🦄 Making great presentations more accessible.
This project aims to enhances multilingual accessibility and discoverability while maintaining the integrity of original content. Detailed transcriptions and keyframes preserve the nuances and technical insights that make each session compelling.
Overview
📖 AWS re:Invent 2025 - Strategy First: Empowering the Business to Lead in the AI Era (AIM348)
In this video, Adrian Estala from Starburst and Raja Palaniswamy from MUFG discuss empowering business teams to drive AI innovation through a three-layer data strategy: federation to access data where it sits, a business-facing semantic layer organizing data into logical products, and adoption in AI solutions. MUFG implemented core and functional data products, enabling business data producers to self-serve without IT dependency, reducing time to market. Raja emphasizes "data needs AI and AI needs data"—using AI for data management tasks like profiling and lineage, while feeding quality data into AI models. The approach allows business teams to experiment with agents and chat with data through Amazon Bedrock, with proper governance and authentication, ultimately enabling ten AI projects in ten weeks versus one in eighteen months.
; This article is entirely auto-generated while preserving the original presentation content as much as possible. Please note that there may be typos or inaccuracies.
Main Part
Building the Foundation: Federation, Semantic Layers, and Business Empowerment in AI
I'm Adrian Estala, the Field Chief Data & AI Officer at Starburst. Before joining Starburst, I held a similar role, so when I look at these types of discussions, I approach them from an ecosystem perspective rather than just a Starburst perspective. I hope to bring you a broader conversation about strategy that includes Starburst. It's great to meet everyone here.
Raja, why don't you tell us about yourself? I am Raja Palaniswamy. I work at MUFG, where I manage the enterprise data and drive the data strategy for the team right now. Fantastic. I had an opportunity before we got started to walk around the room and get a pulse for who's here and what the interests are. I can see we have some chief product officers in charge of product, and we have teams that are really interested in driving adoption on the business side and pushing business self-service. Those are two great topics.
What we want to do today in this conversation is bring to light the importance of letting your business teams feel empowered when they're driving AI solutions. When I say AI, I'm talking more on the GenAI side, the Agentic side, and conversational AI. I'm not talking about machine learning, and I'm not going to let my business teams drive ML ops. However, when we get business teams more involved in being creative and owning the art of the possible with our AI capabilities, magic happens.
A lot of the customers and business teams I work with directly are getting excited and building solutions together. If you sat in on any of the cloud sessions this morning, it's amazing. The ability to create, build, and understand how skills come together to solve a solution—these are all things that our business teams should be able to own and drive on their own. But there's a gap. There's a gap where some companies struggle to drive that empowerment, and that gap really comes down to understanding your data and being able to access your data.
I draw this up in three simple blocks, and Raja is down the journey over at MUFG. What I want to do here is set the foundation for the talk track in three simple layers, and then I want to dig into Raja's experience and what they're doing at MUFG and the work they're doing to drive business-driven AI innovation. Let's start with the foundation first. The first part of any foundation is the challenge of getting the data. What we do is allow our teams to get data where it sits. I don't care if it's in the cloud or on-premises. I prefer it to be in the lake, but it isn't always in a lake.
Federation for a long time is what made Starburst famous. That's what we've been good at for a long time, and that's great. But if I can get to my data where it sits without having to migrate it, whether it's in the cloud or on-premises, this next block is what's most exciting. You're hearing a lot of buzz around the semantic layer. It's this business-facing semantic layer that's the real game changer. If I can access my data where it sits without migrating it and I can start to organize that data into logical data products that my business teams can understand, I told Raja this once before: we want to create an experience where when the business teams walk in the room, they recognize their data. It's like walking into a family. You know your cousins, you know your family. You recognize your data. It's easy for you to understand.
When you can create these logical data products, I don't care where the data sits. I can organize it into packages that my business can understand, and it's a real game changer. It's a true semantic layer where you're abstracting the legacy complexity and giving the business team something they can use. When you've done that, business teams feel a lot more confidence when they enter the room and say, "I want to test out a new agent in Amazon or any system. I just need to get the data."
