Atlan + Snowflake: Building AI Agents That Know Your Business

Charlie Freeman

0:00 Welcome, everyone.

0:02 Where are we all joining in from?

1:16 Give it a, a minute or two here, and then we'll get going. Excited to, to chat with you all here today about the great partnership with Atlan and Snowflake. All right, we got some Brazil, India, Dallas. Very nice. Costa Rica jungle. That's a new one. I like it. Thanks for tuning in. Awesome.

2:05 Well, we can go ahead and get started here while some folks still filter in. Like I said, excited to have you all here today. Fantastic webinar. Certainly high in demand of a topic from our pieces our, our customers. How we're gonna be doing this today. So some of the housekeeping items, webinar is being recorded, and we will share it afterwards. Please use the chat for comments and questions anytime. We'll get to many of them during the course of the webinar, and then we'll also have a Q&A at the end for any deep questions that both Abhinav from Snowflake and Razi from Atlan will be tackling together.

2:44 But really excited to, to have you all here today. Please keep the chat active. As far as content, so we have the great Austin Kronz from Atlan talking through kind of the AI agent's failure path without the right context. He's our, our chief field data officer who's commonly meeting with customers to talk through their problems, and he'll be walking through that story for us. Then we have Abhinav, who has given us the first light look at Cortex Sense, the problems it's set to address for Snowflake customers how we're envisioning that to pair up with the context that's curated in Atlan. And then we'll turn it into a live demo.

3:23 So both Razi and Abhinav will take the wheel and, and show you what it looks like in, in practice for what our customers will be looking to, to solve problems with. And then, like I said, we'll, we'll wrap that up with a Q&A for deep questions, but I'll be hovering the chat in the meantime. So please don't, don't hold them till the end. It'll be helpful if, if you get them in as, as soon as the question comes top of mind. But with that we can kick it over to Austin if you want to take it away for us from the Atlan side.

Austin Kronz

3:54 Yeah. Thanks, Charlie. Charlie, can you just give me a thumbs up if you can hear me okay?

Charlie Freeman

3:59 Yep. Looking good.

Austin Kronz

4:00 Sounds good. Awesome. Yeah. I think before we get into, you know, what this actually looks like in practice, I, I just wanna level set on why are we talking about context. And, you know, most of the conversations that I'm having these days with executives who are thinking about building AI agents goes a little something like this. You know, there's a heterogeneous landscape of systems that lives on one side of the house, which is all of their systems of record. Their systems of data, like their data warehouses and their data cloud, like their BI platforms, and then obviously their systems of work, you know, your Confluence, your SharePoint, your Drives, and things like that.

4:34 And all of these agents whether they're analytical agents, operational agents, your productivity agents, coding agents, you name it there's a growing number on the right side as well, and there's sort of this messy middle where every time we go to build one agent, we're baking context into that particular agent, but we need to find a way to capitalize and reuse the context that we've engineered when we're going into that second, third agent.

4:59 And so when we think about what is an AI context platform and just where does context help, it's really addressing that messy middle area and having one place that we can serve context to agents regardless of how they're built, who's building them, and, and what the purpose is. So if we go to the next slide, Charlie. You know, I have a lot of conversations, like I said, where, you know, you talk to one person and you say, "Context," and I've gotten pretty good at it now, but originally, I would pause and you know, about ten minutes into a conversation, I would say, "Well, what, what do you mean when you say context?"

5:35 And sometimes that means semantic layers. Sometimes, you know, we're talking to an architect, and they're thinking it means ontology. And sometimes you talk to someone else and, you know, they're talking about a knowledge graph that they've tried to build. And so before we get into any of the, what I'll call buzzwords or just actual kind of technologies surrounding context, I wanted to take it back a level and just say, fundamentally, what does an agent need to be successful? And we think about context as this knowledge, expertise, and norms. Knowledge is just that map of your business. What do things actually mean and where are they?

6:11 Expertise starts to get into the procedural knowledge where we need to know how work gets done in an organization. And then the norms are just what can we expect that agent to do? What is it allowed to do? In practice, if you go to the next slide let's think of some very simple questions. You know if, if you attended Snowflake Summit, you may have seen that we built a whole fictitious burger franchise called McContext. And so all of the demo examples that, that I have here are related to, you know, a fake burger franchise. But something as simple as why is drive-through time up this week?

