Governance That Drives AI: From Pilot to Reality, with Atlan and Snowflake

Austin Kronz

0:00 All right. Welcome everybody. Really excited for today's webinar. We have an awesome crew assembled today to take you through Governance That Drives AI from Pilots to Reality. I'm Austin Kronz, the Director of Data and AI Strategy here at Atlan, and let me introduce you to the awesome cast that we have to go through today's session. Sharif, please, start with you.

Sharif Karmali

0:24 Hey everyone, Sharif Karmali, and I run Product Marketing at Atlan. A long career being both a data producer and data consumer at times, amongst other things. So very excited having played a lot of roles, and how at Atlan we get to help bring all of those diverse people together to collaborate and get to success. And I think I'm popcorning it over to Gary.

Gary Patterson

0:48 Good day, everyone. My name is Gary Patterson, and I work in Product here at Atlan. I'm really excited to have the opportunity to show you a little bit about what we've been up to in the product, and how we think that we're gonna be able to help everyone with Snowflake and Atlan work on bringing AI into reality.

Austin Kronz

1:04 Awesome. And Mayur.

Mayur Mahadeshwar

1:06 Sure. Thanks. Welcome, everyone. My name is Mayur Mahadeshwar. I'm a Partner Solutions Engineer at Snowflake. I work very closely with the Atlan team on all things Snowflake and Atlan, so just excited to be here. My background, I have spent over a decade in big data and data engineering, and also very passionate about the governance space. So yeah, look forward to talking to everyone today.

Austin Kronz

1:32 Awesome. Well, we do have a jam-packed schedule, but it's an Atlan webinar, and you know that means we need to start with the audience. So let's get everyone involved right off the bat. I want to hear from you. We're talking about AI today. We're talking about moving things into production. So help me finish this sentence, "If our AI app would be in production today, if we could just figure out," blank. So we want to hear from you. I'm expecting a lot of similar answers, if I'm being honest. This has been a problem that's plagued the space for a long time. Looks like a couple people typing. The meaning of our metrics.

2:10 Yes, a very important ingredient. As always, the slide OID is important to moving models into production. Anyone surprised by this first one? Gary, Mayur, I, I think it's probably what we expected to see here. Metric definitions again, data governance, governance in general.

Gary Patterson

2:32 Not surprised by security— Yeah... Approvals, guardrails in general.

Mayur Mahadeshwar

2:37 Yeah, that, that, that's probably the answer, like nine times out of 10 it's, it's not about the model, it's about the messy data, the information or the lack of it, and the governance blockers.

Austin Kronz

2:48 Makes sense. Awesome. Good data. I, I love these ones that are coming in. So keep this coming. We'll have the poll stay open. I think we're gonna cover and hopefully answer a lot of these questions and, and points that you're bringing up throughout the session today. So with that, just a quick overview of everything we're gonna talk about. We're gonna start with Sharif, who's gonna give us a breakdown and really help clarify this problem statement. And I said it's plagued us for a long time, but I think we have this new way of looking at it. Really excited for Sharif to take us and help us cross this AI value chasm.

3:21 And then we're gonna hand it over to Gary and Mayur, who are really gonna talk about what are Atlan and Snowflake doing together. It truly is a partnership and there's a lot of collaboration that needs to happen to make these things work. But then we want to leave it open to the group, so please continue to use the chat and the Zoom Q&A to submit your questions as we're going through any of this, and we'll do our best to answer them live at the end in an open question and answer. So with that, I won't waste any more time. Let's get to what you came here for. Sharif, please take it away.

Sharif Karmali

3:51 What's really great about the vantage point of running product marketing is I get to talk to a lot of our customers, my favorite part of the job, and understand what is going on inside their companies in a lot of detail, and that's the basis for a lot of what I'm gonna talk about today. The first thing is, and I had one CDO describe this to me as, like, they, they joined just before, their company just before ChatGPT made its debut in the world. And they were like, "I thought I was gonna have one job, and then my job is completely different from what I thought it would be."

4:24 We're sort of in this stage right now we call AI curious, and the best way I've heard this described is today we are thinking about when should a human go ask an AI to help in its workflow. But we're heading toward a new era of business where AI will be thinking, "When do I need to go ask a human to help in this workflow?" And those roles will switch. And the AI native era is what we're calling that, what a lot of people are calling that. And it's not just another technological innovation. It really has the potential to be like electricity, where it will rewire every industry.

5:08 It will be presumed to be there and available, and you have to be an AI native business. Like, the factories that could harness electricity looked nothing like the factories before electricity was invented, and you have to do that in order to plug in machines. You can't slap a Tesla dashboard on a horse-drawn cart. But becoming AI native right now is a struggle, and that's really the topic of this webinar. When folks do studies independently at large scale, we get these sorts of statistics. 85% of all AI projects or pilots fail to make it all the way to production. 42% of companies say they abandon most of their pilot AI initiatives.

5:56 And this quote on the right-hand side is awesome. This is from a CIO, one of our customers. Sums it up so well, "The things that we thought would be hard initially a few years ago were actually easy." Getting people to adopt AI, figuring out the models, figuring out architecture is like RAG. The things that we thought would be easy because we had been working on having data and context and on analytics for a long time were actually really, really hard.

