Skip to main content

Agentic Analytics on Databricks Lakehouse

Summary

  • This video explains why traditional data warehouses designed for human analysts break under AI agent workloads—flawed semantics cause hallucination, fragmented governance exposes unauthorized data, and costs explode when one analyst effectively becomes thousands of AI agents.
  • Genie Ontology is presented as the mechanism for teaching AI agents business-specific terminology and context so they resolve queries correctly without generating inaccurate results, grounded in semantic definitions that data teams have already validated.
  • The end-to-end agentic analytics architecture on the Databricks Data and AI platform covers serverless compute for agent query patterns, Unity Catalog governance guardrails for agent access and outputs, and Unity AI Gateway for consistent agent serving across all consumers.

Agentic Analytics on Databricks Lakehouse

Watch: Agentic Analytics on Databricks Lakehouse
Agents promise to scale instant access to data insights across organizations. However, most data warehouses were built for humans reading dashboards. Agents expose every assumption that breaks. Flawed semantics lead to hallucination, fragmented governance exposes improper data access, and cost/performance issues explode when one analyst becomes a thousand agents. In this video we'll walk through what agentic analytics looks like in practice on Databricks Lakehouse, demoing the patterns customers are running in production today. Following these patterns, you can dramatically increase your data team's impact across your organization while reducing time-to-insight. You'll leave with a clear model of approaches to agentic analytics at scale, and see why Databricks Lakehouse is built for it.
Talk By: Ben Tripp, Associate Product Manager, Databricks ; Moe Steller, Adoption Lead, AI + DBSQL, Databricks ;

Chapters

FAQs

Why do traditional data warehouses fail when AI agents run analytics?

Data warehouses built for humans assume that ambiguous queries will be resolved by a person who understands the business context. When AI agents run those queries at scale, ambiguous semantics lead to hallucinated answers, governance gaps allow agents to access data they should not, and the sheer volume of agent-generated queries creates performance and cost problems that human-driven workloads never caused.

What is Genie Ontology and how does it prevent agent hallucinations?

Genie Ontology is a semantic layer within Databricks that allows data teams to define business terms, metric calculations, and entity relationships that Genie uses when answering analytical questions. By anchoring agent query generation to validated semantic definitions, it ensures that when an agent asks about a business concept, it uses the same definition the data team intended rather than guessing.

How does Unity Catalog governance enforce policies on AI agent analytics?

Unity Catalog's governance framework is extended in this agentic context to control what agents can read, what they can write, and what SQL patterns they can execute. This means the same access controls and data quality policies that govern human analysts also apply to AI agents, preventing agent sprawl from bypassing governance rules.

What is Unity AI Gateway and why is it important for agentic analytics?

Unity AI Gateway provides a consistent serving layer for AI agents across multiple consumers, ensuring that the same governance policies, rate limits, and model routing decisions apply regardless of which application or agent makes a request. This prevents a fragmented landscape where different applications enforce different policies on the same underlying data.