6:49 Some things that are interesting about this, you know, when you get past the what questions and you get more into these why questions, context becomes increasingly more important. In a simple question like why is drive-through time up this week, you need to have knowledge of, first and foremost, what does drive-through time actually mean and where is it and how it's calculated. There's even more subtleties i-involved in this. We say this week, and, you know, an analyst at this particular company may intuitively know that, well, of course, this week means Sunday to Monday.

7:18 Whereas you may have come from another organization where Monday to Friday is typically what you mean by this week, because that's what matters to your business. The expertise on this, this is a diagnostic why type of question, and so again, maybe a human analyst would intuitively know, oh, I should check seasonality. I should check whether or not this particular store has two drive-throughs or whether or not we've launched some new promotion. But an agent doesn't necessarily inherently know that about your business. And so the procedural know-how or this expertise needs to be encoded into what the agent has access to. And then norms. You know, who's asking this question? What can they see?

7:54 And what should they see? And how should we respond, which should be the, the overall type of result. Is this a, is this the CEO asking the question? Is this a store manager? Is this a branch manager? Whatever it may be just the overall acceptable actions of that agent are really important. So I always like to ground in this before I get into, well, what, what is challenging about context right now. So go to the next slide, Charlie. Here's, here's what we have seen. We call this the path to contextual intelligence, because if you think of what makes a great AI agent, there's really only three ingredients. You have the model, which is largely commoditized.

8:34 You know, I have access to the same frontier models as everyone else attending this session right now. You have the harness, which is where it's built, which is where we'll talk more about what Snowflake is doing. And then there's the context. The context is what makes the agent unique and what helps curate information about your organization. The models are very intelligent, but you know, the analogy I've used in the past is picture hiring the most brilliant consultant, has every Ivy League background you could possibly think of, but you put them in your organization, and you don't give them a laptop, you don't onboard them, you don't let them talk to colleagues they become relatively ineffective.

9:09 And so we think providing context to these really intelligent models is, is what is going to actually get value in an organization. What's hard about that? We see these three challenges or these three blockers to getting to that value. We talk about context bootstrapping. You can spin up an agent very quickly. Sometimes it becomes a, a labor of love actually providing context over time to a particular agent. And so just finding what do things mean, what are these definitions, what are our key metrics, how do we define this where does that information live? Does it live in a doc? Does it live in our code?

9:46 You know, bootstrapping context can become pretty challenging for your first few agents. When you cross that hurdle, you start to get into, well, context, much like data, much like everything else in the knowledge management space, has a life cycle. So you might be curating semantic models, you might be curating definitions, but your business changes. There's a new use case, and so there's actually a life cycle in and of itself of managing this. And from a governance perspective, what is the right thing to use? You know, we have a ton of data that might be accessible, but... And a lot of it could look similar.

10:21 So how do we appropriately govern and make sure that this is trustworthy context, this is accurate, and this is up to date. And the last thing we talk about, sprawl and drift. So at scale, if you think about this holistically, you know, you could have a lot of duplication that might emerge. You could have context that was accurate previously but isn't accurate anymore. And there's also not just one place you're building agents.

10:45 I think that's the important piece here when we think of this from, from an Atlan perspective and from just an AI context platform perspective, is you'll have agents that you're absolutely building in Snowflake, but maybe your engineers are building in, you know, Google Cloud, or they're building in OpenAI and Codex or some third-party harness, whatever it might be. And so you could have a lot of sprawl across those systems when context is being baked directly into them. So these are the three challenges we see. If you go to the next slide, what do we think of is the overall context platform of how do you create this shared enterprise brain?

Austin Kronz

11:20 First and foremost, we think of this in three buckets. You're going to have your AI-ready data. How can we pull together all those systems I mentioned before, your databases, data warehouses, BI tools, transformation tools, source systems collaboration tools, et cetera, and just stitch that together, build a solid map of lineage that connects the technical assets to even business concepts like glossaries. And so we have this AI-ready foundation that we can then build on. As you build these more analytical agents, semantics become very important. How can we provide semantic models to make user questions a bit more deterministic in their responses? And, you know, when someone asks, "What is revenue?"

11:59 I should be able to ask it ten times the same way or a different way and get the same answer. And the last thing is really I would say the most emerging paradigm in, in building agents is how do we capture some of that procedural knowledge in the form of skills? Agent skills becoming a large like key type of context for, for building these agents. The one thing I'll talk about it a little bit on the next slide. What we're trying to do is mine as much of this automatically as possible. We feel that there is a lot of information that already exists in that what we call data graph.

