From Power BI DAX to Metric Views: Building Trusted AI with Databricks Genie

Summary

Mercedes-Benz Korea built an automated DAX-to-SQL transcompiler that converts Power BI semantic models into Databricks Unity Catalog Metric Views, automating 40% of DAX measure conversions and establishing a single source of truth for both BI and AI workloads.
A three-stage implementation journey — data foundation with medallion architecture, AI-ready semantics with metric views and Genie, and trusted AI through 44 quality gates and 100% Power BI alignment testing — ensures AI-generated insights match existing reporting definitions.
Persona-based multi-agent systems with Unity Catalog governance enable role-specific insights across sales, marketing, and finance domains, with each agent configured to surface metrics and data relevant to its specific business audience.

From Power BI DAX to Metric Views: Building Trusted AI with Databricks Genie

Mercedes-Benz Korea developed an automated DAX-to-SQL transcompiler to convert 600+ Power BI semantic models into Databricks Unity Catalog Metric Views, enabling trusted "Talk to Your Data" at scale. By moving KPI definitions from Power BI into a unified lakehouse foundation, they created a single source of truth for both BI and AI workloads.
Learn how Mercedes-Benz implemented a three-stage journey: data foundation with medallion architecture, AI-ready semantics with metric views and Genie integration, and trusted AI through quality gates and 100% Power BI alignment testing. Discover how the transcompiler automates 40% of DAX measures while handling complex patterns, and how persona-based multi-agent systems with Unity Catalog governance enable role-specific insights across sales, marketing, and finance domains.
🤝

Chapters

FAQs

What is a DAX-to-SQL transcompiler and why did Mercedes-Benz Korea build one?

A DAX-to-SQL transcompiler automatically converts Power BI DAX measure expressions into equivalent SQL metric view definitions for use in Databricks Unity Catalog. Mercedes-Benz Korea built one to migrate their 600+ Power BI semantic models at scale, automating 40% of the conversion work and enabling a unified semantic layer for both BI and AI workloads without manual rewriting of every measure.

How does Mercedes-Benz Korea ensure Genie answers match Power BI reports?

Mercedes-Benz Korea implemented 44 quality gates and a test suite that validates 100% alignment between Genie answers derived from metric views and the corresponding Power BI report figures. This testing framework gives business users confidence that conversational AI results are consistent with the official reporting they already trust.

What are metric views in Databricks Unity Catalog?

Metric views are governed, reusable KPI definitions stored in Unity Catalog that provide a semantic layer on top of lakehouse tables, standardizing how business metrics are calculated across all analytics and AI consumers. Mercedes-Benz Korea uses metric views to centralize KPI definitions previously scattered across hundreds of Power BI reports, so that Genie queries return results grounded in authoritative business definitions.

How does Mercedes-Benz Korea use persona-based multi-agent systems?

Mercedes-Benz Korea is building multi-agent systems where different AI agents are configured with role-specific context, tools, and metric views tailored to sales, marketing, and finance personas. Each persona-based agent answers domain-specific questions using metrics and data relevant to that role, governed by Unity Catalog access controls that ensure each user sees only the data they are authorized to access.