Full transcript

[00:07] Welcome everyone. Glad everybody's here day three. I hope we are all still awake and not super sleepy yet. But I'm right before lunch or right after lunch, we have this wonderful session here about agentic analytics.
[00:19] So just a little couple of housekeeping things. Forward-looking statement, you have seen this by now I think a gazillion times. I don't have to cover that. Neither the survey, so if you would love to give us some feedback on the session
[00:31] afterwards, we would love to to see your feedback or come up for a little Q&A by the end. I'm going to kick it off right into the content here. So going to introduce myself, Mo Stellar. I run the Lakehouse
[00:44] AI adoption here at Databricks and I have my colleague Ben with me. Ben, do you want to introduce yourself? >> yeah, nice to see everyone out here. I'm a product manager here at Databricks where I work on Lakehouse and integrating Lakehouse into different AI architectures.
[00:57] >> Sweet. Thanks, Ben. And topic of today, agentic analytics. So before we dive into the content, just going to want to see like a quick raise of hands, agentic analytics, who is talking internally in in your teams or
[01:09] in your company about this topic? I'm just going to raise my hand myself since Ben and I we're talking about this since the beginning of the year. I think a little bit more intensely. Okay, very good. So what we want to break down in
[01:22] this session is a little bit the paradigm or the pitfalls, operationalization of agentic analytics. We're going to talk about the products specifically in Databricks, obviously, there are going to be some announcements
[01:34] that are going to be tied in into the end-to-end and we have a wonderful demo that Ben is going to deliver in the end. And that leads us actually to our third person that will be part of today's
[01:48] presentation, Jess. So meet Jess. Jess is here for all of us, I would say. She is the head of analytics at a company called Wonder Bricks. And um
[02:00] Jess is in the unique situation uh that she gets a lot of requests every days from executive level stakeholders. She needs to response and respond in reports. She needs to respond in ad hoc
[02:12] queries. She needs to response to all these analytical questions that she is uh facing on a day-to-day with her team. Um unfortunately, her team is not going to grow. Um so, she needs to meet the
[02:24] demand from executives. And there is a an example here um that she gets asked in her travel company um how the completion rates are in Switzerland. So, she's based in the US. She needs to answer a question based on
[02:37] the the Swiss market. And um as well as questions around um churn in APJ. APJ stands for the Asia Pacific Japan region. So, it's an acronym that is used
[02:49] in her company. You see where the journey is going here what I'm trying to tee up, right? But there are essentially a lot of requests that we she gets all the time. And her team is trying to answer those now with the rise of agents
[03:04] of GenAI over the last years. You probably know that your business stakeholders are very keen to get even more answers from data. And they're trying out all these tools that are out there um
[03:16] to sort of like get ahead of Jess's team number one or just ask her over and over again, "What's the answer for XYZ?" So, ultimately, Jess decides the current state of my analytics platform needs to
[03:30] transition and align to the agentic era. She knows that agents are going to help her to meet the demand of all these questions. So, Jess is smart. And uh she makes that decision. She
[03:43] discusses that in her company with her CIO, who's very important, with a finance team, and so forth. And um everybody gives her thumbs up and, uh, grants her access to this new world of
[03:56] agents. And, um, since Jess is around for decades, I would say she's very experienced as the head of analytics, she went through several of these technical transitions.
[04:09] Beginning of the cloud era in the mid-2000s, rise of ML platforms, 2010s, basically Journey AI when that came around. And so, she's aware of very, very, very fundamental pitfalls that
[04:21] came up over and over and over again. And to get ahead of that, Jen, uh, Jess is thinking about how to get around, basically, those pitfalls.
[04:34] I want to touch on these pitfalls and then we're going to look at our exciting product stack here that Ben is going to cover. So, number one, pitfall number one, accuracy. You probably have heard this now during the course of the summit several times what accuracy means
[04:47] specifically to to agents. Jess's organization, as you mentioned, is very complex. It's multinational, and so there are regional differences and complexities, um, existing globally. And agents
[05:01] usually lack semantic context to answer, based on these differences, questions. So, here's an example. Which of our top hotels are at risk? This is a completely ambiguous question to an an agent that
[05:14] lacks the semantic context. What does top mean? What does hotel mean? What does risk mean? Jess certainly knows this. Her CFO, or I don't know, some other executive knows what that means, but the agent doesn't.
[05:26] So, she needs to figure out a way to improve accuracy, to add a context layer, so the agent can operate, um, at the expected levels.
[05:38] Number two, her CIO is often talking about a unified governance layer. We know agents lack of guardrails. It's a completely different, um, layer within the stack that needs to be implemented. Agents often operate on
[05:53] sensitive data, on internal tools, and on the open internet for reasoning tasks, for research tasks, whatever you grant the agent access to, as well as there are many other people in the company that we
