View Recording: How Data Problems Derail AI Projects—And What to Do About It

Suneer Mehmood 0:46 Welcome everyone. Thank you so much for joining. This is a webinar on getting us AI ready on Azure and the governance strategies. That are ideal and technically required in the future and how we can approach that. So yeah, with that. I welcome you all myself, Sunil Mahmood. I work as a data and AI architect for concurrency. So yeah, like I have slides around like. Roughly 25 slides, so I’m I’m gonna time it and and explain it and also keep a track of you know the time and try to pace it through. That being said, please feel free to ask your questions and I would appreciate if we can keep when I I’ll be attending to those questions towards the end so that we can cover everything that is needed. But I’ll try to make it as quickly as possible as well. So thank you, Paige, for confirming that you can see my screen. So moving on so. What are we planning to cover in this webinar? We’ll talk a little bit about the data and AI maturity curve. Why governance matters right? The key pillars of data and AI governance. Tools and services within Azure. Little bit of deeper look into that. Some architectural patterns process recommendations and and common governance challenges some case studies. In fact, one case study. Some upcoming regulations from current and upcoming regulations, and then we’ll conclude with that and move on to question and answer Q&A in fact. I hope everybody is OK with that ajuna and moving on. A lot of you might have seen this data and AI maturity curve before. If not, I’ll spend some time to explain this. This talks about like you know how an organization starts, right? Like mostly organizations start with. Like minimal data infrastructure, right? So they would have missed business needs to actually analyze some data as they start start off, start something like some manual processes with spreadsheets and so on and so forth. And and and have different methods of. You know, crunching those numbers. Might not have a standard, you know. Like a way of, you know, calculations and and and and so on and so forth. The data and the code base of whatever they do well, there is no nothing to say as code base per SE in this particular phase. But like you know, all of those processes will be spread across across different systems, right? So that’s the first step. And the next step. Of the maturity phase is the descriptive phase where you would go into a traditional. Data warehouse environment, right? So that’s when you get to do some. Like foundational data warehouse? Some facts and dimensions, as you call in a typical data warehouse language. Facts are. The numbers and dimensions are the qualifiers for those numbers, right? Like you know, it can mean the customer a product and so on and so forth. Right. Traditionally a data warehouse would have data coming from 1 system and you would have repeatable visualizations like Power BI on top of it where you could like, you know, update the data and see new numbers. And so on and so forth. So that is more like retrospective or descriptive reporting, right? Like you’re looking back into the history. Now moving forward, like in a few if your data estate gets matured a little bit. You move into a modern data framework, right? So a modern data framework would mean that you’ll get a little bit more of diagnostic capability into your data. That would mean that like you know. You would have a centralized data warehouse. You would have a data lake mostly self-service BI and from you know you will have a holistic look of your your data altogether. You know if. If there is a system, well, if your organization have several Erps and you wanna combine all of that data together, this is where this is the maturity point where you know, you would say that this is you have a modern data frame. Work and you have a holistic look at your data, right? So that would mean that you would create a data cataloging you know you would have data quality rules, manage your lineage documentations and so on and so forth. Now these are all retrospective, right? Like now we’re getting into the predictive world. Like what can we see for the future? Right. We have historical data now. Like what can we see? You know, into the future. So when we talk about machine learning, we have the concepts of like a linear regression which helps us predict, for example sales numbers or or you know. Revenues or whatnot, right? Like you know, we, I have the historical data. I have a machine learning model for my linear regression to actually predict what can happen in the future, right? So I have good enough data. Qualified data to actually train a model. And test that data and make predictions out of it. Going on a step further, like we get into the wisdom or optimization or a prescriptive, you know level. So which would mean that you can further use that data like for sentiment analysis or customer classifications or segmentation in improper machine learning language. We call it logistic regressions. Excuse me. So that is like. We can do several categorizations, right? Like you know, hey, you wanna see if your customer is satisfied whether you wanna see how you know, where do you wanna invest more? What kind of a business segment do you wanna invest more? What kind of a customer category should you focus more if you have not categorized that like you know? And so on and so forth and the past two years, we have seen a significant boost into the innovation part of it, right, like. You know, generative AI is this is a buzzword, right? Like you know LLM, large language models retrieval, augmented generation and all that. We have been been hearing about and it has started influencing us quite