Full transcript

[00:07] Good morning, everyone. Thank you so much for coming. I'm super excited that so many of you are here today. Um I hope you enjoy the keynote this morning and are as excited as we are for the latest Databricks product. And before we get started,
[00:23] um there are some forward-looking statements in our presentation. So, yeah, do not treat them as the guarantee of future performance or uh any outcomes. And we'd love to hear your feedback, so feel free to use our app and then click
[00:39] the button on the top left corner and then, yeah, fill out the service uh at the end of the session. We'd love to hear how you find this session and uh your suggestion for future sessions.
[00:55] My name is Sai. I am a solutions architect at Databricks. Uh I have a machine learning data science background, usually based in Frankfurt in Germany. I work very closely with the Mercedes-Benz account and I have the privilege today to co-present with my co-presenter Faris. So, over to you,
[01:11] Faris. Good morning, also. Warm welcome from my side. My name is Faris Kamal, German by passport, um living in Korea for 5 years and representing Mercedes-Benz Korea. Um I've been in Korea for 5 years, as I said, and I'm
[01:28] the CEO and AI officer. And in my third role, I'm as well the AI lead for 30 markets, any market, basically, except US and China. So, and I've spent 20 years in the company on sales side, always in sales, and actually on the business side and IT
[01:45] side. So, um before we start, actually, we looked at the numbers, who is in here, and we want to make a quick check. We actually have how many data analysts we have here? a quick hands-up, please. Okay,
[02:00] good. And how many um data engineers do we have here? Okay. Good. Then we have on the CTO and CDO executive management.
[02:16] Good. These numbers are completely opposite to what we have in the registrations, but it doesn't matter. Talking about data, right? So, we come later to that. We have to trust these numbers. But we try to split that presentation in terms of a very hands-on um presentation of what Sai is going to
[02:32] show you. And so, let's get started. Um So, our session today is called from Power BI DAX to Metric View KPIs. And actually, we take you through the process how we automate automated that.
[02:48] Uh before we do that, give me a couple of slides where we are coming from. Um So, and actually, Mercedes-Benz Korea, who Who doesn't know Mercedes-Benz? So, I don't have to explain that, I hope. Good. But maybe something you don't know
[03:04] about Korea. I'm quite sure you well acquainted with the food, maybe with K-pop. Uh but maybe three things you don't know about Korea, unless you've lived there, is uh luxury. Um if you if you if you have to guess per capita spend per person per
[03:20] year for luxury goods, on what position would you see Korea? First place worldwide, second or third place? Who's for first place? Second place? Third place? No idea?
[03:37] First place, 350 US dollar per capita. If you see Gucci stores, long line. And they limited Gucci, you can only buy two bags a year. And prices are much higher than anywhere else. Good. So, um actually, as a vision what we
[03:52] have formulated 2023 is we at Mercedes-Benz Korea want to become an AI-driven and data-driven company. But actually when you look at it, we are also a company. We don't produce, we are importing cars, we're selling cars, we're importing parts, we're selling parts to our dealers through to our
[04:07] customers. So, we are quite small organization and that's what you have in all our major markets. Um, so our business reporting to our headquarters in Stuttgart ending up with silos in terms of systems, in terms of data lakes and data solutions. So,
[04:23] actually the first thing what we had to do is to say before we can talk about AI, we have to lay the foundation consolidating the data and automatizing our process. That what we have spent the last 2 years. So, now we are ready to look into AI. We are not fast movers in AI, but I think we have done a good job
[04:39] in laying the foundation and talk to your data was one of the topics we want to prioritize because it is actually one of the topics we see biggest potential in. For limited, um, scope which we started at and I take you through that
[04:55] uh, in a in a minute. What is important is what we build in Korea is actually also what we want to scale to the other markets then. Now, um,
[05:11] besides Korea being luxury market, it is also one of the most, um, uh, digital market in the world. Um, you have um, 100% uh, broadband internet which says um, 100% of the household has broadband
[05:27] internet. You have 5G coverage, underground parking minus three, you still can watch your Netflix. Whenever I go back to Germany, I feel like in the stone age uh, compared to that. And also it has formulated a very broad ambition to become number three when it comes to AI by 2030. And you've seen the KOSPI
[05:44] and the Korean index and you've seen the share price development of Samsung and Hynix. You can see that they're in a very good position to do this. There's another reason for Korea being so strong is they have grown within two generations from
[05:59] agricultural country to this high-tech powerhouse in only two generations. They have double growth in productivity. So, being in the market means you have to be very fast. You have to deliver