6:24 And there is this chasm between the use case in a pilot-constrained limited form and AI in production where it actually can deliver value at scale and help you transform and become a leader in your industry. That chasm is something we've been tackling with a lot of enterprises across a lot of industries, and we feel like we understand it fairly well now. And our customers, some of them are listed here, some of the industries are listed here, but not all of the customers and not all of the industries. If you add them all up, enterprise value over ten trillion, it kind of blows my mind. It's like the GDP of a decent-sized country.

7:07 So we've seen this in a lot of different environments, and we've distilled really three hurdles that make it so difficult to cross this chasm into production. Before I share them with you, we're really curious, you know, you all are seeing this every day as well. What do you see derailing AI pilots most often? Some of these already came up in the Slido, right? Like the dirty missing data, lack of governance. Interested to see which ones come out highest in this poll. Dan, I love that Jennifer, like, if you chose other, write what else is coming up in the chat. That's great to see as well.

7:49 I wanna open it up Mayur, Gary, Austin, reflecting on this, is this what you would have expected?

Mayur Mahadeshwar

7:57 Yeah, yeah, more or less. And based on what I have seen, you know, most of the time in pilots, you have a group of really smart engineers hacking away at a problem statement in a very controlled environment. And when you try to take that to a production scale or an enterprise scale, the solution does not scale. And if it does, there are lack of governance considerations. Like somebody would build a very complex complex model and but the success of that model basically lies in everybody having access to some PDF in a, in a SharePoint drive. Now, that's, that's not going to happen in an enterprise.

8:36 Like how many people actually, you know, have access to that sort of data? Now, many GenAI pilots lack visibility into what data was used, how it was transformed, and whether the model outputs are explainable. This, I feel, is critical for production-grade AI.

Sharif Karmali

8:53 I agree. And it's, it's interesting. I feel like looking at the one, two, and three, you think about almost the order in which organizations tackle some of these problems. And, like, once the data context is sorted, you know, then you start to realize that expanding that to more prompts and more use cases, the, the semantic layer can't actually cover all of that, and business context becomes more of an issue. So it's really fascinating to see what people are facing most and which ones resonate. This lines up pretty well to, you know, as we fill in these three boxes.

9:27 But what hopefully you'll take away from the next few minutes is understanding these more deeply, and it starts to lead into the solution. So the first one this is a quote from a CDO at one of our customers. They're a technology company. "I have a thousand AI use cases in my roadmap, but I don't even know where to start because I don't even know what data we have for those use cases."

Sharif Karmali

9:51 The left-hand side is the pilot world. In a pilot, finding and understanding the data can be relatively easy. And if you can't quite connect the systems or get what you need, you can make compromises, you can pull things into a sheet or an air table. It's sort of like you stitch it together, and you can get there. But on the right-hand side, that is the actual landscape that Matt Turk maps of all the data and AI tools, and that doesn't even cover all of the transactional systems, all of the different places where data context can live. And also some of it is context that people hold that isn't even digital yet.

10:27 So mapping all of that data context that's needed and making that available to AI or to humans working with AI, that's the first big hurdle. For the next one, this is a great quote from the CIO of an investment management firm we work with. "TAM might mean a hundred things in the open internet, but at an investment firm, we have a very particular definition of what TAM means and how to calculate total addressable market. How do I get AI to understand all of that business context?" In a pilot on the left-hand side, business meaning you can constrain and it can be easy. And you probably have, in your organization, seen this.

11:08 It's like the, the pilot version of a GPT or a, you know, interface where they have a few suggested prompts and you can click one of those. Demos incredibly well. But as soon as you start to write prompts outside of that where you have people from different— on the right-hand side in production, you might have marketing asking a question about something like retention, and they're thinking of customer retention. But is that user retention, and how does that relate? Is that... You know, or if someone in HR is writing that, they're talking about talent retention.

11:39 So all of this semantic meaning becomes much more important when you open it up for prompts outside of the few that you planned in the pilot. And so business meaning is the, the nuanced second challenge. Third, this is a quote from a, a CDAO at a public software company, "I am okay with an HR chatbot using my payroll data, but I don't want any other data and AI use case in the company to touch payroll data. How can I be sure?" In a pilot on the left-hand side, you can keep everything narrow from a governance perspective, narrow use case, data set.

12:20 You can scope it so that only certain regulations that already exist that you understand really, really well apply. You can do a narrow set of users of who's allowed to access and use this pilot, and you can keep the impact or potential harm relatively narrow, right? It's an internal Q&A. There's a human in the loop to gut check everything before it makes its way to a customer or a decision. But at scale, if we're really thinking of AI native value delivery, we're rethinking entire business processes. We have unpredictable use cases, broad need for data because quality will be critical, so the more broad data access, the better. Regulations are not only broader, but also changing.

13:02 Diverse users with different varying permissions that need to be considered. And the potential as AI delivering value means AI sometimes not having a human in the loop, taking action that's high potential impact or harm. One VP at a company we work with shared this with me this week actually, which is they had a hackathon, and they had three people make three different bots in two hours. Like, I don't think we can actually stop people from building, but what we can do is control what's certified, what is agreed upon to be trusted and trustworthy to use for action and decisions. And that's how we think about the governance hurdle in more detail.

13:44 Putting all these together, if you cross these three hurdles, it's what I just said. It's that on the other end, you have things that are trusted and trustworthy, and that is how you can scale things into production and truly transform workflows to be AI native. But context, meaning, and governance live in all sorts of places. So the way we think about Atlan's role and its design is there's a few things you have to do. The first at the bottom of this plane is you have to unify context, meaning, and governance from all of the digital places that exist today.