a lot, right. It has democratized this artificial intelligence to, you know, widely, right? All of us use AI in some or the other form. Earlier we used to use AI without our knowledge. Now we actively use AI. One classic example is like we use copilot summaries to understand like you know what, like nobody has to write any summary session, right? Like you get the summarization yourself, you get a complete look into who spoke what. You know, what are the takeaways? What are the to Do’s after that meeting and and so on and so forth? You can actually. Draft A, you know, pretty professional looking e-mail. With the help of copilot. AI, right? And then you can not just text or or framing of words or or drafting drafting emails. You can also create images. You can create videos. You know the content generation industry, like the social media, the content generators have been using this primarily, right? Like you know as well. So it’s getting into an interesting world. So. These are the like you need. You need very good data for for you to have good models for you to actually ground your your models based on your internal knowledge base. So these are the different maturity levels that you would have in our data. As you know, data landscape and then based on your scenario, what do you wanna implement? What do you wanna wanna go first? You know your prioritization, you productize that you know. So I hope that makes sense. And moving on. So let’s talk a little bit about why does government governance matter, right? Like what? What is governance after all, right? Governance means you are really on top of your data quality, right? You are. You know what? Your data estate is inside out. You know, you manage really well, you avoid any risk and compliance right there by building trust and transparency for not just your data process, but your AI process as well, right? So when I talk about AI amplifies data issues. Your your AI model or your machine learning models is only as good as your data. I’m sure a lot of you might have heard garbage and garbage out, right? The data is not good. You you make you know a model that gives you a biased output or or not the right output. Right, it can have unfair, you know, outcomes right based on if based on the amount of data that you’re training on or or the bias data that you train that with, right? And if you have impure data, the output could completely be wrong, right? So. In talking about risk and compliance, you could also have sensitive data. Right in your system. And and mostly personal data is never used or actually train an AI system. But like you know you you should know like you know if on a metro data landscape that do I have sensitive data. What am I using to train my model right. You should have those. Guardrails to prevent any misuse and protect personal data as well and comply with the laws. OK. So if you do all that, then we build the trust, right? Trust with not just the users. But with the world out there as well, right? So. That gives transparency into the AI how it uses data and make decisions and and how do we do that? That governance, like you know, it’s through documentation, lineage tracking and oversight, and we’ll get into that a little bit little bit more in deeper and how Azure approaches that and thereby. You drive business value, right? Like you can discover data silos in your data estate. You can streamline access to high quality data, speed up AI development, right? So. Modern data governance actually enables us to confidently ensure AI readiness, right? So. Speaking of which. Let’s talk about key pillars of the AI and data governance, right? So earlier it used to be called data governance. Now AI is so much a thing. And AI and data is so close together so you know. Most of what? AI governance is derived from data governance because those go go hand in hand, in fact. So the key pillars of governance is having unified data catalog and and discovery. So what does that mean? Means like you know, you have a inventory of all data assets. Data leaks your databases and so on and so forth, right? And and to be talking specifically about Microsoft tools like we’ll get into tools in a detailed fashion, but like Azure purview is something that let’s you scan all your data assets across your landscape. And then let’s say that you know, hey, these are the different landscapes that you. Data assets that you have and you have this kind of data over here you have. You know that data over there and so on and so forth. We’ll get into a specific examples. And there’s also governance key one another key pillar is like data classification. We want to classify what is a publicly categorized data, what is a personal data, what is a sensitive data and so on and so forth, not only just for AI usage. Or train AI also for a data govern data systems perspective, right? We don’t want users accessing. Confidential data or personal data, unless they have to, right? So there has to be control. For example, there could be personally identifiable data, financial data, et cetera. And also as the core topic of this is like we want to make sure that the AI is only trained on the data that it actually needs. Right, not only personal sensitive data and we need to have quality data, quality and lineage tracking. So let’s say for example. While building a while projecting a dashboard or a report where somebody is looking at some metrics. We wanna see like hey. How did we end up calculating this particular sales numbers? Right, we’re all, which all layers. Did it get transformed through what all is undergoing to that particular field, right? There’s all this also applies for a feature engineering. So when I talk about feature engineering, what that means is