[06:05] call Shadow IT that develop agents at scale. So, it's not just Jess. There are other uh stakeholders like Jess that want to pursue exactly the same thing. And so, what's happening here is Jess is negotiating with her CIO to bring in a
[06:18] data and AI governance layer to sort of check on agents, to check on cost controls, to really know what these agents do on a day-to-day.
[06:31] Number three. To me, honestly, this is like a very important point here. Um scaling up. So, there is a paradigm to understand or in in this case here that agent number one is easy to stand up. I think we all went through this
[06:43] journey. We all tried things out. We built an agent with cloud code or your your your favorite tool, essentially. And it's super easy. And you you show success. Your MVP gets a check mark. But then the business comes in and says, "I want to go from
[06:58] one to 10, from 10 to 100, from 100 to 1,000." And it's this typical scaling challenge. We had this with ML Ops, with ML models back in the in the past, or sometimes even now today as well. But um
[07:10] with agents, it's sort of the same thing. So, going from one to 1,000 agent agents is basically an expensive journey. Thousands of queries are ran simultaneously by agents, and compute
[07:23] becomes super expensive due to the concurrent SQL queries. Second, performance starts to go down. Warehouses nowadays are not built for the agentic era naturally, because we
[07:35] just basically moved into this age, right? And so, this uh the support of scale of thousands of agents needs to be guaranteed, and this really like is keeping Jess awake at night, but she
[07:47] wants to solve this with Agentic Analytics. So, to summarize what she's dealing with. She's dealing with accuracy, lack of semantics. She wants to take care of that. Second, governance,
[07:59] access to data and tools. Again, her CEO is very keen on that. And number three, the performance and cost piece is very important to her, as I guess for any platform transformation.
[08:11] And how Jess is going to do that and how she is tackling these risks, Ben is now going to show us, and he will basically lead us through the product stack on how she enables her new platform. >> Cool. >> Yeah, thanks, Mo. So, Jess is super
[08:24] worried about all of these risks that Mo laid out. You know, if you're going to roll out agents to an entire company, you better hope that the responses they're giving to your stakeholders are actually accurate. You don't want the cost to explode. You want to make sure
[08:36] those agents are doing things that they're actually supposed to. So, at the same time, she's super excited about all these benefits. Like, what if all of her business users could self-serve simple questions, so that her analytics team
[08:48] can focus on the actual difficult um high-impact analytics tasks that they want to. Um so, she decides, you know, I see these risks. Let's see, is there a way I can create an architecture to
[09:00] address these issues? So, at the very bottom of the stack, we have to start with compute. Given the risks we were talking about, we know that this needs to be high performance at scale with thousands of queries coming in. You don't want
[09:12] your end users to have a really bad experience where they're waiting for super long for any of these answers, even if the answers are correct. And as it scales up, we need to make sure that the costs stay within the reasonable bounds. So, she needs to find some
[09:24] compute layer that's going to be high performance, low TCO, um to minimize that time to insight. The second layer is going to be this governance layer. So, she needs some way to control and have observability and
[09:36] insight into all of these agent access patterns. Not only to her data in her data warehouse, but to all of the external connections that she also wants to configure to enrich all of this agent
[09:48] response types. She also needs to secure the model spend itself because as you scale up to your whole organization, as I'm sure many of you have been dealing with, those costs really begin to add up.
[10:01] Um and so the semantic layer will come in on top of all of this um which will have some sort of unified dynamic context layer and we'll talk into why this needs to be uh dynamic in a bit. Um that has human
[10:14] overrides to make sure you can actually go in there and if you see something that is semantically incorrect, you can centrally govern it as a data team. So for this bottom compute layer
[10:27] we're going to start with Lakehouse. Um as many of you are familiar with, it was you know part of the name of this talk. Uh Lakehouse has a bunch of incredible options that you're already familiar with um from our you know
[10:39] classic and pro offerings to the serverless offering which is fully managed and elastic and on demand. So that lets you really work um at massive scale. But we have this really exciting new offering that I'm sure you
[10:51] guys heard announced um and that's Lakehouse RT. Lakehouse RT is a perfect fit for these agentic analytics use cases. Um it can scale up to super high concurrency um and very low latency for for
[11:04] millisecond speed at scale uh to really maximize that time to insight. And so this is kind of what it looks like with dashboards. Each of these dashboards is running about eight queries at a time.
[11:16] And you can see as we scale up all the way to 10,000 dashboards running at once, the latency is still around 150 milliseconds for Lakehouse RT. So you can imagine this is
[11:29] a pretty good approximation of what it's like for 10,000 agents to be running queries at the same time uh rolled out to all of your business users at your organization. But, even though we're very excited about
[11:41] Lakehouse RT, we don't, you know, want to abandon our bread and butter. All of our existing serverless SQL offerings have been improving uh drastically over the past year. Every year we improve the performance significantly, and it improved even by almost 2x this year