14:16 Yes, this is the data and AI systems in the pipeline, but it's also places where meaning resides. It's business systems. It's also places where governance or principles or policy reside in privacy and security and policy and those sorts of systems as well. So everything that's digital, if you can unify all of that bidirectionally, so that as you update it in one place, it updates in those systems that are being accessed sometimes directly. That's step one. Step two is you need to capture all the context, meaning, and governance that doesn't live in a system yet. And there that's obviously like a vast collective memory of every human thing. So it's not feasible.

14:57 It doesn't make sense to just capture that for the sake of capturing it. The collaboration needs to be purposeful collaboration around what do I need to build a data or AI product. So that's where we think about facilitating the data and AI product life cycle, where you pull in cross-functional stakeholders to give that human context around a data or AI product. And then finally, once you have that, you need to be able to activate it. If AI is electricity and AI is feeding off of metadata, then metadata needs to be like electricity. You need to be able to bring it up in any experience anywhere.

15:37 When you combine unifying, collaborating, and then activating that context, meaning, and governance, and you think about left to right context, meaning, and governance being the way to bridge the chasm, that's the way we are framing all of the different experiences and capabilities of Atlan going forward. So at the base of it, we have what we call an Atlan Metadata Lakehouse. It's built for the scale of workloads, the diversity and interoperability requirements, the enterprise requirements around security and control. So that Lakehouse foundation is what everything is running on.

16:14 But when we think about unifying, that first step, that's where, yes, really deep, great out-of-the-box connectors, but also a framework to connect to any type of system very quickly. And again, not just data pipeline systems, but any system that could hold meaning or governance as well. Creating a knowledge graph of that which makes end-to-end lineage really quick and having everything sync bidirectionally. So that's unifying. Collaboration needs to happen both in the context realm, data engineers, architects who have context on data and how it's being transformed, analytics engineers, they can collaborate in the catalog embedded in dbt YAML, wherever they're documenting, however they're creating and building across the data product life cycle.

17:02 You have meaning, so you have ways for the business and other stakeholders to engage on what the semantics are, what things mean, and glossary data contracts that can be consumable in a business format as well as a technical format, experiences for collaborating on what quality actually means. And then on the governance side, where stewards and others, legal compliance, privacy teams, IT teams can collaborate and add their context on what are the policies that matter, how should they apply to assets, when should we allow by exception them to apply versus as a rule, and AI governance as an important part of that. And then at the top, activating that.

17:45 We have AI-augmented experiences that live in Atlan. We also embed in the places where people are working, Slack, Teams, GitHub, Jira. And we're thinking about AI native experiences. So really excitingly, if you follow us, you might have seen we just launched MCP Server, which is already being used by many of our customers to feed metadata into their AI native workflows and their chatbots. And we're thinking about that in partnership with Snowflake, with Cortex, and other opportunities as well. There's a lot here to digest, so don't worry. You don't need to digest it all in this two minutes that I've shown it.

18:21 Instead, Gary is gonna walk through the context-meaning-governance journey and what some of those product experiences can actually look like in partnership with Snowflake.

Sharif Karmali

18:32 So Gary, over to you.

Gary Patterson

18:34 Thanks, Sharif. Okay. So thanks everyone. I just wanna walk through a little bit about where we're gonna start, and I'm gonna do this first with the first piece that Sharif mentioned, which is context. And let's start where most of the AI pilots fail with that context. You can think about it as three different layers here, where how we connect to the data, so connecting into Snowflake, pulling all this rich information, how the data connects and flows throughout the environment, and then how we help tie it all back together. So if you go into Atlan, I'm looking at a home screen right now. It's personalized, and I'm gonna go into the assets.

19:19 Now, one of the first things that I wanna do is start looking for my Snowflake information. And if you're like most of us there's a variety of types of informations that you can have in Snowflake, everything from tables and views, databases and schemas, listings, which we'll talk about, even things like iceberg tables. We pull a variety of metadata in, including things like usage and popularity metrics. And if you're like most of us, part of the way that you value data depends on who's using it, how much it's being used. Oops.

19:58 So I'm gonna, I'm gonna pull up and do a filter here on some of the most used data just to find a good representative hero table, and here we go. We can see that this system has massive amount of data, so I'm gonna walk into this. And as we pull up the table, you can see there's a variety of information around descriptions, columns, tags, and so on. But what I really wanna do is walk into the lineage. And I wanna show you how we pull the information in and we trace through everywhere.

20:39 One of the things that's, that's really nice is that if we expand this lineage out, you can see all the way back to deep sources going back through maybe a raw zone, things that are landed into into Snowflake, going through transformations, ending up into this table, but even this table goes into other downstream systems. And if we think about what's happening in this downstream system, one of the things that we may wanna do is be able to have tags to be able to understand when we're using information that might be sensitive. So if I go into this, you can see that there's a number of tags that are possible.

21:18 I'm just gonna go into this edit view, and maybe there's some PI data that we're worried about. We might have confidential information. So I go in and I want to set propagation. One of the things that's really intriguing about the, the Atlan approach is that we can propagate the tags so it doesn't just reside in this one table, but can flow based on parent-child relationships or even up or downstream assets. So we typically set propagation in here, and we'd have bidirectional syncing. Right now it's disabled, but we would set bidirectional syncing and then run this, and it would flow everywhere throughout the the systems.

