AI Pulse: September 2026

Rachel Roundy: Hello, everyone. Welcome to the Snowflake AI Pulse. If you have joined us before, welcome back. If this is your first time joining us, welcome. We're excited to have all of you here. This is our ongoing series where we connect Snowflake's latest AI innovations to what they actually unlock for you in production. Really excited to get going. This is our sixth AI Pulse. Wild that we've had that many so far, but we're so happy to have you all here. Before we jump in, of course, to flash up our safe harbor and disclaimers. Shout out to our legal team, always keeping us organized. And quick peek at our agenda today. We have a lot of incredible innovations to share with you. So first, we will talk about AI gateway, then AI functions, then semantic studio, CoCo, and CoWork. Quick flag on the Q and A section for this webinar. We will do Q and A after each section, so you will have dedicated chance to have your question answered live. We, of course, won't have time to answer every single question, but please put them in the tool. We will prioritize anything we can't answer live. We will answer either in the tool or offline. All right. So let's get into it. I am Rachel Roundy. I am on the product marketing team here at Snowflake for AI. And I am joined today by a couple of my product marketing peers and a lot of incredible product leaders that are going to be talking about our latest innovations. Now, if you joined us last month, you heard us make the case that the hard part of enterprise AI has moved again, right? We've said that a lot of organizations have stopped asking whether AI can work, and they've started asking how to scale it, how to actually run it in production. The models are phenomenal. They stopped being the bottleneck a really long time ago. But what's stopping AI from leaving the production or, excuse me, POC or development environment is really everything around the model. Right? The governance, the cost controls, the context of your business data. And we talked about that 88% stat, how 88% of AI pilots never actually reach production. And that's not because your POC failed. It's just that that surrounding machinery wasn't there. So last month, we started putting that machinery into place. We talked about model choice, ML ops, multimodal AI functions, and CoWork and CoCo. And today, we're going deeper into the operational layer, because we have heard that teams that are going to that next level aren't just moving faster, but they're building more intentionally with the right infrastructure that helps get them to that AI that scales. So you've seen this slide and we'll keep repeating this because really this is the backdrop for everything that we're launching today and everything we're doing here at Snowflake in the AI space. And it comes back to this architecture. So the agentic control plane where you have all of your AI models, your data, your applications governed within a single boundary with RBAC, audit, cost control, security. So this is really one plane that's right next to the data that your business already needs, right? That critical context. And what we're shipping today makes this boundary even more capable and even more operational than it's ever been. When we talk about the agentic control plane, really what you all are trying to accomplish is three outcomes, right? And this first one is trust. You want to trust that your AI is working, know that you can see where it's running, what it's doing, that you have visibility, that you have an ability to understand who is using which models, and also understand that you can trust the cost controls. Right? There's a lot within this. This one is foundational. Beyond that, you want to build. And you can do that by grounding your agents in your actual business context. Right? So AI that actually understands your data. And then, of course, you want to scale it. You want to automate and optimize at scale, but of course, within your budget. And we have a lot on all of these topics here today. So you will see consistently with the features we're sharing today, each one of them kind of falls within these buckets of trust, building, and scaling enterprise AI within your organization. Now, this is often everyone's favorite slide, Every AI Pulse. This is showing the status of all of our features that we will be talking about in the deep dives today. So everything that you'll hear in our deep dive sections can either be sandboxed or leveraged today as soon as you log off. We also have a ton of great resources linked to documentation for all of these in the webinar platform. So I encourage you to check those out. And of course, every time we have one of these, there are way too many amazing features that our product team is shipping to cover in this one hundred ninety minute session. So I want to give a quick nod to these beyond the Pulse features that are either coming soon or too much goodness to contain here. The first one is under machine learning. For machine learning teams that want to develop with their IDE of choice but running against Snowflake compute and governed data, we have just launched, or soon in PUPR, remote development in VS Code that brings your workflow to Snowflake. This is a huge one. We will go deeper on this in a future episode. The next is AI Extract VNext. This is the next iteration of AI Extract with much fewer hallucinations and much more high quality outputs. Under AI gateway, we have some really exciting things, and I won't give too much away there because we do have an expert here to talk about that today. The dynamic model routing and tools coming soon. And Snowflake app runtime is a complete solution going from prompts to a deployed web app inside of your Snowflake account. It is really an incredible tool. I would encourage you all to check it out. On the data side, on data engineering, DCM projects turns the configuration of your Snowflake account into a version controlled code and puts a dry run-in front of every deployment. DBT projects, we have some great updates and new capabilities that make it easier for users that are looking to upgrade from DBT core native infusion support. Lots of exciting updates there. And then Openflow data DCP for Snowflake Openflow helps solve the private data access problem. Lastly, Snowpipe Elastic Channels, let organizations stream data directly from thousands of devices, applications, and services into Snowflake tables or supported Apache Iceberg tables without requiring intermediate storage or message buses. So really exciting stuff. Again, lots of resources linked in the platform if you want to learn about these other features beyond the Pulse. And I also want to give a quick shout out. We have an exciting new launch today of our Snowflake Developer Blog. This is a blog that has tools and solutions and ideas for developers of data intensive applications. There is a lot for the data side of the house, the AI side of the house, machine learning, you name it. Our product line has expanded beyond our traditional data science use cases, which we heart all of you and we're grateful. We're going to have lots of goodness on the data side, but also beyond that. So this is live today. Please go check it out. There's going to be a ton of great information shipping from our teams. Okay. So without further ado, I would like to introduce our first guest. He's a very special guest, the cofounder of Natoma who recently joined Snowflake. Please welcome Paresh. Paresh, I'll hand it over to you.

