join us in person today for the live broadcast. But I was lucky enough to be able to catch up with him last week. Let's roll the video. Ravi, before we get into the technology, can you give us a sense of what your data environment at Medtronic looks like? Yes, sure, Karl. I'm happy to share the history of all the Medtronic. Medtronic being a leader in medtech industry and the mission they're carrying upon is elevate pain, restore health, and extend life. We serve patients across 150 plus countries. So given the company has grown to mergers and acquisition, and in all, you already always know that when mergers acquisition happens, we incur a lot of technical debt getting the data systems and multiple systems. To give you an example about we had a data legacy system prior to Snowflake, like Oracle, Hadoop system, and multiple systems were there in place. And significant data sprawl and fragmented logic and data silos, as you can see about the magnitude where we serve multiple regions in the world, teams spend a lot of time searching, stitching the data instead of acting on it. So even before AI, what was slowing your teams down day to day? As legacy engineering tools and technologies were one of the things increase the turnaround time and limited agility in the space. And this created a persistent bottlenecks and delay in any inefficiency around the period. So team and whenever business comes and ask us time to speed to market, if I may say a time to market was pretty slow. So that was, we were not able to keep up pace with what business was asking for. So how did you actually start to rethink that model with the businesses? Great question. Given the complexity of a data ecosystem, we re-imagined and model around the three core shift with looking at it. One is unified data. What it means is establishing interconnected ecosystem, clean, trusted and AI ready data. The second strategy around that was shift left strategy means build your semantic layers on the data platform rather than building it in your BI intelligence world, which used to be the case. And the third pillar of this one is AI powered BI. Transition from a static dashboard to conversational AI. Those are the three models which we came up with and we are in a mode. A bottom line, what I say that the tagline we use here is that go beyond your dashboard with AI in BI world. How artificial intelligence is going to change the business and that's great. That totally makes sense to me and resonates. So how does Cocoa come into this? Like what changed for your engineering teams? Cocoa accelerated, Cocoa and Co-Works together. I always say this is really a big in our ecosystem brought together at, I always say that data engineering on steroids is Cocoa plus Co-Works, but Cocoa in general accelerated our data consolidation and modernization and enable the migration of semantic layers which were budding in Power BI, Tableau, Business Object, different visualization tool. At this speed, you cannot imagine the past. If you go with 50, 60,000 reports, it takes forever, but that is helping us to really get the semantic logic built onto the data platform which happens to be Snowflake here. Combined with Co-Works, I could say that it unlocked the AI driven BI and agentic capability. That's really cool. So, I mean, when you think about your teams day to day, what does it actually look like for them? Now with the Cocoa and Co-Works, the faster iteration and time to market that instead of days and weeks we spend, it's just few hours you can get it done if the data is in your data platform. And especially I want to call out here, it transformed the data engineering life cycle completely, like from a pipeline development, code generation, code reviews, and automatic data quality checks and continuous deployment. So in all in all, it really has helped us to make this data engineering work more like a commodity, easy to go. And it isn't just engineers, right? How is this changing things for the rest of the business? Given the technologies, if you say Cocoa and all, it's not the hardcore technical aspects. The business team can build in a matter of hours instead of weeks and months, and they don't have to be a technology expert in this space. So it's clearly shift from a dashboard to AI-driven decision-making to conversation AI. Users are now able to talk directly to the data and understand even on the mobile phone. So yeah, at Summit, we talked a lot about the AI era. So how do you think about the shift in that context? It's great question. I can go on and on for talking about AI, but I will brief it here that AI is shifting from experimental to real production scale. It is no longer that we are doing a prototypes or POCs, it is hitting the market and it's real. Delivering actionable insights through agentic AI capability is a second layer of it. Fundamentally, what I think is that AI has transformed how we do software engineering and data engineering landscape completely. Especially I would like to call out evolving from SDLC process, from requirement to deployment, Cocoa and other LLMs have really transformed leveraging this practices like white coding and all. Well, Ravi, it has been an incredible pleasure to have you here today. Thank you so much for joining us and sharing your story. Thank you, Karl, for this opportunity. I'm glad I was