like, you know you. Actually source into the model. Hey, you need to use this particular dimension or you need to use this particular you know field. To train your model, for example, it could be like something like a survey or or demographics information and so on and so forth, right? So so you need you wanna understand looking at a particular field, what all transformation did it go through both from a lineage perspective to understand hey, where is this coming from? What all did it go through? Also, if you go upstream all the way because data flow is like through different pipelines and finally it ends into its pristine form, right? So on the upstream perspective, if I were to make some change over here, where all is it going to get affected, I would want to know, right? So that kind of. You should have that hold on your data, right? And then come it comes to access control and security, right? You don’t want to, you know. Like. Everybody to have all the access across your landscape, you would have roles and permissions for certain set of users or you could have regional based split for users you know you would have, you know managers looking at data from North America versus you would have certain set of. Users looking at European Union data, they’re only responsible for that for that section. So you could have those kind of access controls or you could have privileged space access control as well. Somebody can see the sensitive data. Because their job roles requires, somebody does not have to sees that that right because they don’t need it. And then it comes to lifecycle management and data minimization. So what that means is like adhering to certain, you know, compliances, right. For example, you might have heard of GDPR compliance and so on and so forth, right? Like you know, hey, I should retain only minimal amount of data. You know, I should not have any additional data that is not. Needed for my day-to-day business or reporting purpose. Right I should. Purge the data that is not needed in my system. I should mask the the personally identifiable data. Also something like GDPR has compliance has something like a right to forget if if a customer wants to, you know, want their data to be deleted from a certain system they can request for that and that should be deleted. Right. And on top of this, monitoring all this right monitoring like. How? Who all is accessing the data at different levels? How is that data being used? What are the queries that are throwing like? You know where all are they actually extracting the data into. And so on and so forth. So these are the key pillars. Or best practices of data governance. Moving on. Now let’s talk about the technology, right, like. What is that technology landscape which helps us? Achieve those key pillars of data governance right, which is becoming more and more important. So, per view, we cannot speak data governance without speaking per view in Azure. Microsoft’s purview. I would say it’s a cornerstone of data governance. Right. It’s a unified data governance service that helps us discover, catalog, classify and manage data across the entire data estate. Now, a typical data estate. At least until 2-3 years ago. Even in Azure cloud would mean that you would use multiple services like a service to actually pipeline your data like your data engineering pipeline Alex that’s called Data Factory. By the way, you would have like your operational storage of your data in a relational fashion, something like a. SQL databases or you could have file based storages too. Like you know, you could store image data. You could store other forms of data. You could have semi structured data that is coming from your IoT devices and so on and so forth in a file based storage. So depending on your needs, there are different types of services that is available in in Azure, right Azure Cloud. So you end up having different services that you you use and then you know the data pipeline services which helps move data from one to another by transforming, right, so. If you have all of these, how do you have one common Birds Eye view of what all is out there? Right. So that’s what we used to provide, right. And Parve can actually scan all those data assets and say that you have this particular Azure SQL database and these are the tables that is there. These are the fields that is there and this field is getting used and so on and so forth, right? So and you can also catalog your data right like you want. Might want a catalog one set of data or set of KP is or key performance indicators as your or your sales and revenue data. Or you could. You might wanna categorize one set of files as your customer feedback, right? Once you catalog that. It’s very easy to go in and say I want to find everything about my customer feedback because I’m building a model, you know to do sentiment analysis for example. Right, so, so that helps tremendously when we get log the data. Then there is classification and sensitivity labels as well, right? Like. You can say that. You know what this is? API data. This is a personal data. This is sensitive and then have rules around that. Like you know, for access C accesses and policies and all right. I already mentioned about data lineage. The concept around that. So purview helps with that. And another advantage is that you know you can actually. Create a business glossary, right? Which is definitely required as and when your data estate is maturing, right? Like oh, now you have. You say you have a data warehouse, but I don’t know what is. What is that you have in your data warehouse? I have sales data OK what all do you have in your sales data? Well, I have sales and revenue, OK. So does that mean you have sales margin, your