Paresh Bhaya: Thank you, Rachel, and definitely not a special guest. Rachel's a special guest here. But more importantly, you guys are the special guests. Thank you for taking time today to come learn about Snowflake. The partners like yourself is what makes Snowflake what it is today. Talking about the control plane, of course, a key part of the control plane is AI gateway. Right? It is one unified platform with the single mission to help you maximize your AI ROI. We've been talking actually. Rachel kicked us off. She kind of mentioned that agents and AI use cases are moving from pilot into production, and they are kind of scaling into scaling into, like, broadly across enterprises. We really need to maximize that ROI. And that is where the Cortex AI gateway. Right? Really focused on helping you improve your AI economics, but at the same time, making sure you are able to observe every single AI agent and then securely log on with them. The most important aspect of this right now, the transition that we are going through from pilot to production to scaling is the fact that AI is starting to take actions on your data. And to really scale it enterprise wide, you need to go and connect AI to all of your applications. And that is the single core mission of this unified platform. So what is Cortex AI gateway? It is that, as I said, a single control plane over to govern your agents, models, and MCTs. All of it in one plot. You start by doing that model optimization. Now I'll touch upon all of these things in detail, but model optimization becomes like a very key pillar in today's world. Talking about that, you have cost governance. As we are kind of scaling AI across your enterprise, the cost is kinda getting out of hand. How do you ensure and you control that cost aspects of it? No AI is secure. No AI is going to get production level widely deployed unless you have the right security, right access guardrails, the data protection. As AI is accessing all your data, you need to ensure they're getting it. And often our loop, but the most important aspect in this world is that observability. We are running and trying to get all that ROI, making sure AI is making our life better, making sure AI is giving you all the all the benefits. Observability is often overlooked, and that is the key aspect of it. Before I dive into details of each of these things, one thing that I really want to highlight, that Cortex AI gateway is going to be independent and neutral. Yes, of course, it will be tightly integrated with Snowflake, CoCo, Cover, Cortex agents that you guys love. But we do understand every enterprise has agents beyond just Snowflake environment, beyond just Snowflake platform. So this we're really making it independent and neutral where on your left side, you can bring in any AI agent, AI harness. That could be your Claude Code or Cursor or whatever your developers love, Or it could be any homegrown agents that you might be building on your bedrock or a majority I found. Or maybe just just chat GBT enterprise. But on the right side, you also get 100 plus MCP servers with bring your own MCP servers, and we'll touch upon all of those as well. So again, it is truly an independent and neutral platform, unified governance there across all your agents, models, and MCP. So let's double click into this. Right? Let's talk about dynamic model drop down. This is models have improved significantly over the last few years. In fact, the rate at which new models are coming in is is accelerating as well. But some of our AI use cases do not need that frontier model. When you are trying to go and summarize a Slack channel and, like, automate some of these things, you don't need the best and the greatest models out there. You can actually leverage the older model. Or more importantly, through Snowflake, you can even leverage some of the open weight models. Snowflake supports those open. And the whole idea here of dynamic model routing is you can depending on the task, we leverage efficient open weight models, older models. You can classify which models to use. And when there is a need, when do you need to do that, like, deep vulnerability scan? Sure, go use Fable 5.0. Right? So leverage those frontier models as and when needed. And the key key result of this is that token efficiency. In our internal studies, we have seen great, great token efficiency when we are actually doing these dynamic which brings us to cost governance. You guys are already aware of the the cost governance from Foco and CoWork agents. We are extending those same thing to third party agents as well. Essentially, giving you the capability to go and set limits, budgets based on the role, and depending on that, constantly monitor it, Ensure that users never go crazy with their AI usage, and you can kind of get control of that. We've already heard in news multiple news articles where enterprises have burned their entire year worth of budget in a single month. This is the way that you can set control, set limits, and now you can really get that telemetry on as the usage is high. So really bringing that cost governance into control, which brings me to the other key aspect of it, which is all the MCP servers. MCP servers are not new. I mean, about a year old now or a year and a half old, but technically they're not I'm sure all of you guys are using them, but there is this massive sprawl that's happening. Everyone is just connecting to all these different MCP servers, and you just go about it. This is where we're help you come in and control and manage all of those MCP servers. So you can now go connect your plot port or CoWork with any of these MCP servers and really get that additional benefit. But this is where I like to think of it as a three parts. Right? It really we end up hosting these MCP servers, so you don't have and your engineers, end users do not have to host them on your laptop. But it's not just the hosting part. If there are third party servers, MCP servers hosted by third party vendors, we can proxy to them as well. But the most important is bring your own MCP servers. Every enterprise has their own custom tools. You might have built those your own MCP servers. That's the idea here. You can bring those MCP servers that you had built, and you can centrally control and manage them from Cortex AI gateway. What is that control? You provide those security guardrails. Essentially, leveraging your IDP, RBAC to assign who gets access to which MCP server. So you can go and control that. Either sales guys get access to Salesforce, to Clod, and the data analyst will get access to Snowflake through Core, or engineers get access to GitHub through CloudForm. So you can really control who gets access to which MCP server. But we don't stop there. We take a step further in terms of, like, fine grained authorization. Right? So what do I mean by that? You can not only say sales guys get access to Salesforce, but they can do these 10 things. But on the flip end, you can also say, oh, product team, they also get access to Salesforce, but they can only do these three things. So very fine grained tool development, which takes you to the next aspect, which is how do you ensure the identity and security. Right? So on we ensure that there is that on behalf of token, and we store the access tokens. We store the refresh token. So when a user is actually going and doing list that opportunity in Salesforce, they will only see the opportunity they are. Which brings me to providing that whole observability aspect that we were talking about. Right? So you get that complete tool called audit rate right from who's the user, what model was used, what harness was used, what was the tool call that made to that MCP server, and including that policy enforcement enforcement. Right? So if somebody, let's say you decided, hey, the the users cannot go and delete that MCP server, then if the user does try to make that tool call, you do get that visibility into that MCP server that tool call getting blocked. Point being, you get that centralized observability across all your agents and instances. Which brings me to a much broader open source. As I said, we are truly independent and neutral. So if you do have third party agents, if your third party agent supports hotel, which today's day and age, all of them support hotel Gen AI, you can pipe those logs to the gateway. And now you can really get that end to end visibility in agent fixed table. Right? So who's making that tool call? Which model was used? How many tokens were used? Did that result into an MCP call? If there was so, what was the tool call? What were the tool arguments? So you can really drive that entire use case, and then you can, like, attribute which teams you're driving the usage, where the failures are occurring, or if you need to go into an audit trail and figure out after the fact, you get all of this. With that, as Rachel already called out, we are in PUPR. So some of these things are in PUPR, so you should be able to try out the observability aspects, The MCP tools and model routing will be in private will be in private preview today. So work with your teams and let us know how we can help. That I see bunch of questions in q and a. Rachel, what what kind of questions are we getting?

Rachel Roundy: Yes. Thank you so much, Paresh. This is there's so much goodness here. Oops. And I think we're missing the q and a slide. Okay. So we're just going to leave it on this, and we're going to ask some questions. Okay. Okay. So we have so many good ones in here. It is honestly hard for me to prioritize. But this one, a little bit of a tricky one to start us out. But how does it handle multi hop agent chains? If agent a calls agent b, which calls an MCP tool, does the cost slash security policy apply per hop or only at the entry point?

Paresh Bhaya: It it will initially, right now, we are in PUPR, it will be at the entry point only. So it won't cover those hops. Where we are going with this is in our roadmap is going to be a to b. And that's why we are trying to see when you have an orchestrator agent and then agent A calls agent B, we can enable those as well using A2A. But today, it's going to be at the

Rachel Roundy: end point. Amazing. Thank you. Okay. Next question. Could you speak to the security risk of opening my Snowflake account to the third party agents through the MCP or within CoCo or any other way?

Paresh Bhaya: No. Good great point. Right? I mean, because we are you trusted Snowflake as the data governance platform, you are doing two things here. Right? One is through Snowflake, you will get these 100 plus MCP servers. So you can leverage Poco over to not only talk to Snowflake data, but now you're actually taking that and connecting to your OneDrive and your Salesforce and your Zoom transcribes or whatever tools you are using in your day to day job. Right? So this is where it really helps you. So talking about security, when you are enabling all of these things, you can start really putting guardrails on who can access those things. And not only who can access it, but also leveraging ensuring what tools are allowed. So that is a good point. On the flip end, you could leverage a third party agent like plot code or or GATGPT for that matter, and connect into Snowflake data. Now, do we support that? Yes. In terms of security, yes. It will be based the MCP servers will be based on schemas, and we leverage PoAuth. And we'll piggyback on the R Pack there. So you will be able to the end users will be only be able to see what they are supposed to see. And the data governance that we have, we will respect that and we'll be allowed. And it's completely based on the user's auth and their auth tool. So they will have to come and do a three legged auth, give a consent. We will make sure if the user is allowed to do it and do it. Having said that, the latter use cases, we've seen way, way, way better efficiency. And we've shared this in AI Pulse before as well, where when leveraging POCO and CoWork directly on Snowflake data rather than trying to use like a third party harness, there is much greater token efficiency and much better outcome because we have that semantic view. Great question.

Rachel Roundy: Yeah, and great answer. Thank you so much. And I think you answered this one a little bit, Parish. But just to double down, can you use this to manage agents built outside of the Snowflake ecosystem? And is there any visibility into that?