[06:15] solutions very fast. We as Mercedes have positioned ourselves remarkably well. We have first biggest market, so it's China, it's the US, it's Germany, it's Germany, it's UK, then Korea with 52 million inhabitants. But it's not so much about the sales number what we sell there. It's the the cars we
[06:33] sell there. It's number one market for the E-Class short version. And it's always among the top five when it comes to S-Class and Maybach. So, that's where we sell the good volume for us. So, that's why part of the success is how we turn data into actions. And that's why when we talk about data and
[06:49] we looked at the KPI, the business performance KPIs. We didn't look at knowledge. You heard a lot about knowledge if you've seen the presentation in the morning. You can do a lot with Genie, but we focus actually on business KPIs. Now, actually, how do we talk to data today
[07:05] is I asked about how many of you are data analysts in Power BI. Is what we say is when you look at it as all the companies we have been using Power BI. I mean, if you break it down, you look at it and say KPI for us is how many E-Classes did we sell to a private
[07:21] customer. And in Power BI, you have this KPI quite fixed. So, a sales person know what it is, a controller knows what it is. But however, if you want to change that, if you want to change that, compare it to last year, if you want to change it to give me a different visualization, the
[07:36] team goes back and comes back after a couple of days or weeks and says here's a change and you continue. We have been busy continuously doing these changes. The good thing is you can really trust what you see there. So, on the other side we have talked to your data where you actually can just
[07:53] ask any simple question and you get an immediate answer. The challenge here is can you really trust that answer and is it really the answer which is comparable to the Power BI? And this is the biggest challenge we have faced at the beginning and if you look at Gartner says actually 60% of the
[08:09] AI projects which fail because of the missing AI ready data. So, we have to tackle this. But basically this is something where we say adoption is built by trust and not by technology. And something we have been facing before is even in Power BI we have a lot of
[08:26] challenges to build this trust. And the typical two situations I usually have is you have a Power BI report, you sit in a meeting and the business says, I've got feeling that number doesn't fit. So, and then you go back to your technical team and the technical team said, yeah, we have a technical issue. So, the next time in the next meeting
[08:43] you come back and you see your business doing the X report again. So, this number one. The number two is consistency. So, you do your planning in January, you have certain numbers, you look back in July and the numbers have changed. So, people start comparing again and there also studies which are
[08:58] 70% of managers always validate the data before they actually act. So, now with AI, why is it more challenging with AI is because we've seen it I hope everybody has seen the session in the morning. It was also shown yesterday. Hallucination, I really loved
[09:14] Genie's answer, I strategically fabricated that answer. So, that is one thing. How often have I been with the CFOs who says, okay, I ask a simple question, I get the wrong answer, I don't trust it. Number two and number two is if you build it as a separate product,
[09:30] it gets access directly to your database. It always comes up with an answer, but it's wrong. So, most likely wrong, the other way around. So, now how do we build a stress and that's why we talk about the compilers, we believe that we say whatever is in Power BI is the truth.
[09:47] So, when we talk to data, we look into Power BI, we talk to data and it must be one-to-one. That's the starting point. Otherwise, you don't have to fix one side of the story, right? And before now we will talk about the compiler, um it's
[10:03] All our data is in um lake base. This is number one. Uh as I said before, we have to bring together different databases. Uh second one is we have one source of truth for all the products we use, reporting, automization, and now AI. And number three is we have a KPI layer,
[10:20] what we call KPI layer documented before Unity uh catalog. We had everything documented in Excel. What is so important is you have a clear definition of the KPI and you have a clear coding how it's being calculated. And um we have an data and AI board
[10:38] that's a VP level which decides what's going to be the road map. We have me as an AI and CIO AI officer and CIO making sure that AI and data works in one program, not in two different ones. But more most important is we have one business leader actually who's
[10:53] coordinating all the business colleagues to say the definition of KPIs must be same, no redundant KPIs on different uh KPI reports, one clear definition, right? So, having said that, we started in February 2026 with Talk to Your Data.
[11:09] And as I will now take you through what we have done so far. Important is we are at the early stage. Uh we are at the early stage. We are happy to share what we have been doing. We are at the early stage in Korea to scale it and also to the other markets. Thank you so much, Faris.
[11:25] So, as Faris mentioned, Mercedes-Benz Korea already built a mature and unified data foundation Uh with Lakehouse and Unity Catalog. So, all gold layer data is governed and managed in Unity Catalog in one place. So, the Lakehouse Foundation and their