12:41 So whenever we are stitching together, again, databases to transformation tools to BI tools, data warehouse data quality, all of these signals are actually really rich in telling us what what information or logic you've already baked into these pipelines. And so at Atlan, we're trying to extrapolate as much of this for you using what we call context agents. So you see the first piece there is really just creating that AI-ready data foundation. When we get into number two, this is what we are solving the bootstrapping problem for you.

13:16 We're automatically generating descriptions, readmes, creating glossaries, linking glossaries, even building out entire ontologies of business concepts based on some of the logical concepts in your data as well. We have all the query history. We understand the actual business questions that are being asked of your data, and so we can, you know... I think the, the name of the game these days is text-to-SQL, but I think we've actually been doing a lot of SQL-to-text. So we're actually trying to figure out how that SQL query originated as a business question and what that might mean, and how we can use that to maybe even generate your golden sets of questions or your evals automatically for you.

13:54 Inevitably, i-in stage three, you'll see not everything can be completely bootstrapped. We want to do as much as we possibly can from a bottoms-up standpoint, but there will always be a human in the loop. You'll always be working in your local harnesses if you're using Coco, or if you're using anything in Snowflake, you might be doing some engineering directly in that harness. And so what we want to do is just meet you where you're at and have an ability to still be the context layer, maybe publish and pull context from Atlan, but you don't necessarily have to live inside of Atlan to do this.

14:26 And a big piece of that is just being able to activate context. So we have an MCP. Typically, the way agents are interacting with Atlan context is via MCP, but we're also built fundamentally on an Iceberg native, what we call Context Lakehouse. And so if you want to just do some rich, deep analytical queries across what context you already have, you can run SQL on top of Snowflake, which is hitting our Iceberg Lakehouse behind the scenes as well.

14:52 And the last thing I'll mention here is kind of the holy grail of building effective agents, which is how can we take the traces that your agent is actually emitting, meaning, you know, what is the agent actually doing in production? What is it doing well? Where is it struggling? Where is it providing incorrect answers or inefficient answers? And then use those traces to then feed back into this whole process to say, "Well, maybe we need to update our skills. Maybe we need to update that semantic model. Maybe we need to adjust something in our code or our harness to more effectively deliver answers to users."

15:26 The theory here is you can only bootstrap for so long. I've talked to a lot of customers that want to get everything right before they ship. And I do think you have to have some type of eval threshold that, that exists where, you know, we're going to allow this agent to kind of run in some production environment and actually use traces the rest of the way to improve the agent. So go to the next slide, I think the last slide, Charlie. We can... Let's actually skip through this one for now, just for the sake of time.

15:55 If I were to summarize, you know, how we're seeing customers progress through maybe an AI maturity is they're definitely starting with their AI-ready data. Just how can we, both from humans needed this just as much as agents you know, how do we just find data that's trustworthy in our organization that we can understand and utilize? And then how can we layer on top of that this business meaning? So whether that's adding our semantics, our metrics or like I said, even abstracting entire business ontologies based on some of that structured data.

16:29 Agentic context is things that are specific to skills and tool calling and, you know, we've even talked to customers that are talking, you know, we have human SOPs, but now there's agent AOPs that are becoming prevalent in some of our conversations, just holistically, what does this process look like and how can we provide that to an agent? And lastly, I don't really know any company that's truly at self-improving context. I think they have it in pieces, but the ultimate goal would be how can we just close this loop and have what the agent is actually doing and producing and emitting just automatically feedback and improve so we have the self-improving state. But I'll pause there.

17:09 This is how we are thinking about the space. I'm really excited for this session because we'll actually be able to see what this looks like in practice and how both Atlan and Snowflake are helping to bootstrap and provide that context in a governed and effective way. So Charlie, Razi, team, I'll, I'll turn it back to you.

Charlie Freeman

17:29 Thank you, Austin. Up next we have Abhinav. Abhinav, I can, I can drive for you if you just queue me. And we'll, we'll get started on, on y'all's story.

Abhinav Vadrevu

17:39 Sounds good. Hey, folks. I'm Abhinav. I'm a product manager here at Snowflake. I work on agents and their context and how you build agents and context. So can we go to the next slide, please? And we can probably skip this one. This just says that it's a preview feature which is... It's very early, so maybe I should say that you know, if you're interested, please reach out to your Snowflake rep and we can you know, look at getting you enrolled. So let's, let's talk a little bit about the problem that Cortex Sense solves in particular.