21:55 Maybe just, Mayur, to bring you in for a second, can you tell us a little bit about why bidirectional syncing and tagging is so important in Snowflake?

Mayur Mahadeshwar

22:04 Yeah, sure, and I'm glad you asked. This is actually one of my favorite parts of the integration because it's, it's super practical, right? So within Snowflake we have tags, but we also allow you to define masking policies on the tags. So Gary mentioned about PII data and, you know. So as a customer, you can actually create a masking policy so that if a column is tagged as PII, that's not visible to anybody, you know, based upon their role. And this can actually flow bi-directionally. So what that means is your data is consistent in your metadata layer as well as your data storage layer. And with lineage, it can propagate both upstream and downstream.

22:49 So you don't have to go around, you know, tagging every column and making sure that you are covered. You can just tag it in a central place and make sure that if a column is PII, it's PII across the entire data life cycle.

Gary Patterson

23:03 Great. Thank you. Let me, let me flip back to the slides here. And what we showed was that we're connecting information, showing that movement everywhere as it flows, and then we can sync it back to connect across metadata. Next thing I wanna show you is what Sharif talked about with meaning. And to do that, we really have a few areas. There's data products, which are these highly curated experiences for repeat consumption. We typically show them in marketplaces, and this is where we'll, we'll touch a little bit on a Snowflake listing and what that means. And then we'll get into data quality.

23:40 We, we surfaced earlier about this risk of bad data and how do you have good AI products if you don't have good data. But before I go into the demo Mayur, can you talk about Snowflake listings? What are they and how are people using them?

Mayur Mahadeshwar

23:58 Yeah, yeah, sure. So, so essentially, like, Snowflake listings are available within Snowflake Marketplace. What Snowflake Marketplace is, think of it like an app store for your data and your data tools. When I say data tool, I mean your models or your native apps. So any data that you need, now tying this back to our gen AI, you know, use case, like any third-party data that you need, say, ESG reporting, your healthcare data, can be obtained through the marketplace. Similarly, once you work on that data and create your model or generate some data as part of that model.

24:34 You can share that via listing to your consumers, not just within your company, but also to external users. That way you can actually monetize your data.

Gary Patterson

24:46 Thank you. So let's, let's walk through what this looks like.

Gary Patterson

24:50 We have a concept of data products, and we have domains. You can choose to go through something hierarchically, but a lot of times people aren't sure what they want, and they start browsing through one or more of the domains, and often leverage a search. So in this case, let's assume that I want to find some sales order operational information, and I come up and I see this sales order insights table. It's verified. It has what we call a data product score, which means it's pretty likely that it's gonna be good for us to use.

25:24 So I'm gonna go click into this, and in this product summary view, what you can see is rich information around the assets that are involved, but also as I mentioned, this, this product score, which has a variety of dimensions in which we rate the data. We'll come back to this in a minute. I also wanted to point out this notion of output ports. So it's very easy to put information into a data product. You can think of it as a collection of these, of these systems. And if I want to, I can just go into the assets and add more, more tables, more columns, and link everything up.

26:05 If I go back to the summary view, you can see a few output ports here. Let's look at that lineage concept again. So in this case of a data product, what I'm pulling up is all of the information across different domains. The colors here represent different domains. Imagine that I have different information, possibly from different business units, but that are brought all together into this sales order insights. I'm using it downstream into a dynamic pricing model, perhaps. And now what I really wanna do is be able to understand a little bit more about the output ports.

26:44 So as a user, when I want to consume one of those reusable data elements to that are fit for purpose, how do I go in and, and access it? I can see that there's this Snowflake consumer listing. I just click this button. I can get more information about the listing itself. But essentially, if I want access, I can just click a button. With one click, add a note, press submit, and then we will go through an approval workflow that can be done inside of Atlan. It can connect through approvals through something like Slack or Teams.

27:24 But essentially once completed, it's going to hook into Snowflake directly and give that user access to that marketplace listing that Mayur was just talking about to make it very easy to get to the data. The other thing that we talked about going back to the slide, was data quality. So we're very excited about this most recent announcement where we're now adding native support for data quality through Snowflake.

Gary Patterson

27:53 Snowflake has something called data metric functions that we're using for this. Maybe, Mayur, can you take a second and just explain to people what is a data metric function?

Mayur Mahadeshwar

28:02 Sure. Sure. So data metric functions or DMF are essentially... Think of them as built-in, built-in health checks for your system. Now, these are automated checks which you can actually schedule to run at regular intervals. They monitor your data for things such as number of null values, number of distinct values. Are there any values which break referential integrity constraints? And periodically, these checks run and output their results to an events table. Now, tying this back to GenAI in order to ensure that the output from your model is of consistent quality, you have to ensure that the data that goes into your model is of consistent quality, right? And it's sanitized.

28:47 So you can use these health checks on your data and make sure that your model outputs, puts are always depending upon whatever parameters you have defined. And if there are any violations within your data, you can actually get alerted on that.

Gary Patterson

29:04 Thank you. So let me, let me show you how that looks in here. What I'm gonna do, if you think, you'll, you'll notice that there's a Snowflake data quality area. I'm just gonna search for assets that we've configured with rules. And then maybe just to make it easy, if people are using, there's a, there's something called data contracts. I'm gonna just search for things that have a contract, and we'll, we'll whittle down our space here. And what you'll see is that there's two of these Snowflake tables. They have rules, and in this case, we see that there's a contract violation on this one. So we'll, we'll look at this one. Let's go into beverage orders.