[11:42] own master KPI catalog already prepared them very well for uh implementing enterprise level talk-to-data solution. Now, the real challenge is how can we make all this reporting data AI ready? So, AI can understand the uniqueness and
[11:59] the business logic of all those domain-specific KPIs and return the high-quality answers that users all can all trust. Um to achieve this, AI doesn't need more intelligence. Just as our CEO Ali says this morning. Uh the real challenge is on the context
[12:14] layer, right? So, it needed this more context. To be more precise, uh a unified and governed context where uh any agent can openly access and reuse those contexts without any silos and lockings. And that's what we call the unified semantics.
[12:32] Uh we want those semantic layer to sit as close as to your data, right? So, both BI and AI workloads can capture them in just one place on one platform. And this is the approach that Mercedes-Benz Korea and Databricks worked together to implement a unified
[12:47] architecture for data, semantics, and AI. Mercedes-Benz Korea follows the medallion architecture to build the data pipelines and jobs using Lake Flow. Um that includes the headquarter data and the local market specific data. They
[13:02] are all ingested. They are all ingested into Delta Lake and joined to produce the silver and gold layer tables. So, in the end, you have the BI ready data products um for analytics in the Lakehouse. And all data assets are governed inside Unity Catalog. So, that
[13:17] is the data foundation that Mercedes-Benz Korea already had. Now, we want to build a semantic layer in the same place where the data sits, right? So, as Power BI is the main reporting tool for Mercedes-Benz Korea, all KPIs were
[13:34] defined as DAX measures there um for reporting convenience. Uh to make those Power BI DAX measures available and AI-ready, we need to convert them into Unity Catalog metric views and expose them to and expose that semantic context to AI.
[13:50] Since Mercedes-Benz Korea is already using the master KPI catalog to maintain those definitions for business specifics, uh we use it as the semantic contract uh to curate the metric views inside Unity Catalog, and we define all the governance policies inside Unity
[14:05] Catalog. When we have all those enterprise contexts uh captured as metric views inside Unity, we can easily onboard them now to Genie and let users interact with the enterprise data freely in natural language. And if needed,
[14:22] um you can also use Genie as knowledge source for external AI applications, for example, Copilot to extend the capabilities uh of your own AI solutions or other platform. Now that we have AI-ready semantics uh in Genie interface, we can go one step
[14:37] further, right? To compose those custom multi-agent systems and stateful agentic apps. For example, um we can uh Mercedes-Benz Korea and Databricks are also developing a persona-based mult- um multi-agent systems.
[14:53] And that multi-agent system can automatically pick up the knowledge domains uh based on the user's role and also return the answers tailored to the user's specific persona. So, this unifies uh the architecture and eliminates also the semantic silos. In
[15:09] the end, it provides a reusable and interoperable semantic foundation for agentic AI. Now, the unified architecture is the outcome for trusted talk to data. To get there, Mercedes-Benz Korea and Databricks went through a whole journey
[15:26] with different priorities. On a high level, there are three stages in this journey. Each stage has a different technical focus and implementation best practices you should consider when you develop the journey AI solutions on Databricks. And these are the insights we have
[15:42] distilled from our journey with Mercedes-Benz Korea. And you can apply them also in your own organization and we would like to share them with you to help you also be successful in that. The first stage is data foundation and this is the prerequisite for any successful AI projects.
[15:59] Mercedes-Benz Korea is the best example. They didn't start from complete scratch. So before we started the project, Mercedes-Benz Korea already implemented the Lakehouse architecture, Unity Catalog, and the semantic contracts. Those are the catalysts for a successful
[16:15] agentic AI development cycle. And the second stage is building AI ready semantics. This includes moving semantics from BI reports to Unity Catalog metric views and it also includes curating those metric views for better performance and for better
[16:31] context of AI agents. And the third stage is building trusted AI. An effective effective semantic layer provides AI with context. But the context is only valid if the AI response quality is can consistently
[16:47] match the the user expectations and the Power BI reports in our case. So therefore, in the case of Genie, systematic quality control, answer benchmarking, and user incorporation monitoring, those are the best practices for making talk to data production
[17:03] ready. And we will illustrate this journey with using our experience with Mercedes-Benz Korea. And since they already have a very mature data foundation, and uh will focus on the learnings on the second and third stage, uh building AI ready
[17:20] semantics and trusted AI. Mercedes-Benz Korea defined over 600 KPIs as DAX measures in Power BI reports. Uh making the business logic for Power BI DAX also available in Unity catalog metric view was the key step in building
[17:36] an open semantic layer. So, build rebuilding all those Power BI measures manually inside Unity catalog would be a very tedious and lengthy task. And therefore, my awesome colleague Andreas Yek uh built an