18:11 So if you go to the next slide, basically what we see today is that a lot of customers that we talk to have one of two problems. One sort of flavor of the problem is you have a lot of business users who use various apps like ChatGPT, or maybe it's like Claude Cowork or things like that, and they want to query their data. Now the easiest way to set this up is just to talk to Snowflake MCP. It's actually even the most common way, even though we don't always really recommend that. And what happens is that your agents don't have context about the data they're querying.

18:45 And so if you go to the next slide basically you get these answers that are incorrect but sound sort of correct because maybe the agent decided to sort of look around, try and find some databases or schemas that sound related to the question and then try to answer your question. But, you know, in practice, that, that answer may be wrong. And so at Snowflake, we solve this with semantic views which have sort of worked pretty well actually in terms of, like, at least getting the quality up. But there is a problem right now with semantic views as more and more customers have adopted them.

19:20 If you go to the next slide, you'll see what this is, which is the, the problem that we have at Snowflake when we adopted semantic views is that we have probably around ten thousand tables and because a semantic view is an authored context, like a human-authored context, it's pretty hard to sort of scale this up to all of the tables that exist in Snowflake. And so internally, we have probably less than five percent of the tables actually well-documented with good semantic views, and majority of the data there is actually not exactly super AI-enabled yet.

19:55 So we did find a solution to this, so if we progress to the next slide we have this idea of a more modern context stack. And so in the next slide, basically we have... Imagine, like, all those gaps that existed before with no sort of good context. Those gaps get filled up with context from Cortex Sense. And that's the core idea behind it, is that it fills the gaps with knowledge, and that knowledge comes from a wide variety of sources.

20:26 Today we have a bunch of connectors and we're working on adding more and more, and this is where Atlan comes in where we can sort of work well with with all of the stuff that you have in Atlan, all the context. So if you have, like descriptions and, and semantic information and so on living in Atlan we have plans to build a nice connector between Atlan and Snowflake so that you can get all of that context in. And there may be context living in other places too. Of course, it's in Snowflake already, some of the stuff, but then also it could be in, like, your BI tools and, and maybe there's PDFs and so on.

21:03 And so essentially what Cortex Sense does is it pulls together all of the context that may live in various places to better answer your questions with agents. So if we go to the next slide this is a sort of rough view of essentially how all of this might look. So Cortex Sense is actually part of Cortex Agents or Coco. It's essentially a, a piece of it that just makes it feel smarter. And all it does is it pulls in context from all of the sources that exist that Snowflake knows about.

21:36 So a big part of the breadth that I was talking about that you can, you know, really learn about anything comes from the live connectors. And so one of these would be Atlan, but essentially we can pull in all that information, learn about it, and when you get a question coming in we can find the relevant metrics or the relevant context to answer accurately. And then these other three buckets are things that you might be used to in Snowflake already, such as you know, for structured knowledge, we have semantic views, we have search services, we have also just things in the works right now in private preview like ontology and, and so on.

22:11 So all of this can be a source of context for Cortex Sense. So if we can move on. Basically, when you get a question coming in from an agent you can have the, the sort of governed context that a human wrote that's sort of... That's preferred. So if there's an existing semantic view, an ontology or something like that, that, that a human has written it, it uses that, and then if you have, like, less trusted sources, let's say query history or some Power BI dashboards, it falls back to the slightly less trusted context or the less, I would say, certified context.

22:47 So for the next slide if you could move forward, the, the Cortex Sense has high level three types of knowledge in it. First is what does your data actually mean? Like, what is the semantic information like metrics and, and grain and keys and all this stuff? The next is how does your business connect? And this is very sort of ontological in nature. So how do entities relate to each other and, and what are the relationships and what are the concepts? And then lastly is maybe very similar to skills, I would say. Like what are the steps of things that you have to do to answer a question?

23:25 So basically, with these three types of things, we're able to get really high-quality results. And the way you think about the value of Cortex Sense is one of three things. Like, firstly the most important sort of thing that we think it will drive is the speed with which you can build agents that talk to your data. And so today you have to do a lot of modeling and creating a semantic view before you really can use an agent. That takes time and effort but the idea is with Cortex Sense, because a lot of this is automatically managed, the time it takes to get from zero to one gets a lot faster.

24:01 Let me say that this is not that semantic views will go away. They're still very useful for governed human-curated context. It's just that when you first create an agent, you just want to test if it works and whether the concept makes sense and whether it feels valuable. And so if you want to do that first you shouldn't have to go through weeks and weeks of creating a semantic view or multiple semantic views to describe all your data before you can even test if the concept works.