If they can ideate very quickly because they can say, "I need a little bit of this, a little bit of this, a little bit of this. I want to bring these governed, curated, secure data products into my agent and use them to do an analytical exercise," they can do that because the data is already there and they have the access in terms of security and availability that they need. Those are the three layers. If you want to think about what's the bridge to get to business enablement for AI, it's getting access to the data through Federation. It's creating a business-facing semantic layer that they understand. In many cases, I build these with the business teams. Raja will tell you we've been building this with his business teams. The third layer is adopting these data products and using them directly in BI or AI solutions. That's the framework that I like to talk about. You don't want to hear it from me. You want to hear it from Raja.
So I apologize for the long intro, Raja, but I've set the stage. Before I jump into some questions, is there anything you would correct or anything you would add to what I just talked about in that foundation?
MUFG's Journey: From Physical Consolidation to Business-Driven Data Products
I would say data needs AI and AI needs data. It's a cycle. I think we started the journey with the traditional data strategy where we would consolidate everything into physical consolidation, move the data, curate the data, profile the data, and then do things accordingly. We pivoted that and then said, "Okay, can we create data products in a logical way?" And then those data products, or what Adrian talked about, I will add is the semantic layer is nothing but from a business product lens. In a corporate world, if you're from corporate banking, you have lending, like we have term loans and several of those business products. We aligned our business products and data products to be identical, so when a business user comes in, there are business data producers who create these curated data sets from the semantic layer so that they can create. If, let's say we have commercial loan operations, that team wants to see the data quality of the loan booking. They are able to really curate the data from the core products and then create a functional data model or a data product that is specific for them.
We gave the business users three personas to Adrian's point: business data producers. We call them specifically business data producers, not IT or a CDO. The data elites are in with the business, so they know and speak the same language. They can understand the data, so I would add that for us, it was more towards how do we enable business for self-sufficiency. They should not be coming to IT or data elites to understand what data is there, how it is related to the product, how do I get access to that data. Starburst provided us that capability to pull data from multiple data sources, whether it is a distributed environment, whether it is SaaS, whether it is PaaS, whether it is embedded systems like Salesforce or any of those. So we pulled that data in to create the curated data products. That enabled business for them to be self-service.
Was there a trigger that got you guys to rethink your data strategy? At what point did you guys say, "Look, we've got to do something different. We have to maybe move closer to the business. We have to drive more self-service. We have to find a way to get our business teams to trust more data." What was the trigger at MUFG to kind of force you to rethink your old data strategy?
I think time to market was the key. How do we enable the business? Because businesses are making money as of now, how do we increase our revenue generators' capability, who are sitting within the business? Time to market. How soon can we get the data for them to prepare the data for themselves rather than them coming to IT or CDO? How do we enable the business to prepare the data to be self-sufficient? That was the pivotal point that triggered us to be in this journey.
Let's give the audience an idea of what a semantic layer might look like, without giving away too many of your secrets. When I talk about a semantic layer that business teams can understand, within MUFG, we've got this notion of core data products and functional data products. You can think about the functional data products as derivatives of the core. Core data products are well curated. I got data from maybe a hundred sources and I've organized them into a small set of core data products. I've got one for customer, I've got one for channel, I've got one for banking, I've got one for all my different big things, and then functional, you can imagine having a lending team saying, "I need a little bit of that one, a little bit of that one, a little bit of that one, to make this data product," and they're building their own derivatives in their function to answer their specific functional questions or analytical needs. So that's a model that we're building at MUFG. Now, tell me about your consumer, your business teams on the function side. Do they feel empowered? Are they all jumping in, or how is that working?
It's a journey. We have mixed business users. Some are very excited to see the data. Some are like, "Okay, get me halfway through. I need to understand the data." So we mixed that up with data literacy programs so that we can educate what data is there and how they can access it. It's mixed, like fifty-fifty. I would say some of the users who are power users are very engaged.
Some of our power users say, "I have my data where I need it from fifty different systems. I'm happy," and they can sail on their own. They are self-sufficient. We're empowering business teams to organize data in the way that they need to. Now, let's go back to the third box, the AI box, because that's what we really want to talk about here. If I can build a data product that's curated, the secret is business metadata. It would be one thing if I'm just building data from technical metadata.