29:47 What you'll see in the outline interface is in the overview is some summary information about the, the tests that have been configured, but there's also a data quality tab here. So if we go into this data quality tab, one of the things that really excites me about the partnership and the integration is that now in one place, you can go in and enable self-service for a business user or an analyst directly. Someone who doesn't have to go into the deep technical aspects or do a lot of complex coding. They can go in and, as mentioned, just select one or more of these quality dimensions. So there's certain standard checks, as mentioned here.

30:26 Maybe someone wants to run a null count or add, for example, maybe a freshness check saying the, the data should come in at least once an hour or once a day, whatever the cadence is. And when you create that rule, as, as you heard, it's behind the scenes, it's creating a data metric function in Snowflake. So it's running in Snowflake the information is being processed there, and then Atlan is reading this information back and surfacing the display. Because what we really wanna do is convert a technical check into a signal of trust. A more advanced user, however, may wanna use something like custom SQL.

31:05 In this case, it's, it's again, it's leveraging capabilities within Snowflake, but now maybe someone wants to do a join across three columns for a synthetic key or maybe there's some sort of accuracy dimension with a complex calculation within one column or a reference to other columns. So this is slightly more advanced view, and then they run, as mentioned, on whatever schedule is desired, and you get these outcomes through the UI that you can filter on and see what's passing or failing. We also show you over here, if there's certain impacted assets, what those are, and we visually display this throughout the interface to make it easy, again, to signal those trust.

31:48 So if I go into something like lineage, and now a business user wants to understand what's happening. Again, you, you can see that there's a summary here on this table with the results of tests, and I can use this to really get into things like root cause analysis or downstream impact assessments. So for example, if I hover over this table, I can see that there's an announcement that's been derived from some of the quality checks that says, "Warning, some of the data is not fresh. Pipeline night might need to be updated. You may or may not want to use this information."

32:27 In other cases, you might have a red alert because there's an accuracy known problem that's being resolved. The other thing I wanna show you here is that, again, as before, we can trace this information upstream. Atlan also partners with a variety of vendors. So for example, you could have information from Monte Carlo or Soda or Anomalo, where again, you can see announcements and alerts, or even information that's coming in from transformations, ELT, like Fivetran.

32:57 So a lot of information is available to stitch together and surface display to go from that technical check to signal that trust for that business user so that they have a good understanding of the data that's going to be feeding these models.

33:12 So once again, I wanna circle back and just do a checkpoint of what we showed you, which is we have data products, which is how we can go from raw information and catalogs to these highly curated experiences where we can share through marketplaces direct access very seamlessly if you're already having it in Snowflake as a listing, easy to click and have access to that, and we wanna make sure that that data is trusted that's going to feed into your model.

Gary Patterson

33:41 The last section that we wanted to talk through we saw as a challenge is really governance. How do we make sure that it's embedded, that it's invisible, that we have the right guardrails in place, and that we can monitor this not just once, but over time? The next thing I wanna show you is our AI governance module. It looks somewhat similar to the data products marketplace if, if you understand the listings and the views. And what I wanna point out is that we have both apps in here and models. So I'm gonna take an example that I'm somewhat familiar with here called Road IQ, which is analyzing road and traffic data.

34:19 If I click into this model, what you can see is a summary of information at the top. And then again, if you're, if you're thinking about quality and how you, how you trust what's in this model, you might notice things like the precision or the accuracy or some of this other information. If you're worried about regulatory I happen to be in the EU right now. You might be thinking about EU AI model and how you have the right checks for bias, fairness, and so on. We make it very easy when building these models and apps to bring the data. So again, imagine you have testing and training inference, these various steps that you're using.

34:53 I can just add a couple of these tables that I wanna use to bring into my training set, for example. If I go into lineage, it's a slightly different view of lineage here. What you'll notice is that it's more of a business or conceptual view, view of lineage, but we have, again, the information of how the data is flowing, for example, from training or or testing or validation inference and how it's flowing out. So it's very easy to connect and see all the information. The other next thing is if I go back to the overview, though, how do you make sure that you have the right sorts of controls?

35:35 So are there tags like is the data confidential? Are there policies that I might have linked that I want to make sure are, are the right guardrails against this information? So the last thing I wanted to show you is our policy center. In Atlan's policy center, it's very easy to add and create policies. We have an AI-assisted flow for the purpose, for the descriptions and so on. But I happen to have already created one for AI models and accuracy, so I'm just gonna go to this right now.

36:08 I did this, I did this yesterday, where I went in and created a purpose, and I just said, "Let's assume I'm a financial industry and I care about AI accuracy and precision," and it will start generating a purpose. And what's great about Atlan's approach is the purpose can then be fed to automatically generate a policy. I didn't write this. I just said, "Use Atlan AI to generate a policy based on the purpose I gave," and then I'll add some scope. The scope is anything that's an AI model, and in this case, I was concerned about accuracy and precision, so I said, "I really want at least..." 90%,.9 accuracy,.9 precision.

36:48 And unfortunately for me, when I did that it said, "All the assets in scope don't, don't match this. Are you sure this is what you want to proceed with?" And I said, "Absolutely, I want to make sure I have good data." I put this into an approval workflow and the policy ran. So if I go back to the overview of the policy center what I can see is that there's 10 incidents. And what I, what I created yesterday was this policy. If I open this up, what I can see is that there was a violation because I didn't have precision that was above 90%.

