Mercedes-Benz USA: Data-Driven Granular Decisions with Databricks Medallion Architecture

Summary

Mercedes-Benz USA, in partnership with EPAM, replaced 140 years of legacy mainframe systems with a unified Databricks lakehouse using Delta Live Tables and medallion architecture, achieving 60% cost savings and 50% team productivity gains while serving hundreds of business users.
A shift from ETL to ELT paradigm provided end-to-end data lineage, while a unified semantic layer and governance foundation enabled Genie agent deployment for natural language data access, consolidating 3,000 dashboards into 200.
Localized market intelligence capabilities now allow MBUSA to make granular decisions down to zip code level in the US premium automotive market, with a recommendation suite providing 'where, why, and what if' analytics across multiple user personas.

Mercedes-Benz USA: Data-Driven Granular Decisions with Databricks Medallion Architecture

Mercedes-Benz USA faced the challenge every legacy enterprise encounters: fragmented data systems, expensive infrastructure, and business users unable to access trusted data. In partnership with EPAM, MBUSA built a unified Databricks lakehouse using medallion architecture with Delta Live Tables, replacing 140 years of legacy mainframe systems. The result: 60 percent cost savings, 50 percent team productivity gains, and a platform that serves hundreds of business users.
Discover how MBUSA transformed from static dashboarding to natural language data access using Genie agent, consolidating 3000 dashboards into 200 while enabling granular, local market insights down to zip code level. Learn the three pillars of success: unified semantic layer, ELT paradigm shift for end-to-end lineage, and the critical role of governance as a foundation for agentic AI at scale.

Chapters

FAQs

What did Mercedes-Benz USA achieve by migrating to Databricks medallion architecture?

By building a unified lakehouse on Databricks with Delta Live Tables and a medallion architecture, MBUSA achieved 60% cost savings and 50% team productivity gains while serving hundreds of business users. The migration also enabled the consolidation of 3,000 dashboards into 200 and provided the governed data foundation needed for AI and agentic capabilities.

What is the ELT paradigm shift and why is it important for data lineage?

The ELT approach moves transformation logic into the data platform rather than pre-processing data before loading, enabling end-to-end data lineage that tracks data from source systems through all transformations to final analytics outputs. MBUSA adopted this shift as part of their Databricks migration, replacing a traditional ETL approach that made lineage difficult to trace and govern across a 140-year-old legacy environment.

How does the Genie agent enable natural language data access at MBUSA?

The Genie agent on Databricks allows business users to query the unified lakehouse using natural language questions rather than requiring SQL skills or reliance on a technical team to build custom dashboards. MBUSA deployed Genie on top of a unified semantic layer and governed data foundation, enabling multiple user personas to access data through a single guided experience.

How does MBUSA use granular localization for market decisions?

MBUSA built capabilities that allow market data intelligence to be analyzed down to zip code level across the US premium automotive market, enabling local market managers to make targeted decisions rather than relying on national averages. This video describes a recommendation suite providing 'where, why, and what if' analytics to support decision-making across sales, finance, and strategy functions.