24:28 And so the way I think about this is Cortex Sense increases your speed to, like, test out an agent and see if the concept has legs, and then after that you can choose to create semantic views if you want. Cool. So for the next slide this sort of describes something that we've done internally with our agents, which is in the past we used to have lots of different semantic views describing all the data, but with Cortex Sense, because Cortex Sense has a lot of that information already we can really reduce the amount of semantic views that we have with an agent, so they're actually smaller and fewer.

25:00 And then we lean on Cortex Sense for a lot of the other context. And what that means is that because Cortex Sense is actually live and it auto-updates and it's managed it sort of stays automatically up to date, and you get pretty good answers with much less maintenance. And this is good for agents where maybe you care less about, like having extremely grounded context for every single case. So, like, not really a good use case for, like, the legal or, or CEO assistant, but maybe a great use case for, like your data science folks who maybe have less, like sort of critically...

25:34 It, it's less critical that you get the perfect answer every time, and more important that it stays updated on its own. And then last thing is I think ad hoc questions get a lot better. So in the past, you used to only be able to answer questions that were described by semantic views, and moving forward we'll be able to now sort of answer any kind of question really that Cortex Sense covers. The other thing if you could go to the next slide, is that because Cortex Sense learns about your entire business and not just about stuff in Snowflake we can create SQL about that refers to data that may not be in Snowflake.

26:10 And as long as we can federate the query, which means that you, you should have a horizon connector to that source of data we can have the inference and the sort of intelligence piece run through Cortex Sense in Snowflake, and then you can just federate queries out. Cool. So let's talk a little bit about setting up Cortex Sense, and I'm gonna run through this piece because we will show a demo shortly. But very, very high level basically there is a process of building, and then you can evaluate Cortex Sense and then ground it based on, like, you can give it further information and sort of overall improve the accuracy.

26:45 On the next slide you'll see that this is roughly... If, if you could progress one slide. This is roughly sort of a visual of how Cortex Sense works. It pulls in information and then it maintains its own internal index, and you can query it with Snowflake agents. So that's Coco, Cowork, and Cortex agents. In the next slide we have a quick view of the eval, which I'll also show very soon in a, in a demo. And you can see that you can set up questions and see how Cortex Sense works, and you'll also see, like, with and without Cortex Sense. Next slide shows that we can actually learn from those failed evals.

27:25 So if you choose, like, we can analyze essentially, like, why did it fail an eval question, is there some missing information, and what can we learn? And then next we also have this concept of conflicts. So this is new. It's different from RAG. Something... Sometimes people think Cortex Sense is like RAG. It's not quite the same thing because it actually understands when you have two different conflicting sources of information and, and how that all works. And then overall you can sort of then manually tune Cortex Sense and give it more information and, and improve it over time. Last slide here is proof points. So if you could... Yep. Thank you.

28:06 So very quickly you know, we have Frontier... We benchmarked Frontier agent against Coco without Cortex Sense and then Coco with Cortex Sense. And you can see firstly accuracy just goes up as you use more grounded context. Coco is better than the coding agent with SQLCP because it's, has the ability to search through any existing semantic views and so on. Cortex Sense has even more context at its fingertips. And then on cost, you'll notice that the coding agent is actually the most expensive because what it does is it runs describe table, describe schema, and wastes a lot of tokens on looking around for context.

28:43 Coco is the cheapest because it has fewer input tokens than Cortex Sense. But in my experience, that per month number, we're gonna probably have to update it because we have noticed that with Cortex Sense, people ask fewer questions because you need fewer follow-ups because you're more likely to get the right answer in the first shot. So that, that will probably, I would guess, go down as we continue to measure. Cool. So with that I will pause, and we can get into demo stuff.

Charlie Freeman

29:13 Amazing. Thank you for that, Abhinav.

Charlie Freeman

29:16 Razi, why don't you take over and make it real for us?

Razi Shafiq

29:20 Yeah. Perfect. So what we're gonna start with for the demo is on the Atlan side of how Atlan collects all of that context, enriches it and then can send it to Snowflake in different formats. So here I am on the Atlan homepage, and we're gonna start here in our connector marketplace. As we've talked about during this webinar, the first step here is going to be collecting all of the context from where it lives. So here in our connector marketplace, some place could be Snowflake. This is where we can have unstructured documents, your ERP systems, CRMs, et cetera. Once we connect into all these systems, Atlan actually creates your enterprise data graph.