cost? Yes, so this to award is back and forth if somebody was were to create that business glossary. It it is very time consuming if you actually manually do it. But this classification and labeling of this data and cataloging this data helps users you know, create a glossary out of per view and. And say that you know this is what I have. You know, you can go use it as you want based on your access, right? But we also provides multi cloud hybrid support. It does not just strictly on app, you know it does not only strictly look at Azure. So if there are. Your data estate and other cloud like Awss 3 for example. Or if you have on Prem databases. So we can scan that too. So a one stop shop for data governance. I would say moving on. Yeah, this is some sneak peek into the connectors, right? You can see on the right hand side there is AWS connectivity within Azure itself you have like you know Azure Synapse analytics, Azure Cosmos DB, MySQL databases, BLOB storages, Azure Data Lake Storage N2. You know you can connect to an AWS account or AWSS 3 RDS St. A version of PostgreSQL storages for Amazon. And so on and so forth. Right. So it provides multi cloud support. Apologies for the resolution here, I don’t think it’s that great, but you get the point, right? So. And for scanning, right? Like there are built in classification rules within Azure, right? You wanna say you wanna categorize your data based on government classification rules or finance classification rules, right? Let’s pick the example of finance. I have sales and margin data. That is, I’ve I’ve populated into the data warehouse and I wanna use purview’s built in classification to actually. You know, use that. You know, categorize my data. Right. So look into my data estate and say that, you know this is your sales number and this is your customer date. I mean, sorry not customer name. You know, customer first name, last name, e-mail address. You know it can. I can look it to that that perspective. You know, this is your cost data. This is your margin. These are the products that was sold right? Like we can actually look at that data including the data type and and and tag it accordingly with the built in classification rules. But we also allows for you to create custom rules, right? Sometimes what is there in the standard classification rules might not be tailored to. The business needs that we have so we can create custom rules to scan the data and classify that. This is a screenshot into data lineage, right? I spoke about data lineage so so. I am looking at something in my data warehouse and I don’t know how did they derive this number right. Like you know, although I believe the number that I see, right? Like all the number of customers or the number of products that I see, I don’t know how. Where all of this is getting sourced from, so from from. You know, this typically goes from left to right, you know for. For those of who are relatively new, right? Like you know, so on the right hand side is my final object that I would be looking at and I wanna see even where is it coming from, right? Like in my report, where is it being get used and all and all right? Like what? All transformation is it going? So this is a little bit of. You know a sneak peek into that. This is a view within purview itself. So that I understand like this applies for feature engineering as well like you know did we did my data go through some standardization or you know did it go through some data quality like data imputing, treating nulls and so on and so forth is it is this partic? Field something that I can use for. Excuse me. My model right? Alternatively, I can also look at like if I happen to change something in the source. How a report would get impacted where all is it getting used? Who all is using that right? Like we saw, you know, tables are getting fed to which all reports are getting fed to. I have that kind of a control. You know what kind of machine learning? You know pipelines. Is it going to that kind of a control within purview? Now getting into a excuse me, so getting into. Azure Data Lake Governance and Microsoft Fabric. It’s very important to talk about fabric here because fabric is a strategic direction of Microsoft. For their data warehouse, you know or or their data estate approach engine like in say so. Traditionally if you look at the first bullet points, you would have like you know Azure data look data leak storages. Like adls relation databases, Azure SQL and several different services, right? And and somebody would go with the zone based or medallion architecture. We’ll discuss about what Medallion architecture is in a future slide. But you know, to my earlier point, right? Connecting back to my earlier point, they would have different services and they would have to use something like purview to have this holistic look. Right. Like I have this many objects in my data estate. I know that you know where my sales data is. You know, customer feedback data is they need to have per view per view scan the entire data estate to to see all that, but with fabric. Data governance is not an afterthought. Right it it data governance approach gets baked into fabric, right? Fabric comes with a concept of 1 lake. So what that means is everything that you have within your data estate is present in your one lake, so you don’t have to go to a different service to access that. You have data that is stored in your relation database. It is there in one lake. You have data that is stored like a files that is stored that is there in one lake. You have your pipelines well and good is there in your one lake. Right. You have your machine learning pipelines still there, right? You don’t have to go anywhere else and fabrics lets you actually