37:24 I can click into these linked assets, and what I can see is that this roadmap IQ that I was working on, if I just click into this, it's nice the sidebar keeps me in context here, is that my precision was not high enough. My accuracy was good, my precision was not, and so I probably want to go back to my model, iterate a few times, and that's what's gonna help me ensure that bad data is not the source of my issue. So hopefully what I'm showing you here, if I go back to sort of the takeaways, is that you want to have governance. You want to make sure that you're connecting.

37:56 You want to have the right policies and terms, controls around fairness, transparency the right ways to, to bring the data sets. And for auditability purposes in that lineage view, just quickly understand if I need to check and report on something, I can visually see it and export that out. But that's not just a one-and-done kind of approach. I need to have an ongoing real-time methodology, and that's where Policy Center easing the, the the burden of creating and managing and viewing in real time with notifications helps ensure that you're maintaining this on a constant basis. So at this point, Austin let me turn it back over to you.

Austin Kronz

38:36 Awesome. Yeah.

Austin Kronz

38:37 Thanks, Gary. Thanks, Mayur. I think it's really insightful going through the context, meaning governance. If you have any questions, by the way, please keep them coming either in the Q&A feature. We do have some help from the team answering some of the more tactical questions, but I think with this group, and if I could summarize some of the comments that I've seen coming in, I think a big question is just, hey, can we get a little bit more clarity on we're seeing this inside of Atlan. Where does Snowflake play a role? Where does Atlan play a role? Specifically when it comes to maybe data quality.

39:09 And, and Gary, if you could maybe even shed some more context on, well, is this just an Atlan and Snowflake thing? How does Atlan handle quality maybe more, more broadly or, or trust might be a better way to, to frame that. There's a question of like what tools are being used and, you know, maybe a little bit more— Sure, sure.

Gary Patterson

39:26 And if it, if it helps, I can pull up the instance again and show, but essentially we have a few different ways that we approach trust. We, we believe in visibility. So again, how do you surface and make people aware when trust may not be present? So we have things like announcements and badging that I was showing you, red icons throughout or information. We also have ways to pull in based on usage and context. So we want to activate the metadata, so we want to overlay something like a technical check with a lineage. We want to bring in collaboration and make sure that it's not a one-person show.

40:07 That you can integrate things through Slack or Teams or automate notifications through ServiceNow or Jira or whatever tooling you're using. And then people are increasingly getting to questions of things like contracts, which I apologize I didn't show you, but it's very easy to create a contract and make sure that producers and consumers are aligned in their expectations. So those are some of the ways that people are proactively changing the model around data life cycle, especially with data products. So people are increasingly thinking through that ongoing journey of how do I register information, connect it over time as they're going through a catalog and thinking through the requirements, how quality will impact with contracts.

40:50 They're surfacing it in products and making it available for consumers, and then publishing it with, with access control. Austin, let me pause. I don't want to take too much time, but we're an open ecosystem, so we work with a lot of vendors, and we're just super excited that as Snowflake continues to advance their capabilities for quality, how do we make it seamless and reduce that friction? Again, especially as people are storing more and more data into Snowflake and using things like Cortex and wanting to run experiments with AI.

Austin Kronz

41:23 Yeah, no, that's amazing. I, I think if I could double down on a couple points you made is that quality is a, is an indicator of trust. I think a lot of what our platform offers in a broader context is, you know, I obviously unders... Want to understand quality, but trusting a data product means I want to understand who built it. Is it certified? You know, can I see where this information came from? So there's a broader picture to trust, and then quality is sort of this key ingredient that you know, a lot of organizations are asking us about.

41:55 So I think it, it's, it's two pieces that I want to touch on, which is just the broader trust picture that Atlan kind of delivers and how Snowflake and the innovations at Snowflake are really enabling us to tap into our partner ecosystem to really lean on what you're doing with customers who are already wanting to leverage this kind of out-of-the-box quality and these metric data functions that Snowflake has. Mayur, if you want to touch on, you know, maybe or share an example of maybe what you've seen in this— Yeah, yeah.

Mayur Mahadeshwar

42:21 No, no. You guys summed it up pretty well. But yeah, essentially these data metric functions, they provide visibility into, you know the quality of your data. But eventually you also want to know, like, okay, this table has a bunch of null values, but how does that actually impact, you know, me going forward? Or, or how does that actually im-impact, you know, what, what my business users rely on? So those are the sort of things that, you know, that with this partnership, we, we help our customers achieve, and they get the entire 360 degree view of, like, missing data quality when they are in their, in their life cycle.

Austin Kronz

43:01 Makes sense. Just quick taking a quick scroll through some of the questions and comments.

Gary Patterson

43:09 Maybe I'll just take... There's one asking of how the, how the actual capability works. So if it helps, I can pull up a screen, Austin, and just show something. But essentially to Srinivas, your question is: Does Atlan leverage the Snowflake native quality capabilities and extract the scores, or does it run DQ rules within Atlan? And the answer is that we are leveraging Snowflake. So when I was showing you a dropdown creating a rule, when that rule executes, it work— we, Atlan is creating a data metric function in Snowflake for that table. Depending on the schedule of the table, it's running in Snowflake. It's using Snowflake's compute. The data is never going into Atlan.

43:54 When it runs, Snowflake writes it out to a configuration output table. Atlan reads the metadata from that table, and that's what we're surfacing back into the Atlan UI.