30:06 And I'll show a few nodes on that graph of what it looks like. So here on this page, you can see I have Snowflake tables I have SQL Server here but I also have unstructured documents. So I have PDFs, Markdown files, et cetera. We've collected all this information from all these locations to then have it governed within Atlan. And here I'm gonna show one node on that graph. So here's a Snowflake table, dim account. For this table, we have the description information, columns and their description information— along with more robust context through a README. So this can be use cases examples of where this is u-used, how to use it, et cetera.

30:52 And additionally, we then also create end-to-end column-level cross-system lineage. So here you can see from Postgres into Snowflake via Airflow, a few different transformations here in Snowflake, ultimately ending up in Tableau. So this is all automated for you out of the box in Atlan for that context collection. Now, let's say we've started to collect all this context, but there are pieces that are missing, and what has traditionally been done by a governance team and a team of stewards, we've actually evolved into context agents here in Atlan, and I'll show a couple examples of what that looks like.

31:35 So here I am in Atlan's context agent studio, and this is where we can bucket all of the information that we've collected from the different sources into collections. So here's an example of everything tagged with a certain tag. From here, we can run Atlan's context agents. So here's an example of one of our agents. It's our description writing agent.

32:02 What it actually does is not only look at the schema and the naming conventions, it starts to look at that lineage that has been created, query history that might exist in Snowflake as well as your metrics and your business glossary to start to understand what is this table or what is this report and using all that information, we'll write a description or we'll write a README as well as custom agents can then be created on top of our local harness here to provide all of that rich context. Additionally, we can actually generate metrics from everything that we've collected, whether that be metrics from the BI layer, the data layer, the upstream layer, et cetera.

32:46 The last piece I'm going to show here for the context collection as Abhinav mentioned as well, was mapping out all the different relationships in the business. So here you can see I have customer as an entity and all the different relationships that customer has. So customer places an order, an order ships to an address, et cetera. Here we're mapping out all of the concepts in the business and how they relate to each other for additional context. And as we feed that into Snowflake as we feed that into Sense, it will just continue to make it smarter.

33:22 Now, before we head over to Snowflake, after we've done this collection of all of this context, what can we actually do with it? In Atlan, we have what we call our context engineering studio that will actually put together and create defined locations. In this case, it's a domain-specific repository. So for my Memorial Health team, it's the finance domain. They've collected all this information for their use case, and now in Atlan, we create our model here, which can then be ported to Snowflake. But here's an example.

34:02 We have the semantic model with the tables that exist, but we also have links to many different types of information, markdown files, business terms, dashboards, et cetera. With this collection what we can do today is actually port it directly as a semantic view into Snowflake. And as we continue to evolve this, this information can now be sent to Cortex Sense, whether it be through a markdown file, whether it be through an MCP or a specific Horizon connector that we'll build in conjunction with Snowflake. With that, I'm now going to pass it over to Abhinav to show after this context has been sent to Cortex Sense, what it looks like on the Snowflake side.

Abhinav Vadrevu

34:54 Awesome. Thank you, Razi. Let me share my screen, and here I am going to show my terminal. And here I, I'm showing essentially how you can build Cortex Sense using our skill that exists in Coco CLI. So for private preview it's less of a UI and more of a skill, and so I'll walk through exactly how that works. So essentially what I did here is I attached the Cortex Sense skill with /cortexsense, and I said, "Hey, build Cortex Sense on everything in XYZ schemas," and I just gave it some pointers on where the data lives, and I told it to include query history.

35:37 And with the Atlan connector, we would also be able to sort of point it to a Atlan connector. And what it does is it does a bunch of thinking, and then normally it just sort of explains to you exactly how Cortex Sense works. And you'll see, like, this visual is very similar to the one I presented on the slide. It tells you, "Hey, this is how this works." It will learn the context, and then it even confirms the scope with you. It says, "Hey, these are the tables that I'm gonna include in my index. Here is, like, there's some semantic views, some Streamlit apps," other stuff that it found that's relevant. And normally you would...

36:09 You'll also have the option to, like, change this context or the scope. The other thing you can do, by the way, is even after building it, you can always go back and edit the scope. So I said, "Yep, this looks good," and then it goes ahead and builds Cortex Sense. I'm gonna scroll down here. And I've already pre-run this for the interest of time, otherwise we'd be sitting here waiting for LLMs to run. But basically you can see that it named the domain Healthcare Quality Revenue, which is a good domain for... Or name for what I gave it. And it's gone ahead and kicked off a build. So then what?