Paresh Bhaya: Yes. 100%. We it is truly independent and neutral, so you can bring in your own harness. So you have, like, Copilot agents in Microsoft or you build you are using PlotCore, the developer is using, you can point to the gateway. That's when you'll get all the MCP connectivity, then all the guardrails. But even for observability, as long as they are hotel clients and they support hotel Gen AI, like Claude Code, for example, does support that, you can point those logs to Snowflake, and that's when you can get, like, for every request, what was the token usage and where the token is going and all.

Rachel Roundy: Amazing. All right, last one. This is a meaty one, but important for our audience. Does our data leave Snowflake when we call on these models?

Paresh Bhaya: Absolutely not. The data never leaves Snowflake at that point, right? So you are really if you're pulling in the data and you're actually getting into the third party harness, that's a different thing. But that's the reason why we provide cortex inference. And we provide that. So the data stays there, your AI, the models are hosted by Snowflake, and then data stays. If you're leveraging MCP and you're pulling that data, at that point, the data will leave Snowflake. So you have to be careful on where and what you're trying to achieve.

Rachel Roundy: Awesome. Okay, I said that was the last one, but we have so many good questions. One last one, actually, for real. Does model routing happen automatically, or is there any configuration required?

Paresh Bhaya: Great point. Today, as of today, when you do UPR, you have to configure. And it's very basic configuration that you want. Essentially, you can select which will be your executor model and what frontier models want to choose. Right? So you can go and do that. Where we are working towards is really giving you the auto model route, essentially giving you the capability based on the task, automatically figure out which models to use for that task. So that is something in our roadmap where we are continuously enhancing that aspect as well. But great questions. I know, Rachel, we are on time. I want to answer so many more questions. This is such an exciting topic. As you can see, I'm really excited about it. Let's do this. Reach out to us. Reach out to your account teams. I'd love to set up time to show you a demo and enable this for your tenants.

Rachel Roundy: Thanks for offering that, Paresh. And thanks for joining us. And everyone, so many great questions. We will continue to answer them in the tool. So please keep asking them. We can also follow-up offline. And Paresh, thank you for joining us from the airport while traveling. We appreciate you so much. And we'll pass over to our next guest. Thanks, Paresh. Thank you. Okay. Next, I would love to introduce you all to Hannah on our AI functions team. Hannah has an amazing update to share with you all. Over to you, Hannah.

Hanna Mykula: Thanks, Rachel. And hello again. Probably some of you might seen me from the previous CI calls, but I'm happy to be here back with the September updates. Before we jump into the features that we actually recently released, I want to quickly recap one that we released actually some time ago, but I believe it didn't get enough attention yet. And it's analytical search in cortex agent that's powered by functions. And first, wanna clarify what is actually the differentiating point to, let's say, cortex search for the analytical search. And you probably all know REG. So REG typically finds the best answer within a few passages, and then let's LM read it and generate the response. And this approach works well for the questions where the answer can be located in one place. For example, something like what is my company policy on on this topic? But it breaks when the question is analytical. Let's say how many escalations do we have last quarter based on the tickets report, where the answer cannot be found within one passage and you actually need to search through the whole document corpora and then do some aggregations and calculations on top. And same goes, you know, for for trend analysis. So actually anything that starts with a question, how many? And if you ask that today, CortexSearch will just return the the top k documents and then gives you a number based on this 10 extracted passages. It might look complete, but actually be not because it didn't actually search through the whole corpora of the documents. Analytical search notices when a question is analytical and works slightly different. It does the wide search and decides on its own how deep in the corpora it should go instead of stopping at the top k most similar passages to be retrieved. And after that, everything it finds goes in the table, and then we have an agent running that applies various AI functions like AI filter, AI extract, AI aggregate, or SQL to actually analyze the data. This means that the answer is computed over everything that was found in the Cortex search index and not just from the first top most similar passages. I will later show on the demo how that works. However and this is by the way now in public preview and or how we say it BPR. Analytical search, however, only works if the documents are already in Cortex search in the Cortex search index. So someone has actually to take care of, you know, building the whole pipeline, before that. And that's what I wanna cover next. Over the past month, we actually shipped two Koko skills that take the pipeline building of your shoulders. And the first one, maybe some of you already know, that's called the Function Studio because it was covered also in the earlier AI pulse. Studio helps you get one AI task right if you have, for example, data in the table and you want AI to classify it, extract it, analyze it. And the hard part in this case is not actually, you know, writing the function call. It's knowing if the answer is good enough, if your prompt is optimal, if the model you chose is optimized for quality and the budget that you have. So what Studio does, it actually optimizes the prompt for you and then finds the right model for the task. It tunes both together for the quality and for the cost that you set initially. And the second one is something that we released also not too long ago. It's now a GA for all of you. It's called AI pipeline builder. So that's what you see on the right side. While Studio gets one task right for you, pipeline builder turns that task into a system that runs by itself continuously. So you just need to point it to the stage where your docs live or to the schema where your tables and use live and describe the outcome. And then it composes the AI functions into a ready to deploy to prod pipeline. So as soon as the new file lands on the stage, a plan pipeline picks it up, parses it, does all the extraction and document reprocessing, indexes it, and puts it into Cortex. And what makes it different is that you really need to describe the outcome and not the pipeline. You don't say anymore to to the pipeline builder, hey. Use AI classify and then AI extract and then write this to the table. You say what you want to achieve at the end. And then pipeline builder decides for you which functions to use, in what order, and maybe what functions to skip. And we know that there are a lot of AI functions right now, and finding the right one is actually where the real work sits. So we made pipeline builder an expert on our own toolbox. And, of course, it works on structured and unstructured data together because that's where we believe the real value shows up. And it builds something that's collaborating from the very beginning. So to run the AI functions pipeline builder, you can either call the skill directly in Kalko, or you can just describe in the natural language that you want a pipeline built from scratch, and it will load the skill automatically for you. Then we'll ask you a question. Hey. What are you building? We'll offer you four templates so you can actually see the templates that we're offering as structured extraction, research and analytics, enterprise search, or customer 360. If none of these templates fit your use case, you can also build a custom pipeline from scratch, and then the pipeline builder will compose it block by block for you. After that, it will ask you where your data resides and start building. So let me actually show you how that works. Okay. Hopefully, you see my screen. This is my CoCo. I prefer to use desktop app. So now I'm gonna just put a prompt. Let's say I'm not familiar at all with with the functions, and I just have a bunch of documents and also views for the customer 360 use case. So I'm just gonna ask CoCo to build the pipeline for me for my customer 360 use case, and then it will do a bunch of a bunch of steps to actually view the data after it has asked me, of course, to provide the location of it. So you see now it's, for example, loading the skill. And because usually it takes a little bit of time, I, of course, already prepared it so that we focus only on the most important part. So you see in this case, it prompted me to answer a bunch of questions regarding where the data lives. And I pointed it to my schema, pointed it to the stage where I actually have a bunch of documents with the support tickets. I didn't explicitly name the tables that I've already loaded. I just asked it to check the schema, and I wanna focus on risk scoring, routing, search and rep, and project health landscape. Then it created a plan for me. So this is also cool because I think it it integrates nicely the human in the loop verification because it transparently gives me back what it wants to build, which functions will be used in which dynamic tables. And also, what's very important is the cost section. AR points you to the current AR functions cost so that you can actually verify how many calls and kinda estimate your cost in advance. In this case, it will be using parse doc. And for those who know, AI parse document comes into both in OCR, which is great for simple documents and layout if you have, for example, a figure image rich documents. And in this case, it proposes me to use OCR natively because those are actually simple text documents. And layout is more complex, but also comes with more costs. I answered all the questions. I approved it. And then it just went ahead and built everything for me. And at the end, you can actually see the results of the outlook outcome of the pipeline. So here are the highlighted at risk revenue by product. For example, transform layer is currently at risk. And then here as by tier, 20 of 50 accounts route to escalate, and I have $18,000,000 at risk. And then it also recaps what was exactly built. And what I found cool is that it will ask me at the end, hey. Do you want to actually set up a Cortex automation, which I want? So now I have a Cortex automation running every Monday that produces a customer 360 digest on the weekly basis and puts it in my workspace. So this is how it looks like. This is what I run today. You can actually see all the reports in terms of product health, product briefings that I need to know, summary of the portfolio, and so on. And I also promise you that I will show you how the analytical search works. So it's actually a simple toggle in the Cortex agent configuration. So if you go in configs and tools, you'll see here the analytical search, and you just need to point it to the correct Cortex search over here. And then I asked in the preview to count the accounts that show churn language and give the combined ARR. And just to quickly recap what Cortex agent replied that it used, first of all, analytical search to answer the question and not just simple rack and found 12 accounts and also calculated the combined revenue that is in danger because of the churn. Okay. And we also saw before that cost control is important. Right? AI spend has usually different shape as the warehouse spend. Right? Because a warehouse costs money while at once, while on the app functions, we get the cost per call. That's why we shift, first of all, the user quotas per user, where you can set a credit quota and then the individual user either on a daily or monthly basis. And when the quotas hit, the execution will actually stop. And this is the important part because the user quotas do not just alert you. They actually stop the spend within fifteen or twenty minutes of quota being exceeded. And, also, the second thing that we have is the budgets. Quotas are for people, and budgets are meant for the team. We want you to check the work, set a limit, and when it's crossed, it will trigger your own start procedure. And it's the combination of these two that actually gives you a proactive and very granular cost control. Quattice put a hard ceiling on one person, while budgets tell you where your spend is actually going and let you act upon One balance is kind of the worst case, while the other gives you a bigger picture of where your account spend is going. So we highly encourage you, even if you're experimenting, even if you're playing with the functions builder or your function studio, any other functions to use both. Because this is something that will actually give you a very pleasant product experience and avoid unpleasant surprises, like runaway costs or costs that you didn't account for in your budget allocation. I think that will be it from my side, and I will be happy to take any of your questions.