Austin Kronz

44:08 Yeah, I think that's really helpful.

Gary Patterson

44:08 Mayur, anything to add?

Mayur Mahadeshwar

44:10 Yeah. No, no. There's actually, like, an events table which contains all your metadata, and that's where, you know, Atlan pulls the data from, so...

Austin Kronz

44:18 Yeah. Amazing. There, there were a couple questions about configuration. Maybe, Gary, if you could start just even a step back. When we're talking about these integrations with Snowflake, maybe just give a high-level overview of, like, how do we... When we say connect to Snowflake, I think sometimes we take that for granted, right? Like, what are we getting from Snowflake? How are we doing this column level lineage? Can you just maybe walk us through what it means to set this up? What, what do we need, and what do we need to execute to start bringing in this information?

Gary Patterson

44:48 Yeah. So we have something called a marketplace where we have all of our connections and what we call miners. And the idea is that we make it very easy through a template to connect into Snowflake. So once we're connected in, we have a connection into that Snowflake instance, we start pulling all of the technical, business, operational metadata and such through miners. So we grab the technical metadata, so you can imagine us crawling across the Snowflake instance. We're pulling in all the schemas and tables, pulling in all the various forms of data. And then we also have something we call miners.

45:21 And so, for example, when I was talking through usage and popularity metrics, we want to know what queries are being run against Snowflake, who's running them, right? So we pull that information in, and we overlay it. For things like lineage, we have a variety of approaches where we can look at transformations. We can work with other third party, pull in their metadata. We can look at for example, parse SQL, and we can start stitching together and seeing how is the data flowing. So there's a variety of techniques that's probably deserving of a, of a separate webinar, but we, we, we start looking at the instan-instance and information.

45:58 So I think I mentioned at one point things like ELT that was occurring through Fivetran or maybe people running dbt jobs and running transformations. So we have all the information that we can parse and bring together, and part of what makes Atlan special, i-in my belief, is how we make this mostly seamless for the user, but we bring the metadata to-together in, again, an easy experience to consume. That, that browser sidebar can exist in any on top of any application.

46:27 So if someone's in a BI tool, they have an ability to just stay in that analytics dashboard and start seeing what does TAM mean to that VP that Sharif was talking to, or maybe someone is saying, "I don't understand. There's an announcement that says the pipeline's broken. I may not wanna use this data," right? But they might wanna understand how the TAM was calculated, and in the, in the sidebar, we would show them the, the actual calculation or we would show them up or downstream lineage where they might, might wanna understand can I go get access to something else or send a comment back. So that's how we call activation of metadata.

47:05 Once we, once we first ingest it, we wanna activate that metadata to make it more valuable.

Austin Kronz

47:11 Yeah. That's really helpful. So there, there's a question coming in. I think one of the most exciting pieces about the Snowflake-Atlan partnership is the bidirectional aspect of this. I think before we were limited to doing things kind of farther right in the pipeline, and bidirectional actually allows us to push things back, you know, up a tier in this data life cycle. There's a question from Sandra that's actually asking a, a, a unique question about, you know, if we're having things happen in multiple places, is that maybe a concern for end users?

47:41 I'd actually like to lead off answering this question because I think the beauty is if you are using Snowflake tags further up in your development process, meaning, you know, as data comes into Snowflake, I'm sure there's certain global policies that might want to be tagged at a Snowflake tier. What we've seen with customers is that there's more than likely other utilizations of the tag function for newer policies that may not be these global standards that need to apply to all types of data. Maybe there's a domain-specific policy, a regional policy. There's some owner that's outside of the engineering group that needs to then apply a tag.

48:16 Their interface into the world of data more than likely will be something like Atlan rather than applying something directly in the Snowflake ecosystem. So what's great is you have all these tags that can come and be just immediately pulled in from your Snowflake ecosystem, but then from a bidirectional standpoint, if I ever want to add something in Atlan, I can also push that back into the ecosystem. So now when engineers go to build, I don't know, new views, tables, just do more development on top of the data, those tags are already existing and can actually be inherited in any type of new use case.

48:46 So I totally agree, Sandra, in your comment about— Our whole philosophy is to shift governance as far to the left as we possibly can. And I think bi-directional actually allows new types of personas to participate in that shift left mindset. But Gary, I don't know, Gary, Mayur, I don't know if you have any thing to add related to that.

Gary Patterson

49:06 I, I wanna add one, I wanna add one thought, and it's partially because I didn't show it in the, in the demo, and, and I can pull it back up if it's helpful for people. But I didn't talk about data contracts too much. I, I, I forgot to show that in the in the demo. And what I would say is, yes, we believe in shift left, but we also believe in starting with the right with the business consumer, especially in the, in the context of data. When we, when we say that, that context and that meaning, the value of data is often unlocked by that business consumer who's doing something with it.

49:34 And so a contract is a way to help the producer and the consumer be on the same page. And so things like tags are often... So in, in Atlan, we can automatically generate a YAML file, which is a technical format for the information around assets or data products, and tags is one example of that. So we have that technical format, and we leverage AI to put that into a more human-readable natural language version, but it has all of the same metadata around the description, the users, the if it's a table, the column and information and types, but also things like the tags and retention policy and such.

50:11 So that context is, again, why you wanna have bi-directional syncing, and you wanna be able to connect things like the data engineer who's working maybe in, in Git or some more technical environment with a business user or an approver or steward who might be going through Atlan as part of an approval process to make sure a change is happening, but doesn't break something downstream.