36:48 I'm actually gonna run this part live 'cause it's a little bit faster. This part is brand new. We have a little terminal UI here that allows me to select a Cortex Sense object, and you can actually see you know, what's in there and, and what's the status. And you can see here that, hey, we have the build is ready, and it used five sources. Maybe, you know, I haven't yet run some evals. There's also this concept of inbox, which I'd previously mentioned. It can understand conflicts, and you can address them right here.

37:18 The scoping, as I mentioned, you can always see the scope and sort of update that if you want, add in n-new semantic views or new tables into scope if you want that. Inbox, currently there's actually nothing 'cause it, it was already, I think, all addressed. But essentially this is where you can sort of approve or disapprove things that it's learned that it wants your confirmation on. So oftentimes it will surface conflicts here, so like, "Hey, my streamlet said this, but my semantic view said that. What should I believe is the truth?"

37:47 And you can always answer with like a natural language answer saying like, "Hey, no, you know, you need both these definitions for different use cases." Status, just an update on the status. And then there is an accuracy piece that you can see there are some, you know, random eval questions that I had set up. And so you can also use this to run eval. So that's sort of what's in Cortex Sense. You can always check it with /sense and see all of the other sort of Cortex Sense objects over there. I'm then gonna show you what happens if you ask it to create an eval set.

38:18 So I happen to have some questions and answers that I already had in some folder i-in, in my machine. So I said, "Hey create an eval set using these... This natural language stuff that I wrote, and then go ahead and run it." And it went, and it created an eval set, and you can... I'm just gonna scroll down to the part where it actually went ahead and ran it. So you can see here there's about 15 questions. And so this first column is the score with Cortex Sense.

38:46 It actually scored a perfect score, which is really cool to see 'cause, I mean, I guess it's expected 'cause all of the questions did relate to stuff that I had in my docs and semantic views and so on. But then you also have a baseline, which is here's what Coco used to do before Cortex Sense existed, so before you had the skill to like get the context. And it turns out Coco actually wasn't really getting the right answer a lot of the time. A few of the times you can see here it did get the right answer. That's probably because the tables are well-named and so on. I tried to make it fair.

39:19 But there is a pretty sizable lift here. So, okay, then what? I should be able to use it, right, and ask questions in Coco. So for now in this demo, you just basically use it by typing /cortexsense and ask a question. What I will say is that we actually have something in the works where you don't even have to invoke the skill. You just start using Coco, and it will just get smarter once you have the Cortex Sense object, and it will automatically use Cortex Sense when you ask questions. So it'll just start to feel s-smarter. But here I asked, what's the... Our clean claim rate by payer? And then it answers with something.

39:57 This, this is accurate. I, I did verify earlier, but you'll have to take my word for it. And then actually after that I can, I can actually verify that it's accurate. So I can say, "Hey, tell me about all the context you used to answer this question, and where did that all come from?" And so, you know, the first thing it does is it confirms, "Hey, yes, I did use Cortex Sense." And then it will actually tell you exactly what it had in its knowledge to, to get this answer. So I had this clean claim rate, I had some other stuff, and it will tell you exactly what the definition was.

40:31 And later you can also ask it like where did this KPI even come from? And it should be able to tell you that provenance information. But the other thing you can do is if you don't like the answer, you can always say, "Hey, update your knowledge," and change... You know, "I don't like this answer," and it will update its knowledge based on that. So this is all stuff that you can do, assuming you have the permission on the Cortex Sense object. So with that you know let's say you wanna actually use it with an agent and not just with Coco. I have a little bit of an instruction here that is in our private preview.

41:06 You have to type this. But we're turning this into a skill to make it easier. Essentially you say, "Hey, I have this agent, and help me enable the Cortex Sense tool for this agent." It's just an experimental flag for now essentially that I had to instruct it to add. And it goes ahead and adds it to the agent. And so with that, maybe I'll also show you what it looks like in an agent. So let me switch my screen quickly. I know we're almost at time, so I'll try and be quick. This is I have two agents here side by side.

41:40 This one is an agent without Cortex Sense, and I'm gonna ask it something that it really shouldn't have the answer for, which is what is our clean claim rate by payer? And this is one with Cortex Sense. I'll do this in parallel in the interest of time. What is our clean claim rate by payer? Oh, I need to refresh this page. Hopefully this works. This was just working earlier today. Maybe I'll show you the, the run from earlier. Here you go. So if I ask what's the clean claim rate by payer in this case it actually... You can see here it actually gives me a smart answer. And it's...