[11:54] alone. Um so, for your larger queries, really big table scans, anything over a uh terabyte and above, uh this SQL serverless uh offering is still uh optimal for these use cases. And the
[12:07] great thing about using Lakehouse is you can use any of these compute types all on the same uh data layer underneath without having to, you know, sign on to any logos or move data around at all.
[12:19] And even in our most recent benchmarking, uh we still lead on TCO, especially as we continue to scale up. So, this is super important for managing all
[12:31] of your central admin warehouse spend as you're suddenly uh no longer just an analytics team running queries and supporting dashboards for your users, but everybody in your whole organization is
[12:43] individually running ad hoc queries. So, uh this is an important aspect to just certainly. Cool. So, we have our compute layer. We have this scalable, flexible foundation
[12:55] at the bottom of everything else. Um we have this unbeatable performance when needed with Lakehouse RT, and then we have for more complex agent-written queries, Lakehouse is able to support that uh serverless SQL uh super well.
[13:08] So, the next important part of the stack is figuring out this whole governance issue, which is very sticky. Um and our best solution for this is Unity AI Gateway. Um we've added all of these
[13:21] exciting new features, um and it allows you to enforce this governance across your entire AI estate, not just your data, uh but we think it's really important to have this live with your data. Because fundamentally, these are data governance problems.
[13:34] So, you can enforce your policies and guardrails across all of the models you use and all of the tools that these agents access. You can create MCP services and manage exactly what tools are accessed and how and by which agents.
[13:47] You can govern the APIs. You can even govern the underlying models and the coding agents and route them through AI Gateway, which I'll show you as part of my demo. And through this, you can track all of
[13:59] these requests, all of the actions our agents create, and you can enforce guardrails for the outputs of all of these models as well, all in one place. So, with these capabilities in place,
[14:12] you can be sure that you can you can sleep well at night. Your agents aren't doing anything crazy on your sensitive data. So, I'll I'll go into a couple of concrete examples of what this actually looks like. So, service policies, this
[14:24] is a new concept that we have in Unity AI Gateway. You can have MCP governance very concretely on in this example, disallowing rights to the Google Drive MCP. So, maybe for your
[14:36] agents, they're always using on behalf of authentication. That's great. There are some pitfalls with on behalf of authentication, right? Sometimes you don't want your agent to have the exact same access that the person using the
[14:48] agent has, right? You want the agent to be able to read documents, but you might not want them to be able to edit those important documents that you have access to. So, this is a great example where we define a very simple policy to disallow rights on that MCP endpoint.
[15:04] Another example here within Unity AI Gateway is unified tracing tables. So, through these traces, you can capture all of the AI activity across all of your models and tools and agents and put them in a single table. You can then use that
[15:17] table however you'd use any table in Databricks. So, you can use it um to turn actionable with Genie like debugging any of your agentic workflows. Um you can use MLflow for a quality monitoring of all of your agents. You
[15:30] can even feed these into Lakehouse to make sure that you have uh the maximum possible observability as you're uh locking down the uh security of your entire data estate.
[15:42] Cool. So, now we have layer two. Governance. You have your MCP services governed, all of the tools that your agents are accessing. And then also the govern uh you can govern the models themselves that are running underneath
[15:54] the agents. And again, using our gateway, you can use models from any of the leading model providers and swap them out as you please. Um finally, we need to figure out the semantic layer. This is
[16:06] I would say, when talking to customers, the trickiest piece of this entire puzzle to figure out. Um there haven't been great models for how to do this in the past. It's been really challenging, especially at scale. Um so,
[16:19] how do we do this? Well, our answer, as I'm sure you've already seen, is Genie ontology. So, Genie ontology um is our new solution to this context problem. It provides a unified semantic layer,
[16:32] uh which offers insight across all of the data you have connected to Databricks, whether it's in Lakehouse itself or it's connected via Unity Catalog connections to um external sources. So, I'll show a little bit how that works here.
[16:44] So, Genie ontology is able to extract and and store what we call ontology snippets from all of your existing sources, whether these are uh pieces of of Unity
[16:57] Catalog metadata that you've added uh manually in like in uh semantics and glossaries, other data products and domains you've created, or using partner sources like connections to Google Drive
[17:10] or or Word or any of these other connections we've we've made available. Uh Genie ontology is able to consolidate all of this information and provide it uh uh so that your agents can very efficiently um query this unified
[17:24] semantic layer. And the best part is that because Genieontology snippets are created um and evolve with your organization, you don't have to go in and manually adjust all of the semantics all the time as you