Sharif Karmali

50:33 Yeah, I wanna chime in on that too, especially in the context of a AI world. In addition to the shift left, there's also an important shift right that Gary's talking about. There's a customer I was talking to that they've piloted a analytics chatbot, and they're like, "We actually are not trying to constrain people to only ask about things that are already in our semantic layer. What we actually want is if they ask about a term and it's not well-defined in our glossary, we want first the chatbot to not hallucinate and say, like, 'I don't know.

51:06 I can't, I don't know what that term means.'" And, and they're thinking through like should, if the user suggests like, "Oh, here's what I mean," or, "Here's how I would wanna calculate that," not just taking that and adding it to the semantic layer, 'cause that's very ungoverned. But right to left triggering a workflow in Atlan, which would say, "Hey, here's the suggested term for the glossary. Let's run that through the workflow we've built for review and approval, and then add it in a, in a governed way into our semantic layer."

51:34 So this, in an AI world where you cannot predict everything happening at the right, this right, shifting things right and having this two-way flow of metadata is critical.

Austin Kronz

51:46 That's awesome. Yeah. I think connecting the worlds of business users and technical users, I think has been the name of the game for, for a long time at Atlan. So that's, that's great. I see a couple questions here. One came in from Sumed. We showed a lot of scores, and I think in data quality and just metadata management in general, the concept of scores comes up a lot. Gary, maybe you can shed some light on when we say things like data product score what is that, and how might that relate into the data quality information that we're pulling in from Snowflake?

Gary Patterson

52:23 Sure, sure. As people are, are, are bringing assets together and combining them into data products and, and, and output ports, there's different dimensions that people typically wanna think about, like is the metadata complete? Right? Is the metadata populated? What is the trust of, of the data? What is the accuracy? So we have different dimensions just to, to, to keep it similar, straightforward, different dimensions that we can measure, and then we, we produce a score from a one to five. So, so the idea is kind of like a, a rating that you might have in, you know, in Amazon or Netflix or something else. How, how good is this?

52:59 How you know, and so you can have like usage and other, and other elements, and that's where quality would also fit in. So there's kind of two ways I would think about quality. One is just what is the score of quality on a column or an asset, and how do you wanna think about aggregating that up into maybe something like a glossary term or looking at it from a business unit as a, as a domain or subdomain or within a data product. There's also elements of quality scoring over time. You know, are you looking at things like drift? Is your model... So, so again, we looked at things like accuracy and precision.

53:31 Are we seeing shifts or changes that we need to be mindful of? But we can take that information of quality and, and, and surface that into other things like a product score or when you think about, again, a model, bringing quality and layering, compounding metadata is important to, to keeping not just an initial guardrail, but keeping things under constant monitoring, good governance.

Austin Kronz

54:00 Thank you. So I think we might have time for, for one more question. There were, there were a couple different questions about alerting. Tell us a little bit more about when these data quality rules happen. You shared a little bit of it in the demo, but maybe just rehash for the group, like how do, how do I become aware of them? Is it integrated into something? You know, what, what does it mean to alert or identify something when it comes to data quality?

Gary Patterson

54:28 Sure, sure. So I'll, I'll give you the generic example first, which is in Atlan, we have various methods to communicate and collaborate. So this could be through something real time, where maybe you just wanna chat about something that you're seeing. It's not even based on a test. You just wanna say, "Hey, I, I'm looking at something. I have a question. What does this mean? Is this right?" You might be using Slack or Teams, just having a conversation that you can track, and you can have it monitored or, or collected over time, versus maybe you have certain failures, and now you wanna trigger something off for workflow.

54:59 So we have concepts of workflow and playbook, and that's where you might want to, for example, send something out like a Jira, create a Jira ticket or send an alert to an, an inbox. We have reporting. Reporting is something that you would also look at. So there was a notion about scoring. You might want to look at things like, do I have data products or output ports more specifically, where there are no quality rules configured? You know, that's kind of an indirect measurement that, that should drive action.

55:28 But when you have a failure, especially something that's high severity, like in a contract, especially where it's maybe, maybe it's a, a very much of a break in, in whatever you do for trust or fitness of purpose of that decision, that's where you typically would want to avoid alert fatigue and actually send something. It's very critical that you send something as, you know, "Don't use this." So again, variety of methods. It could be an announcement in a UI, whether it's Atlan or not. It could be something out-of-band third party, like going to a ticketing system or communication channel. Or it could be something stat more... I say static.

56:01 It could be something like a report or a dashboard analytics that people are viewing within the application.

Austin Kronz

56:06 Yeah. That's amazing. Alert fatigue, very real, so I love this idea of a customizable approach or a personalized approach for the type of use case. So, well, I, I think with that I I'm gonna wrap things up here with a Q&A. I cannot thank this group enough. Shareef, thank you for helping us cross this AI value chasm or start to cross the AI value chasm. Mayur and Gary, thank you for showcasing what Atlan and Snowflake are able to do together. I think, you know, there's always just more to come and this has been a, a great great session. So thank you to the audience for attending.

56:39 We'll have some things to follow up with, and of course, we will be at Snowflake Summit, so please come say hi to us whenever we're out in San Francisco, and meet some of our customers and hear how they're bringing a lot of this to life. So with that, thank you everyone, and we'll see you on the next one.