[17:52] automated transcompiler, and it could accelerate the translation of any DAX measure to metric view. So, this is how it works. The transcompiler allows us to migrate the semantic models at scale. The It first parses the Power BI reports
[18:09] uh and the semantic models, extracts each individual DAX measure. It then creates a metadata catalog, such as fact tables, dimension tables, uh join keys, and then um and then for each extracted DAX measure, it maps the DAX measures'
[18:26] source tables with the Unity catalog counterparts, and finally generates the metric view definitions um such as source, joins, dimensions, measures, and then also the ready-to-use SQL statements to create those metric views. And before compiling the metric views,
[18:42] the transcompiler also validates the syntax and also the aggregation logic. Uh the created metric view can then um be further consumed by the downstream applications, such as Genie agents or apps, through the SQL warehouse. So, it's worth noting that the transcompiler
[19:00] will also flag non-automatable measures for manual review, especially for those very complex DAX measures, uh such as row context manipulation. And when the transcompiler pipeline it produces an evaluation report with
[19:15] conversion statistics, gaps, remediation strategies. For KPIs written in simple Power BI DAX, the Transpiler could 100% automate them. Uh it is only the complex DAX measures that require manual effort for for conversion.
[19:31] So, in case of those 600 plus KPIs at Mercedes-Benz Korea, uh 40% of them were fully automatable, and the remaining 60% were complex DAX measures. So, our engineering team has also released a private preview uh for
[19:47] the BI migration tool inside Genie code. So, you can also use now Genie code to accelerate the conversion of those complex DAX measures after the first step of the Transpiler.
[20:03] So, after you create metric views in Unity Catalog, the next step is to curate them into a trusted context layer for your AI consumption. And data quality is the top priority uh for any business users, not just Mercedes-Benz. I think that is the top priority for any organization
[20:20] that would like to implement talk-to-data. Onboarding those KPIs as metric view into Genie space is a key step, but on its own is not sufficient. It does not guarantee reliable answers. A metric view is only valid if the Genie
[20:37] responses are correct and validated. So, how did we build this trusted and AI-ready semantics for Mercedes-Benz in Korea? The internal target, as Varus mentioned, was to fully align um Genie answers with
[20:53] corresponding Power BI reports, 100% match for each KPI in scope. During the pilot with Mercedes-Benz Korea, uh Databricks and Mercedes-Benz Korea jointly documented those best practices uh for curating, optimizing
[21:09] metric views and genius spaces. This includes the use of agent metadata in metric views and benchmarks and these practices support AI ready semantics and consistent genius answers. To make it reusable and scalable across different organizations, we encapsulated
[21:26] those best practices into a standard technical handbook. It includes five phases. First, prepare the KPIs. Second, build the semantic layer with the metric view. And the third one was to also organize your metric views and
[21:43] genius spaces by domain. Then you need to test and incrementally tune your genius space quality to make sure that your answers are correct. And finally, you will validate as you onboard new KPIs and measures into metric views. Test them
[22:00] with internal engineering team and also with the users users before you release them into production. Now, following this iterative five-phase process, our business units can build AI ready semantics on Databricks data intelligence platform.
[22:17] Also, Databricks is developing an an app solution to automate this process which Mercedes-Benz can leverage for further rollouts. I will hand over to Faris for a few more comments on this part. Yeah. Yeah, just to add on. I think it's
[22:32] technically it's feasible. But you have to dig into it. And what we want to make sure is that we have a manual and an execution that we can really execute it in the other markets. So, that's why we have developed 44
[22:47] quality gates. Each one coming with traffic light system which is unless you're green, you have to do something, right? So, ideally you don't have to think much. You run through, you get a traffic light and you get immediately the actions you have to take. So, that's how we really want to give to our market
[23:03] and a very transactional process that they can simply execute. And as as I mentioned, what is very important, we're currently doing this in a in a still manual way having word document, but tomorrow that's going to be an app which ideally is embedded in the ecosystem, right?
[23:19] Good. I think now it's time for demo. Good. Thank you, Faris. So, now we would like to show you a very quick demo on how the solutions work in Genie One.
[23:36] And and maybe one note, um it's public data we're going to show you. I can't show you the internal data we have of sales numbers and so on. So, it's a public data use case. Um so, you're going to see registration data in Korea, uh which we get once a month from the government. Um but just to showcase the capabilities
[23:51] of of Genie here which we have. And we use a data in combination with our internal data, of course. Yeah. Yeah. So, this is a Genie One, and I'm now