Full transcript

[00:09] My name is Giuseppe. I come from Italy and I've been with the brand for 13 years. I work at the intersection between product management, market data intelligence and strategy, and data. But as you can see, I'm not alone today on stage. Therefore, let me introduce you my colleague Anil. Giuseppe, thank you. And everybody, I
[00:25] know it's the third day. It's going to be the last two days have been really grueling but you're going to leave this room quite happy with what you will hear and see because it's a very true story of what we've been able to achieve and what's coming in the future. So, my name is Anil. I'm close to 25 years with
[00:41] Mercedes-Benz, mainly focusing on digital data transformation across multiple markets and regions. And this is what we truly believe. Business strategy and technology execution works together and we've seen that happen and that's what we are here to share and uh
[00:58] feel free to ask questions towards the end of the session. We have reserved some time for that, but in the meanwhile, Giuseppe, let's start. Yes. Let me frame uh the journey for you today. So, Anil will start explaining the why, why the transformation was needed. Then
[01:14] he will move into the how, explaining the platform, the architecture, and the enterprise foundation. I will then take over and from there I will explain the localized insights solution that we built. And then finally, we will move towards the impact that the solution has been
[01:30] driving in our company, the lessons learned, and potential next steps. So, having said that, Anil, I would say let's jump into the contents. Perfect. Uh I would like to spend a little bit of time in order to set the context and if you look at this slide and if any of you
[01:47] are working with large enterprises, especially the legacy of 140 years that Mercedes has, there's quite a bit of complexities, right? This slide sometimes may be painfully familiar to some of you for some of the people who are working in startup companies. What's mainframe? What are you talking about?
[02:03] We don't even know that probably. If I really have to summarize and simplify, fragmented data, legacy architecture, challenges in moving data from one place to another, these are all leading to some of the challenges that
[02:19] business not being able to take the right decisions. And if you really want to bring ML, AI, and agentic AI, I hope all of you attended the last two days. There's really cool stuff that's coming. And we made a tough decision to say, "Let's really transform. Let's really
[02:34] modernize. And let's let's do what's right and ensure Giuseppe is able to take some of the decisions that he did." When you are a market analyst, you really want to rely on data. But sometimes you don't trust data, though it's central in a mainframe or
[02:49] wherever it is. And that's when shadow IT creeps up. That's when you start building workarounds. That's when you try to create your own space, your own frame, where you feel more comfortable. And that's where I would insist that creating the right data set, enabling
[03:05] the business to trust the data, working together is is one of the key aspects of modernization. If I move forward, this is what we have at Mercedes-Benz USA. We enabled enterprise data and AI platform at
[03:20] scale. And Giuseppe really represents the market local insights. So, these are the two areas that we are going to focus on today, but we are also the enterprise platform supports everything in the organization, which is parts, which is warranty, which is HR, finance,
[03:35] everybody. But the focus for today is really to talk about how we are able to enable local market insights. And I have to admit, when we started this journey, it was not easy. And it seemed like really challenging to say, "How can we even make this happen?" But
[03:51] I think a true partnership really made that happen. Correct. Right? Yes, definitely. That was the key. Yeah. And and what we are showing you today is live, it's in action, and it's a proof that you need to take the right decisions, you need to take the right
[04:08] architectural challenges, but you can make it happen, and we've seen it paying off. But keep in mind the foundation is really where everything gets enabled because data is at multiple places, data needs to be consolidated, and data needs to be accessible.
[04:25] Moving on, uh when you embark such large transformations, for sure many organizations rely on vendors. We did not rely on a vendor, we relied on a partner. That's where EPAM comes into picture because they got us the scaling effect, they were able to bring
[04:41] the necessary engineering skills, and we had the business mindset to make this happen. So, always believe in partnership, and that will take you a long way in terms of making things happen. This is the reality for us today. I would walk
[04:57] you from the left to the right, but if you really look at the left, that's where the legacy was. That's where the mainframe was. Fragmented data sitting at multiple places, data being accessed all over the place. Many of us would like to
[05:13] be on the extreme right. That was our initial ambition, but if I bring your attention to the middle and to focus the the middle part of the transformation looks very clean, simple, right? Many of us tend to fall off saying, "Okay, I think we hit the nail. Let's
[05:29] actually just step away and do what you want." And then he finds workarounds, he sees, "Oh, this is still not there." Uh moving to cloud is one part of the journey, but moving into a stack that makes business users talk to data is
[05:44] where the value is. That's what we did. One central foundation, one central governance, and making sure that everybody is able to seamlessly access the data in a unified platform is the key. Some of you may think, "Hey, why do you need to bring everything together?" There are