Rachel Roundy: So much, Hannah. Thanks for hopping on and sharing all of that. We do have some great questions. So let's see. First one, what concerns should we have about the AI functions pipelines not being deterministic for production use?

Hanna Mykula: So nondeterminism is something that we deal when we work with LLMs. Right? This is some one of the risks that we we have to accept when dealing with LLMs and work around it. That's why we also developed AI function studio, and we want to introduce more evaluations and optimization guidelines so that you you can actually practically build evals in your productive workflows. Right? So that if you change a model, you can actually monitor the quality. If there is any deviation from, you know, what what used to run-in prod, we highly encourage you to already start building the evo and have golden datasets that you validate your production pipelines on. I think this is really the the only way to work a lens with nondeterminism.

Rachel Roundy: Awesome. Great. Another good question. Is there is there a clear criteria on when to use analytical analytical search search in and an an agent agent based based on on Cortex Cortex search? Search?

Hanna Mykula: Analytical search is great for the questions that are analytical. Right? So how many, how much when you really need to find the aggregation of of something across the corpus of document. While simple cortex search or rec is gonna be sufficient when the question is more like what or or where, for example, because really then it's sufficient to rely on the similarity scores of the most top 10, let's say, similar passages.

Rachel Roundy: Awesome. Okay. Great. Can I use AI function pipeline builder as an API call? Can I use the CoCo agent from another app?

Hanna Mykula: That's actually a good question. If you can call CocoAgent with a skill, that that should work. But the AI functions to functions pipeline builder is not exposed as an API. It's just a skill for now.

Rachel Roundy: Makes sense. Awesome.

Hanna Mykula: But I would love to know what your use case is, though.

Rachel Roundy: Yeah. And the person who asked that question, if you wanna add more context, that'd be awesome. Okay. Let's see. This is this is probably basic but really important. Do I need to know Python or SQL to use AI functions, or is this something that a business analyst or an ops team could use on their own?

Hanna Mykula: I think the beauty of the age we live in is that no one needs to know Python or or SQL anymore to to get started. Right? I'm not saying that production should should go without without knowing a single line of code. But I believe if your if your business analysts feel comfortable with CoCo, they can really just call functions pipeline builder and and get acquainted with the functions.

Rachel Roundy: Awesome. Okay. This last question is related to the the cost controls. So you had talked about kind of two levels of cost controls. Could you expand a little bit on when you recommend using each one?

Hanna Mykula: So the budgets are meant to to see actually the overview of your of your team spend. Right? Your account spend, and they notify you. Right? So there is no proactive block if you set up the budgets. You rather get the alert, and then you can tie up your your custom procedure to do various actions, after the budget was triggered. While these user quotas are on per user basis, so you as an admin can set for every one of your team members how much they can spend on AI functions. And once they've reached the spent, they're also pretty low latency. Right? So you don't have to wait an hour for it to to take place to for the block to take place. It will take place within fifteen to twenty minutes once the quota is exceeded, and they will actually be blocked from from using AI functions.

Rachel Roundy: Awesome. Amazing. Well, Hannah it up. Yeah. Absolutely. Thank you so much. As always, a joy to have you here, and thank for you sharing all the latest updates on AI functions. We appreciate you. All right. And next, we're going to hand it over to Abhinav to talk to us about Symantec Studio. Abhinav, over to you.