manage your workspaces so that you know you don’t have to worry about. Oh, now I have everything like how do I manage? It’s totally up to you on how you want to manage, right? Like you want to have workspace based on finance or HR or you want a workspace you want to have workspace based on your development slash production or you want to have workspace based on your regional categorization of your data. It’s up to you, you author it. You decide. You design. Yeah, it’s, it’s it’s nimble enough so that you know you can be creative with that. And and it comes with built in per view integration, so with fabric. Data governance, which is getting a lot more significance and importance as AI and the data that is backing the AI, is getting more lot more significant importance. With fabric governance is not an afterthought. Comes baked in with fabric. So let’s look at that. OK. This is I just want to touch base a little bit about you know what all fabric you know has. I had mentioned about data Factory data warehouse. It used to be called a synapse data warehouse. They’re not. They have not renamed it. We’re just using synapse data warehouse within that. So and then your data engineering pipelines. For I I hope data warehouse is is a very common and understood term. With all of the audience over here, but for the people. You know who are relatively new. Data warehouses the final. Pristine data destination for a performant you know, and I’ll and I’ll and well of your data. Excuse me, analysis of your data. You know, you build your reports or dashboards from is a data warehouse. So and then you have data engineering pipelines, right? Your your data science machine learning pipelines, right? Your real time analytics. You have power BI for visualizations and data activator activators in the mix, right? If you wanna do something based on something that happened with your data, like for example I have some new customers that have come to come into the system. They do not have their hierarchy associated with that. They do not have their parent customer associated with that. I wanna alert my data stewards or list of users that I have certain new customers that need to be acted upon so that they can map. To their respective parents. So I can write a data activator rule to actually you know, alert the users. So all of these services, which were actually different services before, comes under one ecosystem of fabric. And all of this can be available in one lake. So these are several components of that. So I’m not gonna go through that in detail, but this is what it is and this is a sneak peek from our own. Developing landscape that I. Just took a screenshot of. For sanity purpose and I’ve just masked the customer’s name, but you can see like in A1 lake. If you look at a one lake you can see if you look at a type over here you can see a mirror database. You can see a warehouse you. Can see a SQL analytics endpoint. All of these have different purpose, right? For example, Lake House, as you see over here. That is something that you can store the data you know you want to store a structured data which has. Fixed columns and you exactly know what is there in every column like in a spreadsheet you can store that in a lake house. You don’t know the structure of the data that is coming in. You call it a semi structured mostly happens in IoT devices and stuff. You know you can store that in a Lake house warehouse. I already explained what a warehouse is. SQL Analytics endpoint which will help you. You know do your for. Analysts who are SQL savvy can, you know, do their SQL based analytics. On on data that is stored in files. Semantic model is something that is created underneath of power BI so that you know it caters to power BI for all the visualizations like for example if there’s a pie chart or something like. Like a graph. Like a bar chart or something? Like what? What is that dimensions or KP is? And how are they actually related that semantic models actually define that so you can see everything which was earlier different services comes under one catalog, right? So. That’s a sneak peek. Again, this is again another sneak peek into fabric lineage, right? We did not have to turn on per view, you know, in this particular project because we were actually working on fabric. And as I said. Purview is baked into fabric. Now I know that like I have a database over here which is getting mirrored without without the moment of data into fabric. I have this over here and that is getting used in multiple places. You can see this small arrows right. Like I’m using that in several data engineering pipelines to actually massage that data and and and feed it to further points, right? So that gives me like where all is my data being used. If I make a small change over here I know where all it’ll get extract. You know where all it’ll get impacted right. If I make a small change over here, I can look at the linkage of this specific item and. See like oh, this could break my next pipeline or this could break my next model or my report so. Again, this is my point that like. Data governance is not an afterthought in fabric. We in my first slide I mentioned about, we’ll talk about certain architectural patterns, right. And I briefly mentioned about Medallion architecture. Typically in a data warehousing or data estate scenario, we go with a bronze, silver, gold pattern. For movement of data so we can see look at it from a left to right perspective right? Like that’s how the data gets. Promoted from one layer to another. It moves forward, right? So typically this architectural pattern is. A widely followed architectural pattern, although the name got name was coined maybe three or four years ago. But the pattern has been