[06:00] different theories, different schools of thought. You could look at a data mesh architecture, you can look at a data fabric architecture. Irrespective of that, a unification of platform of platforms is where the true value comes. And what we've seen in the last 2 days in Data Bricks,
[06:17] I think they're really going to make Agent Tech happen. We are ahead in some parts of the game, and we want to really accelerate that further. This is nothing new, and I have to admit we did not try to reinvent the wheel, though we've been innovators in the
[06:33] automotive space for a long time. It's always important to do what's right, right? And Jussawalla, do you agree to that? Because I know you were pushing us for speed all the time, but I think speed is one part of it, but really doing what's right, what's important is
[06:49] the key to it. In a very high-level nutshell, because of time, and also because of how do we make things happen, we said, "Let's stick to the Medallion architecture. Let's stick to the Medallion framework, because this really worked for us." Bronze is where
[07:05] you really bring all the data from the source systems. It could be multiple source systems, and then you have the silver layer where you curate the data, massage the data, make data structured in in a way where business can use it. This is where we went with the domain-driven architecture approach. And
[07:22] gold, my dear friend Italian, loves gold, so we said, "Okay, let's build something that really makes it useful, that really makes it accessible." And we have been curating more than 30-plus sources of data in a simplified foundation, and today he's actually able
[07:37] to talk to data in English, to be really honest. We will get one day where we will where you can talk to it in Italian, but I think hopefully English is good for now. Yeah. And if you really look, Medallion DLT was one of the game changers for us because that's where the shift happened
[07:53] from the traditional ETL world on the mainframe to the ELT world. If I have to give you an analogy of it, there is apples at the source. In the mainframe world, we had watermelons at the destination. And how do you really bring the lineage? How do you really bring the traceability?
[08:10] When everything works, there's no questions. Nobody asks a question. Everything is fine. But when there are challenges and when you want to really walk backwards and say, "Hey, why is this number even looking like this?" That's when the big issue happens. So, we changed the game. We said, "Let's move from ETL to ELT approach and bring
[08:29] the source as is, curate all the logic, make it available in the gold." And that's where end-to-end lineage, simplified governance, and high observability within the platform gave us a huge lift in terms of what we really wanted to do.
[08:49] Speed is something everybody wants, especially when you're doing such a large-scale transformation. The CFO always asks the question, "Okay, I gave you all the dollars. Where is my platform? When is it available? How do you access it?" We would not be in a position to put the logical and the physical architecture to
[09:05] fit it in one slide. So, we really kind of simplified that to give you a view of how we are leveraging the data. Walking from the left all the way to the right, the data comes into the platform from internal sources, external sources, in batch, in streaming, in all possible
[09:23] ways, right? We have standardized our ingestion framework. We are Azure managed Databricks shop for Mercedes-Benz USA. So, we will see some of the components which are Azure, but that's where we are able to standardize the option for us. Once the data is coming in, we are
[09:40] processing and curating the data in the transformation layer where you have the bronze, silver, gold available. The publish is where the consumption happens. That's where the gold data products, the extended machine learning data products are all sitting in the marketplace.
[09:55] I'm sorry, the marketplace is blank. I think it's an error on the slide, but there are products that we've actually built where you can consume in a very simplified way. Analyze and engage is where our data scientists, machine learning experts, AI experts are leveraging the data. And BI is where we
[10:12] were in the Power BI world for a long time. We are still going to exist, but with AI BI talk to your data capabilities, we can really accelerate. And I want to really emphasize one thing. Uh most of us really forget governance and security when you really want to hit the
[10:29] speed, when you really want to bring the value to the business. And this is where we had a lot of difference of opinion and every day multiple arguments saying, "I don't care how you want to build the governance, but I want to make sure my data is accessible." But do you agree today the delay that we had really
[10:45] helped us to make sure that you are not bothering us on a day-to-day basis, you have everything visible, you trust the data, and you really love the data. Well, yes. I mean, it's it was still a delay, by the way. Yeah. Uh nevertheless, it helps us in being
[11:01] confident that we have a very robust data foundation to So, yes, I can confirm that. Absolutely. Especially in the world of agentic AI, the vulnerabilities are really at scale, and that's why we were really focusing on building the right foundation, not just for the new car
[11:17] sales, but for everybody else because the first product that you really build takes some time. But the acceleration really happens. The acceleration really comes when you start seeing the power of compounding. I I know how many of you believe Warren Buffett. I'm a big