Abhinav Vadrevu: Thank you, Rachel. Hey, everyone. I'm Abhinav. I'm a product manager here at Snowflake, and I work on a lot of things relating to agents and agent context and how you build agents. And most recently, we've been working on Symantec Studio, which I'm really excited to share. It is already public preview and GA very soon. So with that, today, I'd like to start with a quote from a customer that we spend a lot of time with. Jeff Miller is a managing director at BlackRock, and here's what he said. In the financial industry, trusted data and consistent business context are critical to delivering accurate insights and managing risk across global markets. So what he's actually saying is that in order to make the transition from traditional analytics, like BI tools and static dashboards and and so on, to the AI powered analytics, you really need to trust the context that is given to the AI system. And at Snowflake, essentially, what this translates to is having really high quality grounded, semantic views that describe your data really well. And that means all the instructions, the verified queries, the metrics, the dimensions, all of that should be really well encoded in your semantic view. But the the problem is, like, promise and reality may often have some mismatch. You might have seen this in the past. And, really, you want your AI agents to reason over all your data well, and it should align, with all of the answers that you get in your your BI tools. And so excuse the slightly interesting text over here on the side, but, I I think you get the idea with this picture is that, you know, the promise is that different users in different places are are asking different, you know, the same question from different, sources. And ideally, you should get the same answers, but, you know, in in reality, if you have different definitions in different places, that may not always be the case, and I'm sure people have seen, situations like this. But when it does work, you know, you do get the same answer regardless of which tool you use, and, semantic views can help power that. In the past few months, we shipped semantic view and semantic view autopilot, to make that easier to do. But now that we're getting a lot more customers adoption and a lot more scale, we're seeing some new interesting problems. And so we've learned that, scaling this context is actually, really hard. And so broadly, there's sort of four different types of problems that we've heard, from customers. The first one is this whole zero to one thing that we are, continuing to solve with semantic view autopilot. So the question is, like, where should I even start my Semantic View? How do I start if I have hundreds of tables and maybe some, like, ambiguous join paths and it's really hard for me to, like, encode all of this stuff into, you know, YAML or SQL, and maybe there's no existing documentation? Can I leverage, query history and so on? The other problem is with dev life cycles. You have to solve versioning. You have to solve testing, CICD, all of these things. And, so far, we haven't had a lot of great tools to do this. All of this is changing now with Semantic Studio. The other similar related problem is Drift. So what do you do when schemas change and definitions go stale? And then lastly, it's just one of scale, which is like, maybe you do this and it all works fine for one semantic view. But how do you scale this out with 50 semantic views? And what if they have shared definitions, and how can you make all of that process, work really well and really efficiently? So, starting with that zero to one problem that we have done a lot of work on already with Semantic View Autopilot, just as a reminder, you know, there is this iteration loop that we have in the product, within Semantic Studio, where, you know, as you use Snowflake and you generate query history and you use Cortex agents and generate, questions, or users generate questions, we can actually take a look and learn and see, oh, these are the common patterns that exist in Snowflake. So once you get that, there is a suggestions panel, in Semantic Studio that, you can then use to, take a look at those, you know, common patterns and then maybe tune some of the definitions or instructions. So there can be metric suggestions, there can be verified query suggestions, and so on, and you can review that, refine that, and tune the semantic view. And then when you have those verified queries in, those good examples in, we can actually learn from those examples to further improve the semantic views that it's not just leaning on examples and that, the semantic information is actually encoded in the semantic view. And so that is, there's a button in there to to actually learn from those verified queries. So that's, semantic view autopilot in a nutshell. But, now we have semantic studio, which is the surface where where all of this can can happen. Essentially, it's a purpose built IDE, and you can test, you can deploy, Semantic Views and so on. A lot of what I've mentioned until now with Semantic View Autopilot already existed in the legacy editor. There's primarily three new things that we have today in Semantic Studio. The most important one is that we have CoCo built in, and so the integration with CoCo is really, very sort of sleek and works well, today. The second one is that, it's a visual editor in workspaces, which is really nice because, it works right where you're already writing SQL and doing a bunch of other, you know, your data work and so on. And then lastly, very common ask is, can we get versioning for semantic views? And now that it's in WorkSpaces and maintained as a YAML file, you can maintain it as a YAML file in a git stage in WorkSpaces because it has, git and GitHub connection built in. And so you can leverage Git and work WorkSpaces for versioning your files. So we're really excited to have gotten that out into, public preview and, also in GA soon. I have a few short GIFs over here, but, honestly, we also have a full demo video. I'll just quickly run through this. But this is sort of how you can use CoCo for iteration, and you can see there's a diff, and you can accept it and deploy the actual semantic view. This is Git for version control. You can see the panel in workspaces for, you know, making a commit, and then this is now going into GitHub and actually taking a look at the deploy and so on. And then similar thing over here, this is pushing commit. But, essentially, the CICD can run through GitHub in this case. But with that, let me maybe actually jump right into the demo video so that you can get a sense of how to use this. I don't seem to be able to play that. I don't know. Rachel, do you know if I could? Yeah. I Today, I'd like to do a quick walk through of Semantic Studio, which just reached public preview. So Semantic Studio is a new editing surface for creating or editing semantic views in workspaces. You can find workspaces by opening the side panel and going to projects and then workspaces. And you can also open click on any existing semantic view, and it will now open here in deep by default. So, essentially, a semantic view is now just a YAML file in a workspace. And if you wanna create one, you can click on add new over here or the plus button over here. And what you'll notice is there's a few new options. One of them says agent, and then there's a new one for semantic view. If I were to click on semantic view here, it will create a new file for me, and I just name it. And when I'm done naming it, it will give me a few options for how I can get started. The old legacy option of the guided wizard remains, so you can always go back to it and go through the wizard. But we are now recommending that, semantic views are primarily created with CoCo, which is the most flexible and capable of doing exactly what it is that you want. If you happen to have a YAML already and you just wanna start blank, you can always click on start blank. And that, of course, gives you a completely empty semantic view that you can then easily switch between the YAML and the visual editor. And you can see here there's nothing there yet. So for this demo, I'll just quickly look at this existing semantic view over here called growth analytics. And over here, you can see that there's an existing visual editor along with all the stuff that you're used to and that you expect. There's the tables along with all the dimensions and facts and so on. There are verified queries, relationships, metrics, and everything else. It's super easy to just use the visual editor to edit anything that you want. And also, we have our beloved suggestions panel, which has suggestions that come from query history and usage data. And you can see that some of these are still loading, but over here with this verified query, it tells you that similar questions to this were asked previously. It looks like a few more just loaded, so you have some metrics in in this one as well also coming from query history. We've also brought in a few of our, previous panels. So there's a playground here to test the semantic view with Cortex analyst, and, of course, a monitoring panel. And you'll notice that this one, loads pretty quickly. And so if I open up an existing run a previous run, you can see some of the SQL that was produced and the request ID and so on. And so this is this is a pretty performant monitoring panel now. And then we similarly also have an evals panel where you can see past eval runs with Cortex analyst and also dive into the specific answers that were produced during the eval. And so all of this exists, and then the primary new thing that we have now is a really neat integration with Cortex Code. And so what I'll do is I'll open up maybe a metric that like this one, and I'll just ask Cortex Code, rename the metric. And you can really ask for anything you want. I'll just say average MRR amount. How about that? And Cortex Code has the ability to directly, manipulate the YAML in this file, and you'll be able to see the diff right away. The other great new thing that we have in Semantic Studio that didn't exist before is because it's in workspaces, you can have a git backstage, and all of this all of the changes that you make here can be, tracked with git. So great. So we got this change that I just asked for. It came in. You can see that this is a diff, and I could undo it or I could keep it. Let's just say I choose to keep it here. This particular repo just happens to be a git repo, and so I can actually, see the changes and click push, and I'll say, added metric. I'm actually pushing a few other empty YAML files here, which is fine for this demo. But what you can see is that it's actually doing a a git push here, and it will, hopefully, a minute there you go. Push it all to git, and I can actually open the repo page now. And over here, you'll see that there were a few changes that were just made, and this one is it says added metric. One other thing I can do is I can because it's in GitHub, I can see my GitHub actions and, you can actually have CICD running, every time you make a commit. And so in this case, I do a validation and a deploy and you can see that that's still running. Cool. And then, I think the the last thing that I'd like to say is that you don't have to use git. The simplest way to use it, would be even without git. You can just directly deploy, which means that this YAML file can then, become a Snowflake object, and you can actually have dev and prod targets. So it becomes really easy to have the same semantic view and deploy to dev or prod depending on what you're trying to do. We do have a nice little errors panel and warnings panel over here as well now. And the other really nice thing is that we approximate the number of tokens in your semantic view, so that you know, how many tokens it's gonna use up when you use it with Cortex agents. We're really excited for everyone to use this. Please give it a spin. It's now in public preview going GA very, very soon. Thank you. Awesome. Thanks for watching. Yeah. And with that, Rachel, do we have any q and a?

Rachel Roundy: We have so many questions for you, Abhinav. And thank you so much for sharing all of that and the amazing demo. The first one, how is this different from creating semantic views using Cortex Analyst?