able to share my storyline, what AI is experiencing at Medtronic. Give him a round of applause. So Ravi's story is actually one of the best examples that I've seen of what happens when an organization actually operationalizes AI across their entire business. It's not just a tool that people are playing with, experimenting with, or tinkering with, it's actually changing the way they do their work. When we say that you, the builders, are powering the agentic enterprise, this is exactly what we're talking about. That's what it means to build with AI as your force multiplier. So now one, we can build across all of our data, and two, we get to use AI as our force multiplier. But frankly, at the end of the day, the most satisfying part of building is when what you made actually works and has an impact. When something you create actually helps solve a real business problem, helps the business move faster, actually make better decisions, and in the end, create more value for all of our customers. Because frankly, the opportunity here isn't just generating more. isn't just generating more insights. It's actually shortening the distance between that idea and having meaningful impact in your organization. Historically, turning these insights into action could have taken weeks or months of coordination across many different teams. But in the agentic enterprise, builders can actually move from identifying that pattern and opportunity to delivering value dramatically faster. And that frankly means more amazing moments when you see an idea that you had come to life and you get to say, wow, I built that thing. That is absolutely incredible. So, now let's circle back to our customer churn project and put it to work. Welcome back, James. Thanks, Carl. So, this is what we're expecting to see. You want to make sure that our build is good. So, we've asked actually Coco to go ahead and do some verifications with a number of sub-agents. We also have a long-running agent that's actually monitoring our support pipeline to prevent churn. And then it does research on each customer on how to actually save that customer in a personalized way, not in a templatized way. And then frankly, the agent actually goes beyond those insights and it provides us with actions that we can take with our customers that are likely to churn. James, why don't you go ahead and walk us through what Coco built? Yeah, it sounds great. So, we have our three iceberg tables that we wanted. So, the one in Snowflake, the one on Google Cloud Storage, as well as the one in Databricks. They're all connected with catalog-linked databases. So again, no copying data over. And then we have the dynamic iceberg table with our churn features. Then the AI features enrichment step with AI completes. Our ML model is ready that predicts the likelihood of churn. Next is our hot path that scores tickets as they come in and triggers a Slack message as if it passes a threshold. Then we have the tools for Coco. So, the Cortex search service, as well as the Cortex Analyst. And then finally, we have the Coco Agent SDK that will do the analysis of the customer and select the retention team with actions to take to save the customer. So, it sounds like we got what we wanted. I want to call something out here, actually. I think it's really cool. We can see that the Revisor actually caught several issues as it was actually building out the plan. Yeah, and having that plan reviewed by a different agent than the one who wrote the plan provides an opportunity for Coco to be critical of what is being built. And that's the whole reason why teams work. So, it sounds like we got our Slack message. Let's go ahead and show everybody the results here. Yeah, let's go and do it. All right, so this is the retention team alerts Slack channel. Every card in this channel was generated by the Coco Agent SDK. And reasons over the customer data, and let's check this first one out. So, it's a P0 that needs to be acted on immediately. Coco has a confidence level of 92% that this particular customer is going to churn. Yeah, and let's check out the customer identity strip as well. So, the churn score went from 71 to 100, pretty bad. And the trigger was the customer comment they made. Yeah, it's not good. So, we have our top three signals here. So, let's go ahead and take a look at those. So, cancellation intent is plus one, which is the strongest possible number we can get here. Service experience is minus one. That must have been where they're just having the worst experience in the world. And then the customer satisfaction score actually recently dropped and collapsed. Yeah, it went down to zero today. So, real bad. So, let's go ahead and take a look at the root causes. So, it looks like there were some quality issues that weren't taken care of as to satisfy the customer. And then it created an action plan to solve these issues. So, this is the critical part. There are action items split into what to do in the 24 hours, three days, and seven days. And each one has a named owner, and the expected impact as well. So, along with these action items, there's a matched offer so we can retain the customer. So, we should have a cohort anomaly section here. Can we scroll down to it and take a look at it? Yeah, here it is. So, we see in this section that the central region has a rising risk