followed for several years. Right. Like you know, for years in fact, I would say ever since the data warehouse concept has been there. So bronze is like the the first layer of landing your data as is from your source systems. You could have different source systems as you can see on the left hand side there’s files, there are databases, there are sensors, there are business apps. You land in exactly as is. Don’t worry about. Like any cleanup, don’t worry about like, you know what you’re extracting, how you’re, you know, not not how. But what? You’re extracting. Whether you wanna transform that data or not, right? And silver is like, as you know, you can. The name implies, right? Like the next best or better, you know, medallion layer, right? Like OK, now the data gets into a much more refined or quality control data, right? You take care of any duplicates or you take care of any nulls. You make sure that the data integrity. Is there right? That’s a integrity is an arguable standpoint, but still again, like you know you you clean up the data a little bit as it progresses from silver and that when the data moves to gold it’s your pristine level you know for storing the data for analytics there the data. Should be in in 100% purity for not just your. Reports to get used, but also to use for. You know for your. Models and and so on and so forth. Or you can. You can arguably say that like I can use the data from from silver itself, but the the pattern here of medallion the concept of medallion is that you improve the data quality as you move forward. You know, you make sure that the data is has integrity between them and I say integrity like you know, hey I am I representing any any revenue generation or a sales generation without saying like which customer. Contributed to that right. Which product did I sell to get that sales number and what was the quantity? So in order to qualify certain numbers, I had need to have integrity to the different data aspects, right? So that referential integrity enforcement and all happens at the data moves forward. So typically this is a a recommended architectural pattern so that you don’t have. Silos of your data processing, storage. Right. You don’t have disparate. Ways. So Speaking of which, I’ll quickly move on to Azure policy. So what is Azure policy mean right? Like. It’s a tool for enforcing governance right across Azure Resources, right? So well purviews for data. I mean governing your data content, Azure policy actually governs a cloud infrastructure or configuration, right? So you can create a Azure policy to say that you know what. I don’t wanna, you know, any storage which does not have encryption enabled. That could be a compliance requirement for your system in a Microsoft by you know, by standard it encrypts the data at rest and in transit, by the way. And and you know I only wanna deploy my Azure open AI service in in a particular region, right? So those could be your Azure policies, right? What are the use cases for that? Right. You can make sure that a particular AI service that you’re deploying for a particular customer base in a specific region can. Can remain in that particular region when you actually create services within Azure cloud, you can say that it it needs to be in this particular region and if you have that policy running and if you create it otherwise, that policy will not let you create it. You can. So require tagging right? Like you know, hey, I am the owner of this AAI service that I’m creating. I know that I’ll delete it after a while or like you know, this is for this particular purpose. This is for this particular client, right? I can actually ensure that these these tags are there or I can write Paul guidelines that these need to be there. If not, let it fail the validation while it is creating right? It also let’s. Have Azure blue blueprints or have infrastructure as a code right means like once you have all these policies written in some landscape, you can move it as a code. You don’t have to write it again into your next landscape like your test or your production or if you. Creating a new subscription altogether. You know you don’t have to do that from the scratch, right? It also helps us to. You know. Monitor this continuously, right? Like is there any other object that is not? Adhering to this policy right can have dashboards like you know and say that like you know what 75% of your objects are adhering to these policies. And and 25 not. And these are the objects you know these these fail the policy that, you know some of your rules are not. You know, honored in that objects, right, so. Let me see if I yeah. So I for the sake of it, I tried creating. An object in our landscape without, you know, following the Azure policy requirements so that I could show you. Right. So you can see on the left hand side that you know what it enforcing the primary owner TAG compliance has failed. So this is what happens when I create a resource within Azure. Right, enforcing the client TAG Compliance has failed. Enforcing the expected delete tag has failed and and there is on the right hand side there is a written explanation. Beautiful explanation by copilot, right. How beautifully a generative AI works to explain you that. You know what? This is why your creation is failing. This is the bash command that is written inside which is failing it. This is a partial command, right? And what you should do, right? It’s it’s giving you a descriptive, you know. Information in into why it failed. So this is. This is because there is a active Azure policy there. I hope that makes sense. This is based. This is for the infrastructure. No. Let’s talk a little bit about responsible AI. You know. So Azure actually hosts open AI, right? A