[11:32] believer. He's sitting on 300 billion dollars of cash at this point of time because he's waiting for the right investment and the right compounding. That's what we did, too. We now are able to bring certain acceleration into new use cases and new technology, and that's what I'm going to talk about.
[11:49] This is where I get to brag a little bit before I hand over to Giuseppe because we have been able to bring close to 60% of savings in the current platform from the legacy world because of all the modernization initiatives that we took. Not just that, we are able to see now
[12:06] 50% of productivity gains from the team itself. They're able to bring data faster. They're able to curate it. The framework is fully automated. You have a new source identified. You can see value. Giuseppe, yesterday you were talking about a data. By the time you're
[12:22] back tomorrow, you will have it. Hopefully. Yeah. I'll check. Here's what I really want you to take away. The efficiency gains, and this is something we are speaking very honest because we we've been through this. The efficiency gains are really not the goal. You have to look at how you can
[12:38] fuel it because the speed, the acceleration, and the value is what you're going to get. In the legacy world, this was simply not possible. Even though we moved in an interim state to the cloud, the middle part that you saw, we didn't get complacent, but really wanted to
[12:54] accelerate and move on. And with the new framework, the new setup, we can make things happen. The platform made it possible. The real ROI is the story in the next few minutes from a business perspective in terms of how you can gain. Let me hand it over to Giuseppe. Yes. Thank you so much, Anil. Thank you
[13:10] for this technical overview. And um I'll take you now in the journey from the business side and what is the really localize inside solution looks like. But before that, let me let me give you an overview. I'm in I'm in US since 3 and 1/2 years now. I just moved from Europe, and I've
[13:27] learned many things. But, one the first thing that I've learned working in the automotive industry is that the US premium auto market is incredibly complex, even in stable conditions. You have millions of vehicles in operations,
[13:43] thousands of units sold every year by different OEMs, a huge variety of models and trims, a very powerful dealer network, and as well land which is very robust. So, we are talking about uh
[13:59] something which is very big. On the other side, you have microeconomical headwinds. So, you guys name it, tariffs, geopolitical situation, inflation, and so on. All of this permit us to work in an environment
[14:16] which is very extremely complex. And in this type of environment, the margin of errors is extremely small. So, in my humble opinion, this is what makes the look the localization make a process not a luxury, but a necessity for the
[14:31] business. So, let's go to the point, and I love this slide because it goes straight to the point. Please bear with me. We have one dealer, two zip codes, both zip codes 8 miles apart. On one side, on the left side, you see that you
[14:47] can conquest a luxury customer. On the other side, you can see that the competitive set, the customer profile and demographics, but also the product mix looks completely different. So, this means that for an OEM, the national message,
[15:04] the pricing, the inventory strategy, the conquest strategy, whatever, doesn't work when it comes to local level. And this is exactly the heart of the problem that we've tried to solve. The one size fits all breaks down at local level.
[15:21] So, our vision was to move from good field decision-making towards a strategic data-driven granular decisions. And this meant bringing together a multitude of data sets, new granularities, standardized mappings,
[15:36] cross domains all into one unique single environment. And it also meant going beyond reporting. We just didn't want to do simple reporting. We wanted to go more and and provide we wanted to give more and provide them more to the other users. So, we wanted basically to build
[15:52] an experience that was actually simple for users to navigate. Absolutely. In short, what we tried to do was to build one tool powered to a connected and robust data foundation. So, getting there was definitely not
[16:09] easy, definitely not a walk in the park. We had, as I mentioned, multiple systems, fact tables, levels of granularity. Our external our external external sources uh the sources didn't match with our internal sources. So, a lot of mapping, a lot of grouping,
[16:25] a lot of data engineering processes were uh were done. And um at the same time we had different personas to serve. So, it was not just a dashboarding problem that we had, but we had even analytical translation problem. And for
[16:41] me as a product owner and project lead, the main question was how do I take a very complex market, harmonize data, multiple users at multiple levels, and serve them with data and AI all at the same time?
[16:57] This was the main question. So, and this was something that we tried to achieve. On top of that, we also had the US user experience and as well the design a design challenge to solve. Let me show you now how we provided a kind of guided journey to our
[17:13] users. So, we started with cockpits, Sorry. That um able to identify areas of focus, where we should focus, what we should tackle. Then we give the possibility to different users to go into domain deep
[17:29] dives. Here is where it's possible to understand specific drivers, what caused that situation. Then for more data-heavy users, data specialists, we had our advanced analytics area. This is where our machine learning models are living