of churn, and currently there are 3,600 customers that are high risk, and there's a potential LTV risk of $900 million. These must be the most expensive donuts in the world. It flagged the cohort, and it actually recommended a 30-minute review with the segment lead as soon as possible. Frankly, it's amazing that every piece of this particular application was built with a team of agents and just one prompt. Right, the team itself is what really brings it to life. The review step caught real issues before the code was shipped, an independent agent read through the plan cold and found two consumers were sharing one stream, which would have likely been a soundly break and within the hot path. And then the AI call inside of dynamic tables, the iceberg table, would have been very costly across the board. And when the implementer took the wrong path for AutoML, it go in and caught it. All of this was caught by an agent that didn't write the original code. And then, what we built runs on its own. A customer is monitored, scored, and routed into workflows automatically. And the agent posts a Slack proactively, so humans don't have to remember to do an investigation or pull reports. Yeah, that real-time thing is super powerful, actually. And I love what the retention team actually gets to see. They receive a brief. They see who the customer is, why the customer is at risk in their own words, what we need to do about it, and actually specific action items and each person that owns each one of those steps there. Yep, and that's right. And if no one acts on it within 24 hours, it pings the director. I don't want my tickets getting pinged, so. Today, we looked at this example using a fictional company, Data Donuts. But the solution that James and Ana built is actually very real and production-ready today. James, thank you so much for joining us on stage today and showing us the art of the possible. Thanks, Carl. It's been a blast. So, James is a professional developer, but what he demoed today is actually possible and no longer reserved for just experts. With Cocoa and a number of other tools, we're moving into a world where, frankly, anyone can become a builder. Right now, we're seeing an entirely next and new generation of development platforms emerge. Platforms that are designed to actually help more people go from a startup to a business to a startup. From an idea they have to production and application faster than ever before. Platforms like Vercel, Superblocks, and Retool have fundamentally changed who gets to build. It's no longer just professional developers like James. Now, anyone can actually have an idea, create, iterate, and ship value to their customers. But until now, customers have been missing a critical piece, a direct connection to trusted and governed data. That's where Snowflake comes in. We are thrilled to be partnering with these platforms to meet developers and builders where they are. In the tools and workflows they already know, while connecting them to that secure, governed data foundation. So now teams never have to choose between speed and trust. They actually end up getting both of them together. And frankly, one of these partners has taken the integration to an entirely different level. By integrating Vercel V0 with Snowflake App Runtime, teams can now build application in the ways they always have. And with the click of a single button, actually deploy that application into their Snowflake account. And after the deployment, their application runs inside of their account, inside your security perimeter. Query results stay inside of Snowflake. The governance, identity, and access controls that are on your data are automatically inherited with the application. This is what it looks like when modern application development actually meets governed data together. To share more about this partnership and how it's helping democratize building applications across the entire enterprise, please welcome Guillermo Rauch, the CEO and founder of Vercel. Hi, Guillermo. It's so great to be here. Good to see you, man, thanks. So Vercel has shaped how actually modern teams are building software. But this feels different now. It's not just developers that are building anymore. What changes when, frankly, the entire company can start building things? Yeah, look, for us, AI has changed the economics of software creation altogether. The distance between an idea to production have collapsed. And software is now becoming dramatically easier to create. Organizations are moving from a mindset of software scarcity, like software is rare. How can I get a developer to help me? To software abundance. More people can participate in creating world-class solutions. And I think ultimately, the pace of experimentation is accelerating. AI is democratizing software creation the way I think the internet democratized publishing and access to information. So it sounds like inside Vercel, you've completely embraced that everybody can be a builder now. So what changed inside the company once you actually leaned into this? Yeah, our product development philosophy is to always dog food our solutions first. So internally, what we noticed at Vercel was that teams started building their own internal tools and workflows with the advent of coding agents. Internal agents became part of the day-to-day work. You