Abhinav Vadrevu: Yeah. Great question. Largely, what we've done is we've moved the editor into Workspaces. So a lot of the functionality that you're used to will continue to exist. But the benefit of Workspaces, you get three new things. One is you get the CoCo integration in WorkSpaces so that, you ask CoCo for things, and you get a nice diff, rendered in your WorkSpaces. That's one new thing. The second new thing is WorkSpaces has Git, already built in, so you get versioning. And then the third big new thing, is all of the CICD stuff that you get with the GitHub integration, in Workspaces. But besides that, a lot of the previous functionality, we just ported it over so you get the same stuff.

Rachel Roundy: Awesome. Thank you. Okay. This is kind of a long one, but a good question. Does the shared database and schema location of an existing semantic view matter when editing or saving the YAML to our GitHub repo? We're currently saving the semantic view YAML manually to our GitHub repo, but it would be great if we could do it all within Snowflake Workspaces.

Abhinav Vadrevu: Yes. Great question. So basically, the answer is you have two separate things. You have the sort of database schema, like object location in Snowflake, and then you have the file location in your Git repo. And those two things can be totally separate and unrelated. There is a file that we generate called Cortex project dot YAML. It's basically like a mapping file that maps, automatically maps sort of your, YAML file in your Git workspace to the Snowflake object in Snowflake. And you can go in and change it and move things around to wherever you want. And so you can basically put things where you like.

Rachel Roundy: Awesome. Okay. And sort of related to this, is there a way to import your Git repository into your workspaces?

Abhinav Vadrevu: Good question. I believe the answer is yes. There's a workspace function that allows you to just open an existing Git repo as long as you have the connection set up.

Rachel Roundy: Awesome. Okay. Oh, there's so many coming in that it keeps refreshing my view. Okay. Is Symantec Studio available through CoCo desktop?

Abhinav Vadrevu: Good question. Not yet. The the UI pieces are not yet available. One thing that we do have, though, is there is a skill that comes that accompanies Semantic Studio. It's actually called Agent Studio because it handles both agents and semantic views. The Agent Studio skill is available in all surfaces. That includes Snowsight CoCo, it includes CoCo Desktop, and it includes CLI. So you can use that skill to manipulate semantic views and agents in both locations Awesome. Or all three locations, I should say.

Rachel Roundy: Great. Okay, another long one but good one. Semantic views are excellent. I'm stumped on how to make editing a smoother process because CICD owns all objects in our pipeline. Any suggestions on how to use monitor when my role cannot own the object?

Abhinav Vadrevu: Your role cannot own the actual, semantic view objects, I guess. Yeah. Yeah. Good question. I think we should probably take this offline and look at some other specific options. I I need to review the exact roles which Monitor allows you to have there. But one possible thing you could do is, have two separate locations for your dev and prod semantic views. And so a lot of the use case for monitor is, like, I wanna know what it does when I ask a certain question. So with Semantic Studio, that actually makes it super easy. So you have two different targets, like my dev target and my prod target. What you do is you make a change, you deploy the dev, and then you can immediately, with your own role, see what the traces are and see exactly how it behaves. And then just run your regular CICD workflow with your evals and so on and deploy to prod. And then maybe in prod, it might require a different role potentially to see everything.

Rachel Roundy: Okay. Awesome. Thank you. Does Semantic Studio work seamlessly with DBT as well?

Abhinav Vadrevu: There are a couple of, paths to get DBT based, semantic models into Snowflake and and depending on what you wanna do. So one of them is, we can actually ingest, manifest dot JSON with CoCo. CoCo's pretty good at parsing any kind of files, so you just stick it into CoCo, and it can create a semantic view for you. The other thing is to materialize a DVT semantic view in Snowflake. That you don't get the Semantic Studio UI because the format in DVT is very different to what we maintain. The UI component will work only if you maintain a Snowflake semantic view YAML. And so if you wanna use the UI pieces in workspaces, it would require that YAML to be set up. But we're continuously working to improve the integration there. So I would say stay tuned as we, continue to develop and and make more things possible.

Rachel Roundy: Awesome. Okay. I I love this next question. It's a little bit, existential, but how does Snowflake think about ontology in relation to semantics?

Abhinav Vadrevu: Yeah. Great question. Ontology is a little bit early today, and so we are in private preview and still defining roughly, like, how the whole thing, works well together. I will say that, we should defer this till we get ontology a little further along. And not that, you know, we we have some ideas on how ontology can help, but I think that's still being shaped and proven in our private preview. And so if you're interested in, like, learning more about how we're thinking about ontology, I'd say it's a great time to ask to enroll in our ontology private preview. But I don't wanna share any broad guidance until we get a little further along with ontology.

Rachel Roundy: Yeah. But great great answer, Abhinav. And, yeah, more to come. Watch this space, everyone. Okay. Last one for you, Abhinav. How does the iteration loop when different business teams disagree on what a metric means? Is there a way to flag or resolve conflicts?

Abhinav Vadrevu: Great question. I think, one of the best ways to do this is just to ask CoCo, to well, I guess the question depends on how things are set up currently. So assuming you have different semantic views from different teams, you can ask CoCo to help resolve those conflicts and say, hey. Like, what are the semantic views that exist that have different metrics, for similar terms? And then discover those and then, sort of reconcile. We do have a few other new products in the works in private preview that will help, sort of improve, the whole, like, reconciliation problem where you have different teams with different, definitions. And then from a maintenance perspective, what I would do is generally, treat those different use cases as separate semantic views, and make sure that the, descriptions of the metrics and the descriptions of the semantic views are very clear on the reasons why they're different. Because an agent, that has access to all these things needs to know exactly when to use which one, and so that's really critical. But in terms of discovery and so on, we have a lot more in the works, and CoCo is actually a really great way to get started today.

Rachel Roundy: Awesome. Well, Abhinav, thank you so much. This is an incredible update. We're so glad to have you here today and appreciate the time. Thank you, Abhinav.

Abhinav Vadrevu: Thanks, Rachel.

Rachel Roundy: Alright. And next, I would like pass it over to my colleague, Sid, who's going to talk to us about Snowflake CoCo. Sid, over to you.

Siddharth Dwivedi: Awesome. Thank you so much, Rachel. Hello, everyone. My name is Siddharth Dwivedi. I'm one of the product marketers here at Snowflake, and my focus is Snowflake CoCo. So in Snowflake CoCo, we have two exciting announcements about the features that we are bringing for you all. So let's dive into it. First thing first, CoCo Automation. Okay. I was waiting for my slide to load. Okay. So CoCo Automation. So what CoCo Automation essentially allow you to do is run the Cloud Agents platform that you have announced a couple of months ago that you all might already be familiar with on a cron schedule. So what that means is essentially all the features that you are used to using across CoCo and Snowsight and CLI like the Snowflake managed sandbox, your SQL, MCPs, and all the other tools and very soon Git available to you would all be available to be scheduled as a crop. Right? So think about the task that you have to do on a regular basis. It could be checking the health of your pipeline every day, pulling up a report for your stakeholders on regular basis, or just understanding what changed in your databases that you work on over a specific period of time on a regular basis without necessarily logging into CoCo and running the prompt yourself. And essentially, we allow you to do all of that while you're away from your machine. CoCo, whenever the automation is triggered, wakes up, completes the task, and then goes back to sleep. And we have built a very simplified workflow for to allow you to set up these automations. You just define the prompt and the task that you want Koku to do, define what cadence you want to set up for the task to run through some very simple steps that are embedded in the workflow, and you are all set for the automations to run regularly and complete these tasks for you. And while CoCo runs these automations for you, it ensures that all the user level permissions, the RSS, RBAC, the quotas that has been set by you or your organization is being taken care of. And like anything else, Snowflake, it ensures that the security is kept at the very, very foundation of its automation, and it ensures that the credentials are safe and it ensures, as I mentioned before, the RSS and other policies set by your are being taken care of. Right. So the next announcement that I wanna share with you all around are the model tiers that are that we are bringing within CoCo across all all of our surfaces. So you all might already be familiar with the auto model selection. So auto model selection was designed based on our internal benchmarks as a medium level intelligence which was sufficient for even the most complex styles that our customers were sending our way. And it utilized the models that the org had made available for the for the employees and all the practitioners and user had the ability to use auto mode instead of selecting a specific model. However, today, as we announced the public preview of auto intelligent and auto efficient, we have done further optimization on our auto mode. And this is kind of our first step towards the automated model routing that is in works to allow you to select a specific type of auto model selection depending on the task that you're running. So the first one in this list is auto intelligent. And auto intelligent is designed for the most complex tasks that you, your team, or your organization has to complete. So think about debugging a very complex pipeline or debugging through a very complex set of data or running a complex migration. AutoIntelligent is optimized for those tasks where you do need a very high degree intelligence. And it ensures that it uses depending on the rules that we have built for the model selection, some of the best models that your organization has made available to you. On the opposite side, auto efficient is optimized for the cost performance. So auto efficient is a great automatic model selection method for your day to day task, and it ensures that for any task that you want to run-in Gokul, you have the frontier model level quality at fraction of cost. And this is something that we have optimized and evaluated internally quite rigorously. And as you can see on the slide, we saw up to 75% reduction in the cost for similar tasks using frontier models at a fraction of cost without compromising the quality of the task itself. So, that'll be all from my end. I'll pass it over to Effie to talk about cohort updates.