If I build a data product with business metadata at the column level, at the table level, at the role level, I have real business descriptions of that data product. When you put that in front of an LLM, you get very accurate analytics, you get very performant analytics. We can sit down with a business team, a lending team for example, and say, "Here's your lending data product. Go for it." You're not just giving them a solution that they're going to own forever. You're giving them a capability that they're going to evolve.
You want your business teams to be able to look at the capability and say, "Wait a minute, can it do this? Can it do this? What if I add this skill? What if I add that skill?" The magic happens when the business teams understand the data and they get the confidence to work with AI. They're not afraid to break something because in a data product, you're only exposing this data. That's it. Nothing else. In a locked-in system, there's really clear guidance, guardrails, and governance. You're giving them the guardrails, the governance, and the guidance to experiment.
At the end of the day, you're decreasing the time to experiment and you're lowering the cost to experiment. I always say I want to do ten AI projects in ten weeks versus one AI project in eighteen months. They're all going to succeed, but you can succeed on three and keep going forward and fail on seven. That's probably not the right failure that we all want, but that's the reality when you're experimenting. Not everything works.
Enabling Agentic AI: Data for AI and AI for Data in Practice
On the AI side, I know that we're just starting to build that foundation for those data products. How are you looking at the Agentic front end to this in the future? I think we are looking at the AI landscape in two forms. One is to make use of all the data management and enable the data management capabilities, whether it is profiling, whether it is lineage, or any of those. Second, we want our business users to start chatting with the data. We want our business users to say, "Okay, can you give me all the middle-market banking customers?"
We want to provide them that capability and then provide them that list. We also make sure that we integrate that with our authentication mechanism. We don't want middle-market banking customers to be exposed to those who are not authorized to do it. It's a mix of getting the chat, using AI to enable data products, get the lineage, get the profiling, get the cataloging, tagging, those kinds of things. That's part one. By having Starburst, what's happening is it's enabling us to go into one location rather than fifty different systems that we have to go across.
The other part is that as soon as people come to Starburst, they can click on it. We want to invoke the chatting with that data using the Agentic features that are available to us, using Amazon Bedrock and other features, where everything is in one box so that they can come to something called a real data marketplace. That's the objective of that.
There's a point that I wanted to underscore there because it's agility on two sides. If I have a semantic layer that I built with my business teams, and the semantic layer logically doesn't care where the data comes from, you can leave it there. You can connect to it. It's also the agility on the front of that semantic layer, saying I can build an agent. If you look around here, there are so many cool AI or Agentic-type solutions. If I'm a CEO, I want to try them all, but if every one of them is attached to some core platform, it's expensive.
If every one of them requires me to model data a certain way, it's expensive. If I can deliver data products to any agent and I can tell my business teams I have the guardrails, the governance, and the guidance, I will deliver the data to you. Maybe you can build and try any agent you want on the front end. I don't care. Build any agent. I'm going to make sure you get the right data. Use any source. I'm going to make sure I connect to it in the right way. It's incredible flexibility and incredible agility and ultimately offers you longer-term opportunities to make better choices.
What I would add is that in this journey, business plays a pivotal role.
Business is in partnership with us to enhance data capabilities across the board. This comes through a trust-building scenario where we allow them to prepare data, and it's a centralized location where you have all the data that you can really curate and get to the next level. In fact, we improve our data quality indexes by pulling data through Starburst so that when it's fed into any AI for modeling, it has much richer data quality.
I like what you said earlier. You mentioned, "I use AI for data and I use data for AI." Can you explain that a little bit? When I say I use data for AI, it's to learn through the LLM models the entire rich content that we have in the organization. Whether it is structured data sitting in our deal file or unstructured data that is sitting in our deal file, or structured data that is in our systems, we can feed that data into AI to learn more about the insights.
The other side is where we need AI for data. We wanted to use AI for tagging purposes. We want to use AI for lineage creation. We want to use AI for profiling. Those are the insights that we can create. So data for AI and AI for data. I love it. Fantastic. I'm going to pause to see if I have any questions.
Thank you very much for your time. I really appreciate your interest. We have a booth just down the hall here at 550. I'm happy to talk some more. Thank you. Thank you.
; This article is entirely auto-generated using Amazon Bedrock.
Top comments (0)