find them in Slack. You find them where conversation is happening. And as a result, experimentation and iteration velocity sped up dramatically. And customers noticed. They told us, like, how come you're moving so fast? And I think these internal agents are to thank. So if you look at the numbers, Vercel has built over 100 internal apps and agents that support this iteration velocity. And teams, crucially, became less dependent on centralized engineering queues. You no longer have to submit a ticket or bug a developer. You can just build on your own. And building really is becoming embedded into the way that the organization operates. So I think that the transformation is not just technical. It's also organizational. Yeah, I bet the organizational part is actually probably more powerful than just the technical piece. So I think that, you know, as you started to actually embrace the fact that everybody's now a builder, why did partnering with Snowflake make a lot of sense for you and your customers, actually? Yeah, I feel like Snowflake was an unfair advantage. So more builders means more apps connected to business data. So the trusted enterprise data becomes crucial in this AI era. Snowflake already houses critical business context for enterprises and for Vercel. So the partnership lets builders create directly on top of the governed data that already exists. Builders can stay inside modern workflows and tools that they already love. And so you don't need to move around or duplicate data. And there's a simpler path from idea to production application. Faster iteration without sacrificing trust or governance. And I think the combination really unlocks entirely new ways for organizations to build. I mean, the thing I love about this partnership is that it doesn't force anybody into a new workflow, actually. They get to do this where they actually do their work. Why was this so important to Vercel? Yeah, I've always believed that the best tools meet builders where they already are. Developers and builders want flow state. No more platforms to learn. Modern teams want fast, seamless experiences. Enterprise development has historically introduced, I think, way too much friction. So this partnership is about removing that friction, removing a huge amount of operational complexity. Builders can now stay focused on creating, on shipping, instead of stitching systems together. And my litmus test is, if it didn't feel magical, we wouldn't be using it ourselves. The experience has to feel natural. And you know, let's actually just take a look at how it works. Let's just demo it. So you can start building a Snowflake dashboard in V0 by just mentioning Snowflake in your prompt. That's literally all you need. Just write it in, mention Snowflake, the agent figures out the rest. That's awesome. It's pretty great. So in this example, we're building a dashboard in V0 showing AI gateway usage by model. AI gateway is one of our fastest growing products. We want to understand how it works, so we connect to Snowflake. We roll in through the secure OAuth access controls. So keep in mind, V0 is a secure coding platform. Everyone connects with their own Snowflake account and attaches their own specific role to each chat, bringing their entire Snowflake permission model into V0. So the chat shows you what steps it takes to build this dashboard. It's fully introspectable, auditable. Behind the scenes, we're spinning up a Vercel sandbox, initializing the project using Snowflake's native tooling, which you saw Snow CLI be kind of a key protagonist here. So all of this stuff that this coding agent is invoking is using this Snow app CLI under the hood. I mean, look at this. It's pretty fantastic. The agent is automatically listening, exploring all of the data. Keep in mind, this is following the design patterns of Vercel. So V0 has amazing design system integration. So what you're producing is actually highly interactive, beautiful applications with filters. It honors your design system, your brand language. I don't think I've ever seen a data visualization this beautiful before. Oh, it's beautiful, absolutely beautiful. Notice the share model here, which you called out. We have a share model and we have a publishing model and deployment model that is governed within Snowflake. We're deploying directly to Snowflake here. And it's just as seamless as you've ever seen with V0 before. Just one click and you're ready to go. Wow, that is absolutely incredible here. So when you watch this happen live from a prompt to a deploy, govern application in minutes, and by the way, the design system looks absolutely gorgeous. I mean, the app in less than a minute looks incredible. It looks like a real production application. What does this represent to you and the team? Well, look, organizations can now move from idea to execution dramatically faster. More people building inside companies can contribute directly to solving really large problems. So I think this changes how organizations operate. not just how developers work. AI is making software creation immediate and interactive. We just saw it. And trusted enterprise data keeps these experiences grounded in real business context. So to me, this feels like a completely new operating model for enterprises. So for everyone here, and frankly, the tons of people that are