[24:08] also in the workspace of Mercedes-Benz Korea. But as Faris mentioned, we only use public data set to demonstrate the capabilities of Genie One today. Um so, let me first start the question, and as Genie reasons through that, I
[24:23] will explain a bit context about um this assignment. So, KAIDA database, Korean Automotive Importer and Distributor Association database, is a public database for Korea
[24:38] uh vehicle administration. And there we'll find all the vehicle registration data for all brands, for all automotive brands that are registered in the Korea markets. And it has already the different um dimensions and source data covering uh
[24:54] different regions and also the the registration volume. Mercedes-Benz Korea used Databricks Lake Flow to ingest those source data from uh Kaida database. And and I also use uh Spark and Scala
[25:10] pipeline to transform and clean those source data and to um transform them into silver and gold tables. And in the end, following also the processes we mentioned before, uh they transform those data into gold layer table and also on-board the metric views on top of
[25:27] that into Genie space. So, now we are actually using Genie 1 to interact with those gold layer data inside the relevant Genie spaces to gain more insights. For example, in our case, that would be competitive uh analytics.
[25:44] And as you can see here, Genie's thinking through um how what kind of KPIs are available inside this Kaida database. And Kaida is also a group creation. And Genie understands because it has the context, right? It has the this those definitions
[26:01] we have provided Genie with. And it comes up with a set of um KPIs you can reason through and you can aggregate uh on top of those raw layer that it has identified and also in those Genie spaces it has access to.
[26:17] And for example, he summarizes those KPIs into different categories. He has registration volume KPIs, he has market share KPIs uh across all brands of automotives, and he also has pricing discount KPIs. Okay. And also it it tells you those uh
[26:35] segmentation dimensions that are available for slice and dice those KPIs. Um now we are going to ask a little bit uh deeper question to Genie 1. So, we want to understand uh the registration
[26:51] distribution by geography for 2026 YTD. And also Genie understand what YTD means because we have defined those measures, those metric views with those business definition year-to-date and Genie is now able also to reason
[27:06] through that and provide us with a distribution for for the registration volume by region and in the year of 2026 until now. And as you can see here that is starting to queries through different dimension
[27:23] tables we have because Genie 1 does not just have access to the individual Genie spaces you have created. It has access to all the underlying data sets that I can access based on my identity. So if you as a admin restrict my access to
[27:39] specific data sets, you I can only gain access to those data sets with my scope privileges. And that would mean different people using the same Genie 1 will get different answers based on their privileges and permissions to those data. So now Genie 1 is still thinking through
[27:56] that. Um So this exercise here is to first understand the distribution of the um of the volume for registration and then we also want to understand for example in the primary usage address, right? And
[28:12] how does the registration volume compare with the competitors? So it's taking a little bit longer today and but we'll get there.
[28:28] And then once we have this distribution analysis, um we can ask Genie to further build a heat map for example to help us understand the performance between uh competitors for example. Um Mercedes-Benz versus BMW and also by region.
[28:44] Now it looks like that Genie is providing some answer here. Yeah, so this is the 2026 YTD registration overview by geography. All data coming from the public's data space Kaida. As you can see here, it
[29:01] tells you the total number of registration in 2026 YTD in the Korean market. And it also highlights the Mercedes-Benz market share. Um and then you can quickly understand where Mercedes-Benz has a stronger presence in
[29:17] individual regions in Korea. And also it compares the also the the numbers of total vehicle registrations among those different competitors. Now as I said, we want to take a deeper look because there are different ways to uh there are
[29:33] different registration um addresses inside the database. One registration address is the primary usage address. And we want to use that and want to understand for those vehicles uh used in this primary registration usage address, we want to understand the
[29:50] registration volume among the competitors and build a heat map based on their performance.
[30:08] As you can see here again, uh Genie wants to using the vehicle um registration data set as the dimension table and then also it purely identify the address usage field because there as I said, there are different ways uh to represent the the registration location of a vehicle.
[30:24] Uh address usage is one of them. And now Genie is uh executing the SQL query. We will soon see a heat map of comparing the market presence between Mercedes-Benz and its competitors.
[30:48] And once we have that, we will resume to our presentation. And um yeah, and talk about how we build also build a persona agents on top of that as a outlook. Now, I think it's the last step of submitting the SQL query and we should be able to see that in a minute for the