Effie Goenawan: All right. Can you all hear me?

Rachel Roundy: Thought I was talking. Was on mute. Before Effie jumps in, we heard from all of you that you wanted some clarification on what is the difference between CoCo and CoWork. How should I be thinking about these different tools? Some of you probably use them fluently right between your different teams and others. Maybe this is brand new. So we have a very special session here, Effie, as the expert on CoWork. And CoCo is going to walk us through a little bit of that plus some CoWork updates. Then we'll do a combined Q and A for CoCo and CoWork. Over to you, Effie.

Effie Goenawan: All right. Hi, everyone. I'm Effie. I lead product for Snowflake CoWork. I'm really excited to give you an overview of the frequently asked question, what is the difference between CoWork and CoCo. With that said, can anyone see my next slide? I'm assuming the slide is moving.

Rachel Roundy: You're good, Effie.

Effie Goenawan: Okay. So CoWork versus CoCo, what's the difference? The real question is not really which product. It is who in the company can what they can do with with your data. So you can think of CoCo as expansive. It it's able to read and write. It's where the building happens. It's and cohort is essentially CoCo on rails with some guardrails, some governance. It's the same robust agent capability, narrow to a domain, tune for accuracy, deploy for everyone, including your business user, knowledge worker, people who expect higher accuracy, and then nontechnical. They are not the people who are working on the data, like data science or data engineers. And then the spectrum here is we used to say CoCo is for builders and COWERX for consumption for business users. We've seen move off of that. As you've heard, Snowflake CoWork is now a lot more capable with new capabilities like code execution, including, like, generating documents like PDF, PPT, executing scripts. So now in some ways, CoWork is a lot more powerful. It's able to build some collaterals. However, it's it's about the accuracy and the level of governance. So if you want, like, high stake, high governance curated experience, then CoWork is the answer. CoCo is a lot more expansive, like I mentioned, where it's you are you looking for, like, reading and writing as well as exploring different datasets. So concrete pair here is data engineering team versus field team asking for status offline or, like, executive team asking for updates on the business health. And both are going to grow reach, meeting people where they are, whether they're on the laptop, with web, mobile, with iOS apps, or Android apps coming soon. And then MCP and API apps, both will have this robust access to your apps and services. And for cohort specifically, given the potential reach to a lot more users, essentially, have more people in the field teams or business users than your data team. We do have cohort only interface so they don't have access to Snowsight. So think about, again, governed curated experience, that's where cohort is. So and and to wrap up here, one person rarely needs both CoWork and CoCo. The company usually does, though. Like, we've seen customers deploy both CoWork and CoCo. So, really, the interface is the point. And the benefit here is analyst speeds up without everyone publishing to everyone. So then the noise is not added to the CoWork interface. So with that said, gonna go next to what is the latest on Snowflake CoWork. So this is something I'm really excited about. The team has been working really hard to deliver all of your request, feedback, making sure that Snowflake CoWork is meeting your needs to be the one AI workspace for every context, any outcome where you can work get the work done. And, again, this is not just about answering questions or getting insights, but also completing the actions, whether that's creating PowerPoint, PDF, documents, and soon enough, HTML. We know that HTML reports have been popular, and, like, HTML reports allow you to create beautiful, more interactive visuals, and that is coming very soon. And then most recently is we have automation in general made made generally available. This automation allows you to get work done proactively. You don't have to you can get the agents to start working while you're asleep and you're away from the keyboard. So super excited. We've seen tremendous growth on these automation features. If you haven't already, search our docs and check it out so you can schedule any prompt, any task overnight at specific time. For me, I have many, many automations to prep me for my day, tell me about accounts health, things that I should watch out for my product and my business. So it's been really powerful. So that's a brief update on CoWork. And then another frequently questions on CoWork and CoCo is the governance. How do quotas work across both services? If I have one limit, does it impact ability to both? So, essentially, it depends it depends on how you you set it up. We have many ways that you can set up the budgets, which is essentially either, a, you can set up user quota or resources quota, and you can even set it up at the agent level. So you can set up per user the quota that they have or for specific agent, like, whether specific to CoWork or CoCo. And then second is the admin controls. We know that governance is really critical for your deployment, whether that's CoCo or CoWork. Most frequently asked request is my business users should not have access to Snowsight. So for CoWork, you do have the user level or a role level access to control what interface they're allowed to go. So for most customers, business users, allowed interface, you can set to CoWork so they don't have access to Snowsight and potentially risk making simple changes or being to things they're not supposed to. But, ultimately, RBAC remains to be the foundation, and that doesn't change. And the the thing here is we add many a few more controls on top of our back, whether that's agent that on CoWork agent on CoWork governs what data, what tables, semantic views it has access to. And then, also, we have at assess with this restricted restricted session scope for agents to govern what the agents are allowed to access at execution time. So I think that should cover most of the questions. Yeah.

Rachel Roundy: Thank you, Effie. So thank you for covering. These were these are really our our two most frequently asked questions, and we really appreciate you taking some time on these. We've also have some others that have come through the tool. So if you can hang out, Effie, so we can keep your brain, I'd also like to invite Arun and Sid to join us. And we're going to tackle some of the questions that came in from our customers live here. So let's see. What is our first question from the audience? So this one, I think, would be a great question for you, Effie. Is cohort going to look at WorkIQ similar to Copilot? The email piece is a great enhancement.

Effie Goenawan: I need to take a look at what exactly WorkIQ is. Essentially, the automation that we support is email brief, so to your to the user's email. We can follow-up later on what what you meant by WorkIQ and similar to Copilot. I think the email piece that the the email that we have right now is email to the Snowflake the user email that is registered in Snowflake account. But we are looking into right operations so you can either, a, send a Slack message as part of the automation or email or set up triggers of MCP actions. So that will be supported coming soon.

Rachel Roundy: Amazing. Okay. This is a CoCo question. In the auto mode, do we still have visibility into what model is being used and the token usage, etcetera?