[17:45] basically. And then last but not least, because of the dynamic of the market and because of the way of working and running the business, we also had to build uh special use cases tab. A kind of on-demand self-service area
[18:00] where customers or users can go and take the data that they need whenever they need to to answer a search question. So, the same solution can serve multiple personas without forcing everyone to follow the same analytical path.
[18:15] But that's where we got to eventually, right? Yes, correct. And more concretely, this is how it looks like. And uh sorry for uh the blurry uh screenshots, but um I can't I can't share everything with you today. So, we built uh
[18:31] summary snapshots ready to use, visual geographical analytics, market market dynamics and key metrics, portfolio insights, distribution analytics, and also a kind of in-dealer in in in user inbuilt user guide, sorry,
[18:47] for um to provide to the users the answers that they need while they're navigating through the data and not to panic in a kind of way. But um yeah, this was the tool experience that designed It was designed to help users to move quickly from
[19:03] diagnosis to actionable insights. If I just interrupt you here, right? Uh how many iterations did we take to get there? More than a hundred? I I think we can hit the four digits, I guess. I don't know. I don't really But
[19:19] it was a lot a lot of conversation. And as you can imagine, when you try to build this type of solutions that um uh meant to serve multiple people, whenever you used to user whenever you talk to users, everybody wants to have it in his own way, right? So, I Could we have done in the old world?
[19:34] I I I don't think so. But I had to be, let's say, let me say very diplomatic in taking this the decision of what I put down there for the business users, what I put there for other users, and so on. Nevertheless, what we've done in on top of that, uh we built a recommendation suite. So, we
[19:51] didn't want to show only the data to the to the users. We wanted to provide also a space where recommendation was possible. And the recommendation suite was built on three main question questions. The where, the why, and the what if. The where tells you where to localize,
[20:07] where to focus. The why explain what is the pattern that you see in certain areas, and what is the problem the origin of the problem. And the what if allows the users to have to build scenarios and simulations, for
[20:23] or follow simply the recommendation that the tool is providing. And all of this was possible because under the hood, we had a very robust data foundation and machine learning models that are backing basically the the the information. Absolutely.
[20:39] So, we were not happy with that. We wanted to do more. And that's why we said Sorry, guys. Yeah. And that's why we said, why don't we let the users talk to the data as well? And
[20:56] that's why I introduced Genie, the assistant the agent. This was a breaking moment in the all project. Why that? Because we let the user users talk to the data and feel more confident. Use the agent in order to
[21:11] prove what they were saying and you know that what they were seeing on the tool and in order to understand better the data. This has been grammatically lowering the user entry point entry barrier. Also users that are not really data a friend, let me say, or they are
[21:26] not used to work with data on a daily base are now more confident confident to go in the tool and retrieve the information that they have. And in order to guarantee that information are robust and valid, we built controls for routings, permissions. We avoided that the agent
[21:44] was hallucinating, obviously. We put some guardrails because we don't want to let the users go outside of of of what is possible to be seen or interpret the data in a certain way. And last but not least, we've been also looking at the latency and the cost of it because
[22:00] we said we are building something, we don't want to have explosion or high exponential cost coming in. So, the speed the experience feels feels easy for the users, but again, under under the under under the experience or behind the
[22:16] experience there there's a very bold and robust data foundation that is driving now the impact. And now, talking about the impact, I'll pass it back to you, Anil, for an overview of the impact that we reached. Thank you, Giuseppe. Just another moment on this slide because introducing Genie
[22:32] was not easy and it happened at one of the conversations where we were working on something in a developer's PC. Giuseppe happened to see it is like, I want this and I want it now. How can we make this happen? That's where Databricks was helpful in bringing the speed and the acceleration, the way
[22:49] they're pushing new components, the way they're bringing new technology at scale. We are able to adapt. We have some of our colleagues here from Germany who really make that happen as well, one of our best buddies, and that's how partnership really really accelerates. And if I really have to brag about some
[23:05] numbers, it took us a little bit of approval and challenge from the calm department to put this out, but we want to give you something realistic that you can take back, something that you can believe. And this is what we've seen. Uh we were able to save close to a million dollars when we migrated out of the
[23:21] legacy infrastructure. And this in terms of the scale itself is huge. We are talking about terabytes of data. And in order to do that, we had to accelerate because we were on a very, very critical time path. That's where Agentic AI helped us. We built several
[23:38] migration agents, which gave us close to 70% acceleration so that we hit the speed. Though we were a little delayed, we made it happen because we covered all the bells and whistles in terms of security, in terms of governance, in terms of building the foundation right,