[31:04] result.
[31:21] Our cloud data different factors that could impact the performance of your uh Genie one. All right, usually there will be the data layout, how optimized is your data set. Second, that will be the query complexity. If there are a lot of steps here involved, that will also impact the speed. And finally, it's also
[31:36] the compute size, all right, the SQL warehouse you are using for your individual genius spaces. Uh and that could be the reason impacting also the speed. So, now you see that there's a heat map of all the competitors registered in the
[31:51] database, and then we can clearly identify that Mercedes-Benz has a good market dominance in the top regions, but also there are some uh for example, Busan region could be a potential improvement area where Mercedes-Benz has lower registration compared to BMW.
[32:07] And also, in the end, it gives you the key findings, uh summaries, competitive insights. Um and also, you can identify all the data sources it has been it has used. So, now let's switch back to our presentation.
[32:44] Right. So, this is the journey of how Mercedes-Benz Korea built trusted talk to data at scale. It includes three stages as I mentioned. Remember to follow them when building the production grade service self-service AI solution by yourself. Data foundation, AI ready semantics, and trusted AI.
[33:13] So, a single genius base can answer questions about a single domain, but the real talk to data experience from Mercedes-Benz Korean um needs to span across different departments, business units, sales, product, marketing, customer service, and so on. So, this requires tailoring what each persona
[33:30] uh can see based on the their role and permissions. And therefore, Mercedes-Benz Korea and Data Bricks are also working together to develop the next generation of talk to data, the persona-based agentic system. So, we would like to share uh an outlook for this solution with you, and I'll hand
[33:45] over to Faris to give you a high-level outlook on that. Yeah, in a nutshell, what you see is talk to your data still looks at users using the data. And I mean, you've seen the speed, and you said it's slower, but if you compare it, okay, that question asked today in a traditional way, it
[34:00] takes days, okay? So, yeah, the speed gets there. I mean, that's a discussion I also had with the business. But we're still looking at users using the data, right? And the next step is to say, what do you do with the data? You analyze it, you formulate an action, you execute. And that's the next step what
[34:16] we want to do. It's great jump forward right now getting this data, but we want to build agent persona agents. I'm CFO, I'm VP sales, and we want to analyze what they actually do, and then mimic it with an agent to say, "Actually, if you see no report, everything is fine. But if
[34:32] certain numbers hit a threshold, you will get a notification. And maybe then the one agent talks to the other agent asking to see if it was eight, asking why the number is going to this direction, and the sales agent answers. So, that's the next step we want to do. Um we are we're working on this um till
[34:47] July, we want to have the first pilot. Um technically speaking, of course, it's challenging. I think it's also a challenge on the on the change management side. How now to identify these thresholds? Because you have that in the app, and we talked about ontology in the morning. I think that's a natural
[35:02] next step that we're going to look at. But, we are very excited, and maybe next time we can also share on this journey what is the experience. And again, I think many companies have built that. For us, it's a natural next step. Yeah.
[35:22] Yeah. So, um Thanks, Marius, for clarifying that. Now, we I would like to quickly show you uh how this structure of this multi uh agent system looks like. So, we combined the Genie Agent Bricks and Databricks app to develop this persona-based agent. Uh the
[35:37] user can interact with the agent system through a custom interface, a custom front end provided by Databricks app. And then, the parent supervisor agent will route the question to a persona agent with uh
[35:53] role-specific instructions defined for each individual persona agent. And that persona agent will consolidate the insights from different Genie spaces and across relevant domains. Um and all those um access and control is governed by Unity
[36:09] Catalog with row and column level uh access defined for each user. So, therefore, this agent system can just leverage the knowledge across multiple domains, and also within the building uh governance provided by Unity Catalog based on the user role.
[36:27] So, Marius, would you like to maybe have some comments to reference our another session that we are having this afternoon? So, we look very much in Korea, Korea being a sales organization. My colleagues here having an an MB-wide view on it, they're actually
[36:43] building what we call the agent gone where they building a platform where you can actually any agent from any organization can communicate with each other. So, please, if you have the chance, join Marcus and the team. They have a session today as well with Sai presenting this. And my agents, which I
[36:58] built, going to be there so that anyone who can use it can use it outside of Korea. Good. That would conclude our presentation.

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.