[17:36] scale and become more complex. But, we're not totally abandoning explicit semantics, right? So, all of your explicitly modeled semantics are integrated directly into Genieontology.
[17:50] This isn't a choice you have to make. So, if you've already done a bunch of great work on making metric views and table descriptions, that will just enrich and improve the quality even more. Um and we have some exciting new features here, business glossaries,
[18:02] multi-fact modeling, and we're also supporting the open-source um OSI protocol for for these semantics. And this is all fed into that onto rank algorithm, so you can provide the right context at the right time to your
[18:15] agents. Awesome. So, now we have this third layer. Uh we have the semantics here with Genieontology, and that again is combining learned knowledge from these sources, both from Databricks itself and
[18:27] from your connected sources, as well as the manual modeling in metric views semantic metadata, even um skills. Great. So, now Jess has this wonderful foundation, but she's wondering, "How do I actually
[18:39] serve this to my end users?" And so, for that, I'll pass it back off to Mo. >> That's right, Ben. So, as you may know, in any company pretty much worldwide, there are different end users for different tools. And usually there needs
[18:51] to be a mapping from tool to end user, and Jess is actually in the same situation. So, she needs to build an agent layer on top that serves these end users. They are BDMs and TDMs, business
[19:04] decision makers and technical decision makers often. And they require different interfaces to interact with agents. So, since I mentioned in the beginning, Jess went through many different technical transformations over the
[19:16] years, she wants to align with an agent layer to those stakeholders. So, number one, for business users, she actually uses Genie 1. Genie 1 packages up an environment and kind of off off
[19:29] offuscates for those users in front end that they can interact with on the semantic layer that Manhas had described. It's very easy and it's very straightforward to use and the entry bar is relatively low, perfect for any type
[19:43] of executive that needs quick, correct answers. Number two, there are a lot of data and AI teams within her company. So, she chooses Agent Bricks. With Agent Bricks, agents can be customized, the
[19:56] nitty-gritty can be adjusted, supervisor agents or agents of agents can be built, and those teams love that. They have the full control over these agents and they can basically on a developer platform build many agents interacting on the
[20:09] semantic layer. And number three, her company has contracts with OpenAI, with Anthropic, with other model providers, and those models want to access to the same type of semantics. So, she chooses she
[20:22] chooses Genie MCP here to give basically those models that required access. That helps her to overall get a check mark for all these end user groups. And they're super happy with that. Um
[20:35] they can build their individual UIs, whatever they prefer, how they want to get the data to be served. But Jess overall tackled four or three plus one, you can call it, challenges that have been aligned to these risks
[20:47] that we covered in the beginning. With compute, she covers the scale. With the governance, she covers the governance, so to speak. So, in this case, Unity AI Gateway covers the governance. Semantics, she covers the um accuracy
[21:00] problem. And with agent choice, she brings different types of um end user tools like um, Genie 1, the custom agents, um, or Genie MCP to the table.
[21:13] All right, I want to wrap this up. I want to really dive into the demo Ben had prepared. I think this is going to be super cool so you can see how this all works and actions out of the the slides that we went through here. Ben, do you want to share?
[21:25] >> Yeah, sure, cool. Uh, I guess I don't need that anymore. Cool, so we're going to work through a a quick live demo here. Let's see if I can get out of this. Uh,
[21:41] I need to find my cursor again. Oh, here we go. Okay. Cool. I think it's not sure is it Oh, it is showing it. It's just not showing it to me there. Yeah. Okay, cool. Can everybody see the two terminal windows I
[21:53] have pulled up? Okay, great. Okay. So, um, first what I want to show off is how cool it is to use Unity AI Gateway with whatever coding agent you want. So, imagine that,
[22:07] you know, I'm a I'm a more technical persona. I'm a software developer, a data engineer, um, and I love using these CLI-based agent tools. I don't want to have to go into uh, a different UI and context switch to
[22:21] get good data answers. So, one cool thing I can do is use this utility that we've created to use Unity AI Gateway with all of our, um, agents. And I can do UCode Claude.
[22:35] At the same time, I can do UCode Codex. And you see this is actually all channeled through Unity AI Gateway. Um, I'm going to say yes, I trust this folder. So, this is not just a regular Claude or
[22:47] a regular Codex. All of the model, uh, usage is being channeled through Unity AI Gateway where you can govern it completely. Um, so, in both of these, I have set up the
[23:00] Genie MCP, which will allow me to leverage that same ontology layer regardless of what agent I'm using. This is a super important issue with a lot of the customers I speak to. Uh different teams have different
[23:12] preferences for the agents that they're using. They try to come up with some sort of unified semantics layer. What this honestly looks like in practice is a bunch of random markdown. It's like skills that people have written locally.
[23:24] Maybe they have some GitHub repository with a bunch of markdown in it. Um and it's extremely challenging to scale that up across an organization. As the data evolves, the semantics don't catch up to