online watching this now and later on, what's the best way for them to get started with this? Yeah, so start small. This is how we started. Pick one workflow, one dashboard, one internal process. Hop on to vZero, you can rapidly prototype and generate the application experience. You connect directly to Snowflake, just mention Snowflake, to ground apps in trusted business context and real world data. So this integration makes deployment and governance dramatically simpler. You could stitch this out on your own, but this is just completely streamlined and you don't need a massive transformation process within your enterprise to begin experimenting with this stuff. So the barrier to building has really never been lower and the quality bar higher. So the fastest way to understand this shift is to build something yourself. Go and do it. Just go and connect vZero Snowflake today. Yeah, I mean, it's a one minute to wow experience. Like this is absolutely incredible. And it's great that so many more people in the enterprise can actually build things with this partnership that we have today. So thank you so much for stopping by and walking us through all of this, Guillermo. Thank you for having me. This was amazing. Thank you all. What Guillermo just showed us all is a glimpse into what building looks like in today's agentic enterprise. Teams can build tools that they already know and love and deploy those applications directly into Snowflake where those applications, as Guillermo said, inherit the access controls and the security of the data itself. That means you're never ever gonna have to choose between empowering builders and maintaining enterprise control. You end up getting both. And that matters because AI is dramatically expanding who in your organization can create software, workflows and experiences for you and your customers. More people than ever will actually be able to identify a problem, build solutions and deliver value to the business faster than ever before. That's what build for impact really means. Not just generating insights, but empowering people across your organization to turn those ideas into workflows, applications and outcomes that actually move your business forward. Now, if we take a step back, there's a bigger shift that all of this is pointing towards. You saw what it means to build across, across systems, across tools, across formats, across workflows without needing to move any of it. You also saw what it meant to build with AI as your force multiplier, to be able to go from an idea to production faster than we've ever, ever thought possible. And three, you just saw what it means to build for impact, to actually turn ideas into real outcomes and impact by empowering even more people inside of your organization to go build on top of trusted data. Because here's the thing that's really amazing and it's happening right now. For the longest time, building a transformative solution inside of an organization took a ton of coordination, time and specialized effort, but that's actually completely changed. AI makes it dramatically easier for people in your organization to take an idea, explore it, build on it and turn it into something real and incredible. And it's not because expertise doesn't matter. Those people with those ideas in their head, they're experts in their domain. They understand it more deeply than anybody else, but it's because the barrier of entry has completely disappeared. And today, all it takes is a little bit of ambition, a good idea and the right tools. Like Vivek said earlier, I mentioned, we're entering a new renaissance and everyone in your organization can actually be part of it. And builders like you are the ones that's actually making this all possible. You're the ones that are actually taking data and turning them into decisions. You're turning AI into real outcomes. You're the ones that are powering the agentic enterprise. As Ana showed earlier, magic happens when people come together across teams, across disciplines and build incredible things. And if someone in the Snowflake community has impacted and made a meaningful help, has been meaningfully helpful in your journey through education, innovation, open source contributions, community leadership, you can now recognize and amplify their efforts by nominating them for a Snowflake Community Award now through July 1st. You just need to go to snowflake.com slash community dash awards and submit their name today. And in that spirit, I'd like to take a moment to recognize all of the incredible and amazing community groups we have here at Snowflake. Our Snowflake squad, all of the incredible and varied user groups, all those amazing open source contributors, those creative and crazy Streamlet Connect creators building incredible things, and of course, our Snowflake Data Superheroes. We have a few amazing community members here in the room today, and to all of you here and online, from the bottom of my heart, I wanna thank you for being such incredible and amazing examples of what is possible with a bit of curiosity and teamwork. Snowflake is genuinely grateful for each and every one of you. I wanna invite a few friends on stage because frankly, none of this ever happens in isolation. And with that, I just have one thing left to say, let's friggin' build! Woo! Woo! Woo! Woo! Woo! Woo!