42:20 It says, "Hey, you know, like this is the table." By the way, this is exactly what Coco gave me earlier, the same numbers like Cigna, Humana, whatever else. And then it... You can also ask it still like, "Hey, where did you get this information from?" In the case of this one without Cortex Sense it actually has no idea what you're talking about and sort of needs a bit more information. So you can see that even the... A-and I'll... I guess I'll show you really quickly the configuration. There's no real instructions besides you have a Cortex Sense tool, and then in the tools there's no semantic views, no search services, no other source.

42:58 So it actually just knows all of this information from the one Cortex Sense object.

Abhinav Vadrevu

43:04 So I know we're close to time. I'm gonna pause here and maybe if there's time for any questions at all, we can get into that. But yeah.

Charlie Freeman

43:13 Fantastic. Thank you both for, for walking through this. I think this was amazing to, to see live. So one question that came up in the chat from Shannon Hughes, which I think is a, a good question specific to where we see a lot of kind of joint value for this partnership is, how does the governance work of what Cortex Sense is providing if it auto-generates information how is that validated, certified for use? So naturally, I I see that as a, a great position for the two of us to work together, but Abhinav, would, would love your thoughts.

Abhinav Vadrevu

43:45 Yeah, absolutely. So I think there's two here. Like, one of them is you wanna get from zero to one, and you want ad hoc questions answered, and you want something that's very managed. So it's like two sides of a spectrum. It's like you get more managed, but you get a little less human control. And then on the other side is, like, all of the human-managed context that you get that level of control and governance. And the more you give up human control, the more managed it gets. But the less managed it gets, the more human you have in the loop.

44:14 And so it's a bit of a you need to use both for different use cases, and the way I see it is that semantic views and maybe even, like, the context in Atlan and stuff that you sort of human curate really exists for that governed, validated, certified use case. And then today, if you're using agents with SQL Exec MCP, Cortex Sense is strictly better and will just raise the baseline where you don't have that sort of governed context. And so most customers that look like us, the ninety-five percent of their data has no real semantic user context, and that's where you can really raise the baseline.

44:49 Or even testing new agents for new data that doesn't have a semantic view, that's where you can really get value.

Charlie Freeman

44:58 Spot on. Thank you, Abhinav. So we had another question in chat. I'll take a, a quick first stab at it and then Razi if you wanna cover anything I missed there. So Vishal Kumar question was, help me understand i-if there's context already built into Atlan using all of Atlan's connectors, and Cortex may have similar connectors to build the context why would we need both the products in an enterprise landscape? And I think it, it always depends on the, the use case. As Razi kind of walked us through, one of the, the big pieces that our customers have been obsessed with us for has been those context agents.

45:33 So how can we accelerate the AI-ready data that then Cortex Sense could pick up and use for its inference can be the first piece. And then there will be other components like governance workflows or a rich business glossary attached to that context that Snowflake will, will not have that we could pipe in there for said connectors that will also kind of give you an additive layer on, on what you're working on. But Razi, I don't know if you feel like I may have missed anything there. That was certainly my attempt.

Razi Shafiq

46:03 Yeah. I answered the, the question in the chat as well. For that, connectors are just the first step of, of context and governance. And as Charlie mentioned there's going to be workflows, versioning certifications, et cetera. And so Vishal, I'm happy to work with you on a call to discuss this deeper. But it-it's kind of, especially at an enterprise layer, connectors again are, are just that first step of collecting the context. Then it's the generation and enrichment as well as the governance around the approvals of who's approving changes what does it actually mean, as well as certifications.

Charlie Freeman

46:45 Fantastic. And, and as we wrap up on questions if any final ones, please feel free to put them in the chat. We do wanna highlight our Atlan Context Conference some of our amazing speakers some extra folks from Snowflake joining that one as well. That'll be October twenty-eighth at eleven AM Eastern. Please join virtual. It's a, it's a fantastic event. You'll pick up a lot of great nuggets into how you can better think about Atlan, Snowflake, and other great partners into the problems that exist in your, your enterprise today. Looking at any final questions to come through. I think we've covered most of them. We'll certainly share a follow-up of the recording with everybody.

47:27 If we missed any questions and we, we find them later, we'll make sure that we address them directly and, and reach out to those. But other than that we thank you greatly for, for your time here today with us. Thank you Abhinav, Razi, Austin. It's an exciting time for the partnership, and, and we're really excited to, to build with everyone here. Thank you, folks.