[23:37] it because it doesn't live together, right? They're inherently separate. Um so, they're excited that they can use something like uh you know, Genie MCP to access that same ontology,
[23:49] same consistent results from wherever they are. So, I'll just hop into Unity AI Gateway quickly and show you what this MCP looks like. So, very simple um and we'll actually be adding more and more tools as we go, but it essentially
[24:01] allows for an agent-to-agent interaction between your top-level agent and Genie. Um so, you can get those consistent responses from wherever you are. So, what I'm going to do here is I have a, you know, very basic prompt here.
[24:15] Um I'm going to just ask this to both of these top-level agents and hopefully they invoke the Genie MCP and we'll start to see Okay, there. Yeah. So, it's
[24:27] started. Look. Okay, great. They both decided to ask the Genie MCP. And as you can see, as it communicates, it will start to feed back all of its incremental responses and all of its
[24:39] thinking traces to the top-level agent, so you have full insight into what's going on. Um and I've actually run this in advance, so we wouldn't have to wait through the entire process, but here's a
[24:51] cool example. See, it's surfacing this to the the end user. It found this existing coffee model, um table. So, both of these have found the same appropriate semantic model,
[25:04] even though it's coming from different entry points and might have totally different context as an agent. So, they're able to identify the same the same thing and I'm just going to switch over
[25:16] to where I've already run this in advance, but you can see it's just showing all the incremental progress as it goes. You're not just like waiting for something to complete for a long time. Oh, this one actually already completed. Okay, that's great.
[25:28] Cool. So, you can see here this response with the top breakfast sandwich $2,494 last week.
[25:41] Let me just switch over here so I can show. So, yeah, this is the same thing. This one already completed, too. So, I have the same response in both of these. And you can see the link back into the
[25:53] Databricks UI. And what's cool about going back into the Databricks UI for this is let's see if it will let me click this. Oh, looks like it actually popped up on the wrong window. Let me just add it to this browser manually.
[26:12] I can see in the Genie 1 UI the exact same trace. So, all of this is centralized for your data teams. You can go and see no matter what entry point anybody used, you can audit the response, you can make
[26:24] sure everything looks correct in there, and all of your business users can access it from this UI as well. You can see that while it computed this response, it leveraged the Genie ontology.
[26:38] So, right here you can see what we call an ontology snippet. And this ontology snippet is identifying this coffee model as the source of truth for any New York City coffee shop sales. And it has a couple of rules about how
[26:51] different things are aggregated and summed up. And you can see that the author of this ontology snippet is Alex. So, we can see like, you know, how how much authority do we think this
[27:03] ontology snippet has? And you can see the onto rank algorithm has decided this is a particularly authoritative signal. And if we open Alex's domains, we can see, oh, okay, his his domains do seem to match with expertise here. We can see
[27:16] trusted assets he's created and the ontology snippets. Um and so, if you click in here, you can see that ontology snippet was accessed. And then also, if we click in here
[27:29] or let me do click in here even though this was a different chat, the same ontology snippet was used. So, in that way, you can access the Genie
[27:41] ontology from whatever surface you want um and get these consistent and audible audible responses. And so, cool. Now, I have this response. Now, what can I do with it? Um well, we're adding all sorts of business user
[27:54] tooling to Genie 1 itself. Um one cool thing we can do is say like, "Hey, write uh me up a report on this."
[28:06] And so, it'll start working away. Uh for the sake of time, I've I've already done this right here, too. So, just make this into a report. It creates an artifact. Let's see what it did here. Yeah, so it added um
[28:19] a number of different visualizations um and then it's all again cited here and using that same ontology source. And then one cool thing I can do is I can just share this like any other
[28:31] Databricks asset with any of the same Databricks permissions. It doesn't have to be what we found with, you know, a lot of these current agent platforms where it's either just shared uh to you alone or to
[28:44] everyone in your organization with a link. So, I can add my friend Chow if I want to. Uh And yeah, then he can take a look. So this is just an example of how we can
[28:56] unify from all these different access patterns a single trusted semantic layer across whatever your users prefer. And then you can use your explicit
[29:08] modeling in Unity Catalog to improve any semantic results you want to own that as a central data team, own the quality, and then um audit and improve as you go from all of
[29:20] your feedback you get from your various stakeholders. Um yeah, so that's that's my demo. Moe and I will be up here to like happy to talk through any use
[29:32] cases, any questions, um anything else you guys are working on, any problems you're running into in in your organization. So we'll be we'll be right up here. >> We're right here. We We have about 10 minutes before we all going to get kicked out. So just come up basically
[29:45] and Ben and I will will share the question. >> Yeah, thank you all. >> Yeah, thanks. >>

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.