[23:54] and getting the value. The time itself can vary for different people depending on how skilled you are or how knowledgeable you are. That's where we gave a window of 10 to 100x because some of the users when they talk to data,
[24:09] they're able to get answers properly. A lot of prompt engineering techniques that we also put in the back end with the Genie space made talk to data really happen. We have hundreds of business users today who are talking to data. We cut down more than 3,000 plus dashboards
[24:24] that we have to 200 today. And our goal is to really see if we can minimize that as well and really just live in the new world of talk to data on your phone where you don't even need a PC. And 40% time to market in terms of speed is what
[24:40] we are seeing now with the new use cases that we are able to develop and bring the value. We just don't want to stop here like we said because my analogy of how I deal with in my personal life is whenever there's a new
[24:56] iPhone, I always try to get the latest from a electronics perspective because technology is changing every single day. And that's where we are working with Databricks very closely to bring in the latest and the greatest as soon as possible. And in some cases beforehand
[25:12] as as partners so that we can enable those value. And this is verbatim from our teams. It is both from the business colleagues. This is where you want to talk about somebody. Yes, definitely. Um the feedback that we collect was really for me the measurement of success. Um
[25:27] our colleagues, they they tell us that they have these different experiences now utilizing the tool and the agent. And uh the the the thing that I hear most commonly is like, "Finally." It's this comment here. "Finally, we have everything in one place." So, this has been very very key for our users because
[25:45] now they can prepare daily meetings, they can prepare prepare uh analysis, reporting, and everything uh just utilizing the agent, for example. And this saves them a lot of time. And also the confidence that they have in looking into the data because they have they can talk to the data is uh uh definitely
[26:02] higher. And this, of course, has also uh contributed to the exponential adoption that we had there. So, but also from the tech side, I guess said you have some feedbacks. Absolutely. It was a big challenge to bring newer technology because every day the data engineers had to adopt, the
[26:19] data scientists had to adopt. But within the ecosystem itself, we've built so much automation now to that acceleration happens at speed and scale, and we are able to see value coming out of at the end of the day. We're closing coming nearing to our end
[26:34] of the slides, but I also want to give you a little bit of perspective that we do not want to stop here. And if you are on such a journey, I would say continue to accelerate or if you're embarking such a journey, we are happy to share any insights, but we really want to bring AI at scale, agentic at scale, and
[26:51] that's where we are working very closely to ensure that there is continuous development, continuous monitoring of quality with zero ops that was interest yesterday. It was really amazing in terms of what's possible and we are rolling it out across enterprise Mercedes-Benz at scale, not just at
[27:07] Mercedes-Benz USA. That's where we want to see the big value and the enablement of being the best always. So, yes, you said it correctly. We are coming to the end of our our presentation. But before that, please spare another minute with me. I want to
[27:24] show you what were the lessons learned that we collected along the road. So, first thing is a unified semantic layer is non-negotiable. Please build or the recommendation is build a unique data semantic layer foundation in order to avoid that
[27:41] users, they look at different data at the same time and confusion is always there. Secondly, from our experience, the introduction of the knee, the introduction of an agent was really a breaking moment in the in the in the project. This has been increasing the speed, has made people
[27:58] more users more confident in utilizing the tool and also understanding which data they have, right? Sometimes when they do not understand, they ask directly to the agent, which data are you running the analysis on, for example. And then third, last but not least, bring the logic to the data and not the
[28:13] data to the logic. We have modern platforms such as Databricks that allows us to bring this this type of logics to the data. So, let's keep doing that and pushing that because this will be bring us a very business value, I believe. So, those are those were for us not just
[28:30] abstract lessons that we learned, but were practical enablers of success. If I have to add before we conclude, this is what is the power of bringing the technology in the right space. You have your business partners talking about technology. I don't need to talk
[28:47] about it anymore. The business partners are AI-enabled and they're pushing us now to really bring the latest and the greatest. And from a IT perspective, I'm really happy that I can go back, sit in the beach, and my agents are going to do the work for you. You still have to sell the cars, but we're going to get you the
[29:03] agents who can make that happen. I will check that definitely as well. So, guys, thank you so much for being with us today. We hope we gave you We hope we gave you a perspective on how business and
[29:19] tech comes together and what's the value added.

Learn more about the Databricks Data and AI platform.

The information provided herein is for general informational purposes only and may not reflect the most current product capabilities or configurations.