lot of people have questions like, hey, what if I use Azure hosted open AI, right? Does that does my data get sent to the public Internet or open AI? Open AI itself. The answer is no. Microsoft Azure Environment keeps that data within and it does not send to the public Internet or or it does. It is not used for training open AI models at all. Your data remains your data. So you can, if you actually use Azure open AI service. It’s not. It’s not gonna be sent anywhere else. It’s gonna be within your your subscription, your cloud. It’s not gonna be available to anyone else. So. So yeah, like something like GDPR. You know which I mentioned about, you know, earlier these kind of complaints can be adhered to if you, you know if you know this right and there is like also another aspect of is like you might want to have content FIL. Right. You might want to avoid any hate speech. Or you might want to make sure that there is no profanity or. Anything undesirable content as an output of your model. Right there are constant filters available within Azure so that you can. Tweak the knobs of different sensitivity levels and and let the model adhere to that right. Yeah, and. And talking about controlling the A model usage right there is AAI Foundry model catalog by the way. Like you know, if you look at open AI, there are different models that could be used with an open AI. There is photo, you know, 03 model 04, mini 04, mini high. All of this like you know, all of these models have different use cases. So if you want to be, if you want to be, if you are researching on something you would rather use. One model if you’re writing codes and asking feedback about the code, you would rather. Use another model so you can actually decide what model should your organization use based on the use case. It could be based on the use case or it could be based on the. Performance of that performance metrics of that model. Or it could be based on how recent and how wet it out that model is, right? So that is there, you can have a controlled AM model usage as well. This is an example of content safety studio, right? Like you know, if you get into a content safety studio, you can see that there is a moderate text content section, right? I’m gonna quickly show you that example. A screenshot of that right. So there is a safe content example, right? Multiple risk categories. Example, if I were to read that out of 51 year old man was found dead in his car, you know, blood stains and so on and so forth, which probably you might not want to output as a part of your model, right? Like so, you can tweak those. You know the violence level, the self harm levels. Hatred, sexual level like, you know, all of that into different threshold whether you want to allow it or block it or you know keep it medium and then test your output and then. Let your model run accordingly so that that means you’re like, you know, actually monitoring and controlling. You know the output of that, so this is an. This is from. A snapshot from the window itself. Yeah. And I spoke about like in a different models that could be used, right. So you can see that this is. There are different GPT models like O3, Mini, 4.5. There are even deep seek models, right? And and and you can decide in your foundry. You know what are the models that you would let your organization? Use so that is possible. Moving on. What are the common data governance challenges, right? So we discuss about. The excuse me? Yeah. We discuss about the silos, right? Like if the data is this you know in several different places you don’t have no control, you don’t know if you have valuable data assets in some places which you could use very well for your model or your reports, or if you you would have sensitive data here. And there you don’t know that like you know. So you like in the sense like. These you know you you need to have a proper data governance to actually get a hold of all this, right? And and ensuring data quality and consistency. We already covered that, right? Like you want to make sure that all your data is pristine towards the usage. The final usage of it, but we at hysterical reporting analytics be it for predictive. Or AI model research usage so on and so forth. You know, privacy and compliance concerns. Not just for adhering to compliances, but also for gaining trust right bias or ethical issues. You know, if you’re training a model based on bias data, it’s gonna, you know, output biased. You know outputs, right? And and if you have your data estate across multiple multi clouds or all different. On Prem verses cloud, you know that that’s a problem, right? So. You know, and and then there are cultural and organizational hurdles. Mostly these teams like. AI teams and on mission learning teams want to move quickly. So. If there are a lot of garners practices, you know that could, there is a perception that it could slow them down. Actually, if you have proper garments, it only eases all of this because you know you know where your pristine data is. You don’t have control about where your data is coming from. You have trust on your data and you can project that trust to your clients and also to the. Compliance audits and so on and so forth. A use case that I wanted to just bring in is like, you know, we have a global manufacturing company specializing in. Motion activation control technologies and they’re using Azure fabric to build analytics sourcing data from waiting ERP systems. This company is not a single company, it’s umbrella optics. Different companies, right? So they are actually creating an end to end analytics where their process is not merged together like you know when I say this actually in actuality this is like striking data from 12 different