Siddharth Dwivedi: Yeah. So currently, the model that is being used, you cannot see it in the UI during your session while interacting with CoCo. However, your admin teams, they they should have access to understand which model is being used at different levels, and they should be able to monitor the usage of a specific model depending on which model was selected during the session by one of the other models.

Rachel Roundy: Great. Okay. Effie, on your slides, it says that data from CoWork can be trusted. Could you please elaborate on that further? This is an audience question. Maybe around the governance of CoWork.

Effie Goenawan: Yeah. So, basically, CoWork allows governed QD experience, meaning, like, the admin creates the agent associated with the data the agent is supposed to access associated with tools like analyst cortex analyst and the Symantec views. And further, CoWork allows the builder to specify the verified queries. So when I ask a question, what is my revenue? And then that revenue happens to come from the Symantec views and the verified queries, then it will show the green check mark. So me as a business user, I can trust. So think about correct more like a govern boundaries, clear boundaries, and also verified queries, semantic views, so then the answer is a lot more trustworthy for business users.

Rachel Roundy: Awesome. Thank you, Effie, for elaborating on that. Okay. This is a CoCo question. I have tasks running in Snowflake that are scheduled. Will there be any advantage of moving that to CoCo Automation?

Siddharth Dwivedi: Yeah. That's a that's a great question. So the current scheduling Snowflake generally uses the task primitive that we always had in Snowflake. So think about CoCo Automation as an advanced version of it. So CoCo Automation, they still use the task primitive under the hood. But at the same time, when you use CoCo Automation, you get the full functionality of Cloud Agents. So as I mentioned during the presentation, you have all the capabilities including the tools, MCPs, SQL that is a part of Cloud Agents or the Snowflake managed sandbox essentially really elevating the amount of things or broadening the scope of things that you can automate for your day to day tasks.

Rachel Roundy: Awesome. Okay. For for team CoWork, is there a way to auto select an agent rather than letting the user pick? We want CoWork to auto select the agent depending on the question and not allow users to select one.

Effie Goenawan: So good news is this is something we've been working on since our announcement at Summit with automatic routing. So this is coming very, very soon. It's currently in private preview. So essentially, the idea here is user just needs to ask a question. We automatically route based on the question to the appropriate agent.

Rachel Roundy: Okay. Great. And follow-up on the on CoWork. Can I customize a CoWork space for a team that would have some out of the box dashboard slash reports to see at a glance with the Cortex agent also available for users to dive deeper?

Effie Goenawan: If I understand the question correctly, essentially, are other application that the user wants to display on cohort. Is that the case? And I would say on cohort, we do have MCP and skills. So with MCP, if the app support, like, MCP, for in for instance, Slack or, like, Google Drive that I use all the time, then you can ask cohort to connect. You can set up as long as the admin allows it. You can ask cohort summarize reports in my Google Drive or, like, Excel spreadsheet or PowerPoint presentation and generate a report, and you can save and share on core guest artifacts. So that is possible, but it requires AAP admins allowing the MCP connection and the MCP connection being available.

Rachel Roundy: Awesome. Okay. Question for all of you. I'll let you all choose who's going to take this one. So we talked about how we're positioning CoCo for developers and data teams and cohort for knowledge workers. If you're a business analyst who writes SQL, which product should you use?

Effie Goenawan: I can take this one too. It really depends, like, what I mentioned. It's it's use it's less about, like, whether you're a builder or you are a consumer. Because, like like we mentioned, Coart is now allowing you to sort of build things, including collaterals, like HTML reports or PDF or PPT. And soon enough, like, maybe Lightweight apps, I think the difference here is based off of the accuracy or the governance, the curation that you would expect. For instance, I could be that business analyst. If I wanna know how is my financial reports, I would go to cohort because it's pretty sensitive. I want the highest security possible. But then if I wanna do things like exploration, data exploration, or experimentation prototyping, I would do CoCo because it's a lot more robust and it allows the agent to go off and explore different data sources. But at the potential expense of maybe the data is not as highly covered or curated by somebody else, domain expert. Awesome.

Rachel Roundy: Thank you so much, Effie. I know we've had a lot of questions around this one, so I wanted to make sure we prioritized it. When will CoWork generate PDFs or Word docs? And does PowerPoint generation support templates? So questions around docs.

Effie Goenawan: They are in public preview today, so check it out. If you have your accounts enabled for preview, they should be available. You just just ask. Actually, this is good feedback because people might not even know the feature is available, so we'll work on feature discovery. In the meantime, we do update our docs and release notes. If you don't know where they are, they're search for, like, Snowflake documentation. So we always publish release notes every time a feature hits public preview, and then PDF, PPT, and document x generation are already in public preview. And we even have some instruction how you can upload template to customize according to your brand appearances.

Rachel Roundy: Great. And we did have a couple of questions about this. So at a high level team, does it cost extra to use coworker CoCo? Can you explain how that works in addition to your stuff?

Arun Agarwal: Yeah, I can take that. So basically, yeah, of course. Hi, everyone. Both CoCo and CoWork uses AI credits, which is slightly different than the warehouse compute credits that you get for Snowflake. They will all be using the standard AI pricing, which is available online. So you can look up the table and get the details there.

Rachel Roundy: Awesome. Thank you very much, Arun Agarwal. Okay. Great. Well, let's see. Looking through the the q and a to look for some other questions. There are a ton. And everyone, we will also answer some of these offline, some of these for previous speakers. Let's see. Another CoCo question to close us out here with automations. What happens if the automation fails mid run? Is there a retry policy alerting, or how do you see logs of what it actually did?

Siddharth Dwivedi: Yeah. That's a great question. So if CoCo automation stops in between in the process, it would actually keep a session alive for you. So whenever you log back in, you should have access to the thinking that CoCo had done up to the point of failure or up to a point where it actually requires from you as the user. And based on the context, you should be able to troubleshoot or ask CoCo how to troubleshoot for the next steps.

Rachel Roundy: Awesome. Okay. And there's one last question. I said that was gonna close us out, but this one's too good. Okay. Can you build agents in both CoCo and CoWork, or would you recommend building agents only in CoCo? Do you wanna take that one, Effie?

Effie Goenawan: Yes. So agent building should be on CoCo. We don't support yet building agents on CoWork, but it doesn't mean you cannot personalize the CoWork experience. You can do that by building user skills. So for me, I have many, many user skills to customize the experience. Things like I have specific ways I want the agent respond to me, do not be too verbose, and also the the writing style as well as, like, meeting prep or meeting follow-up. So you can always create the user skill. So next time you ask, it remembers your workflow. So yeah.

Rachel Roundy: Awesome. Anything you would add, Arun or CoCo? Arun or Sid from? I'm just calling you the product name now, Sid.

Siddharth Dwivedi: No. This is I think if he, like, hit the bull's eye, CoCo is definitely optimized and the harness is built for all the builder related task and all the engineering tasks that are related to data. So, yes, you can build highly optimized agents and high performance agents using CoCo. And, yeah, and the best way to find out is to give it a try and let us know how it goes. Absolutely.

Arun Agarwal: Yeah. There's maybe a quick shout out that we also have Cortex agents and APIs, so you can also use those if you depending on your use cases.

Rachel Roundy: Yes. Thank you, Arun. Awesome. Well, you, Effie, Sid, Arun, for the incredible session on CoCo and CoWork. And thank you all for joining us today for the Snowflake AI Pulse. It was a jam packed session. There are a ton of resources linked below in the webinar platform, so please check those out. Yes, this will be recorded. It will be available on demand. And please contact your Snowflake reps if you have any questions. We'll continue to answer your questions in the tool. Thank you very much, and we'll see you next month.