companies, right. And and forming a proper medallion architectural pattern to one common structure and using fabric. And thereby with fabric we have absolute control about where the data is coming from in which from which source and how is a particular number being formed and so on and so forth. So we are building the analytics layer and and this will be very you know since since governance is baked into the solution which is fabric, it’ll very well be used and and very. It’ll provide a lot of advantage going into the predictive realm generative elements on and so forth, even though they have. Not merged the. You know processes and ERP systems across these different companies. All right. So some evolving regulations. So GDPR general Data protection regulation. I already covered that a little bit. Like you know there is right to forget if a customer wants to delete their data from the system, they can. I mean they can request and we are bound to do that and minimal data retention in our system. These are all some examples of GDPR and European Union AI act is forthcoming. So. We need to. Keep an eye out for that. If your business is in the European Union. You know, if again this first four bullet points are like a international, you know regulations that are out there and some of them are forthcoming, right on, for example, personal data protection law by Kingdom of Saudi Arabia and Qatar cloud computing regulation law, right. So if you are doing business internationally, you need to adhere to all of this and and there is absolutely no way other way around like you know. Other than having proper data governance on your system, right? Similarly, California Consumer Protect Privacy Act that is already an enacted law. If you’re doing business in the state of California and Wisconsin, Data Privacy Act, that is not enacted yet. There was a news that it would be enacted in as of January 1st, 2025, but that has not been enacted. But I wouldn’t be surprised if that is enacted in the future. So for the businesses who are doing business locally in Wisconsin, that’s something that to be aware of. Yeah. So Long story short. You know. I think we are kind of wrapping up so. Yeah, like you know. Microsoft purview, you know, there is there is this concept of compliance manager which will help you you know us. All this, some of the standard regulations like GDPR, ISO 27001 and all. There are compliance managers to actually run rules based on that within purview. So that’s that’s like all of these standard compliances are there out there in Microsoft itself? So you can actually run your. You know, activate those particular rules and run. You know your compliance checks against your data estate in. Provides end to end data lineage you know and and you can have double encryption of your data instead of your application manage keys you can have a customer manage key. So that’s an added security that the customer holds the keys and and and you can you know have that and then identity and accountability means like you can. Only give the required people a right level of access, and Microsoft always continuously updates their certifications, right? So whether it is GDPR or some new law that comes, Microsoft always continuously keeps updating. So if you’re using Microsoft tools like Power view or Fabric, you are safe because that they get certified and those new standard tools get will get get added to the policies that is well within Microsoft. Yeah. So and also provides per view data sharing, right? You don’t have to let users extract the data from your system. You can actually. Open up the govern data if you want to share to your clients so that that is also there, right? So and and if you are on top of that governance, you can actually project that trust and transparency to the users who would use that data. So definitely, definitely a tool set of tools to go after. Let’s see what is next. Yeah, like you know, for anybody who’s new or relatively new into gardens, I would recommend the next steps to be performing your data estate assessment. Identify what the quick wins are. I covered a lot. You know you cannot boil the whole ocean, right? Like you know, you have to see. What are your quick wins and what’s your priority, you know, to build a road map of your data governance. Data governance is not an overnight thing. It’s a maturity. Step by step maturity practice that you do one step at a time. Right. And then stay informed on on the new governance and regulations. So that you can ensure that you adhere to it. Or or have a have that in your road map. Yeah. So. That’s. A slide you can engage with concurrency experts to ensure that your data is secure, structured and AI ready. And and we can engage with the leadership team on the priorities to define a strategic road map and and leverage up to you know. Semi final for funding from Microsoft. In fact, if you’re going with Microsoft Tools. That’s all everybody I know. Just we just have 3 minutes. You know, I don’t know if I can cover all the questions, but like, you know, I’m definitely gonna get back to questions if I’m not able to cover them. Are there any questions that I could help answer in this next 3 minutes? Your hand. I couldn’t see because I had teams in a minimized fashion so that I was I could share the screen. But. Paige Wamser 56:46 I didn’t see any hands raised, so I think we are good and if anyone has questions they can fill out the survey and we can get back to you. Suneer Mehmood 56:55 Awesome. Awesome. I thank you so much everyone for patiently listening in. Hopefully that helped a little bit to understand about governance and the practices. And looking forward to see you all in the next session. Paige Wamser 57:10 Thank you so much.