Skip to main content

Getting Data AI Ready: Testimonial of Good Governance Practices Constructing Accurate Genie Spaces

Summary

  • Genie Spaces Success: T-Mobile built successful Genie Spaces by combining declarative metadata approaches with robust governance frameworks.
  • Four-Component Framework: The solution includes UC tables deployed on a robust framework, crowdsourced user feedback, LLM creativity balanced with human wisdom, and fine-grained access control.
  • Practical Governance Lessons: The video provides generalizable lessons for getting the most out of Genie Spaces deployment.

Getting Data AI Ready: Testimonial of Good Governance Practices Constructing Accurate Genie Spaces

Watch: Getting Data AI Ready: Testimonial of Good Governance Practices Constructing Accurate Genie Spaces
Talk by Arvindram Krishnamoorthy, Principal Architect, ML Solutions, T-Mobile ; Brian Schober, Sr Business Intelligence Mgr, T-Mobile
Genie Rooms have played an integral role in democratizing important datasets like Cell Tower and Lease Information. However, in order to ensure that this exciting new release from Databricks was configured as optimally as possible from development to deployment, we needed additional scaffolding around governance. In this talk we will describe the four main components we used in conjunction with the Genie Room to build a successful product and will provide generalizable lessons to help others get the most out of this object. At the core are a declarative, metadata approach to creating UC tables deployed on a robust framework. Second, a platform that efficiently crowdsourced targeted feedback from different user groups. Third, a tool that balances the LLM’s creativity with human wisdom. And finally, a platform that enforces our principle of separating Storage from Compute to manage access to the room at a fine-grained level and enables a whole host of interesting use-cases.
Intelligent analytics for everyone: https://www.databricks.com/product/business-intelligence
BI meets AI: https://www.databricks.com/resources/ebook/business-intelligence-meets-ai
Explore the only unified platform for agent systems: https://www.databricks.com/product/artificial-intelligence
See all the product announcements from Data + AI Summit: https://www.databricks.com/events/dataaisummit-2025-announcements

FAQs

What is a telecom lease portfolio?

The lease portfolio represents the data points that correspond to every cell site and all the lease data associated with those sites, constantly changing as leases go through their lifecycle.

Why is the telecom lease portfolio important?

The spend to lease all cell sites is tremendous and represents one of the biggest expenses for telecom companies, typically with 5-10 year commitments per site.

What are the four main components used with Genie Spaces?

A declarative metadata approach to creating UC tables deployed on a robust framework, a platform for crowdsourced user feedback, a tool balancing LLM creativity with human wisdom, and separation of storage from compute for fine-grained access control.

How does Genie Rooms democratize datasets?

Genie Rooms play an integral role in democratizing important datasets like Cell Tower and Lease Information by making them more accessible.

Full transcript

All right. Hi everyone. Welcome to our talk, Getting Data AI Ready. Um, please take a moment to look through this now probably familiar forward-looking statement. I'm the principal AI architect for the AI center of excellence at T-Mobile. It's underneath the finance procurement and supply chain organization. Also work under finance. I work on a team that supports uh critical infrastructure that includes sellite leasing uh back hall and utilities.
And I'm here today to kind of give some background on the data set we'll be focusing on which is the telecom lease portfolio. And so uh you think about what is a telecom lease portfolio? Um basically everywhere you go you have cell phone coverage. So that indicates there's a cell a cell site nearby. And for uh literally every one of those uh cell sites, there's going to be a cell site lease. Sometimes there's going to be more than one.
You're going to have um uh extra lease if you have to have like a generator or ground space going on. So what the lease portfolio is is essentially the data points that represent um these sites and all of these the lease data uh that goes along with that. Uh so this creates a data set and it's extremely dynamic because um as each one of these leases goes through the lease life cycle uh hits many different events. So the data is always changing.
For example, uh rent's going to increase. uh you're going to have um renewals, non-renewals, changes of ownership, and all of this needs to be reflected in the data set and available for the folks that manage uh this portfolio in order for them to do it properly and um keep everything, you know, running as it should. Um and why is this so important? Well, the uh the spend for to keep a network running and to lease all these sites is tremendous.
It's going to be one of the biggest expenses for any telecom company. Um, and as we mentioned, they all have to be leased for the most part. Um, and these sites are expensive to build. So, when you go into these, you're making a big commitment. Normally, it's going to be a five or a 10-year commitment. So, if you miss on just one key term, you could end up locking into a 5-year rent commitment, which is extremely expensive, and it could be for a site that you'd no longer need.
Um, additionally, uh, this information is relied on crossf functionally across different parts of the company that are very vital such as, uh, most notably like site development and the RF engineers. Uh, they use it as part of their planning to plan and manipulate the network. Uh so thinking about the um the users that are going to be really heavily involved with using this data to manage the portfolio um you're going to have a user base that has a very deep uh domain expertise and a lot of niche uh knowledge that you really only learn by doing the job.
It's very hard to emulate. Um it's kind of that human factor. uh they're going to be very strong strategic thinkers and they have to collaborate with all the other uh crossf functional teams to keep everything in sync and keep things uh you know running properly and along with all the projects that are being upgraded on the network. Uh so but what we find is a lot of times these particular users are not going to have the strongest technical background.
Um and in this case we mean that in terms of they're not going to know SQL to access the data directly on their own and explore it. So they become basically essentially dependent on either analysts or canned reports or dashboards uh to interpret the data for them and then they have to work within that. They can still get some good insights but they're not really exploring on their own. And that's where we we want to bring data fluency to them which is basically uh we see it as them being comfortable exploring the data using their own boundaries.
They're not locked into somebody else's interpretation. um any transformations that need to be take done already there and so that requires a gold standard data set and so that's something that is going to touch on a little bit more. Yeah, thank you so much for going through that user profile. Um before we continue, I'd like to just take a minute to talk about the data set that we wanted to democratize as well as some stats about it.
So in the beginning there was reporting chaos. Uh data was coming from many different sources and we didn't know which to trust. One report would say X widget sold, another one would say Y. By adopting a set of comprehensive data engineering standards, that is by modeling the business as a series of interrelated products and then getting stakeholders to be active in the life cycle of those products, we were able to reconcile the data.
But we still wanted to get to this end state where people could interact with it without needing specialized skills because as mentioned, many of these tables were fundamentally still SQL accessible. The focus of this talk will be lease beast. Um it's an effectially named gold level data set that has 500 distinct columns. It's comprised of four different data domains and exposes the endto-end life cycle for all leases.
Um you can think of upstream as basically what happens when you uh sign the lease. So that's the landlord, the date the building permit was approved and whatnot. And the downstream is basically things that happened after the lease is signed. So that's the rights to sub lease, the ongoing rent payments, various other related data sets. Um, and so what we've built is a consolidated picture for this uh leasing um leasing view.
And we don't need to have individual uh copies of that data set if we could just go to this one central point of location. Um all of these attributes make lease piece great for developers but unfriendly to domain experts. And that's pretty much the agenda that we're going to go through today. will talk about the proof of concept and how we were able to democratize the data and then I will go over the um governance principles that led us to be able to do so.
And so we talked a little bit about uh data fluency and bringing an individual to be comfortable in the data. Um being able to do that for basically all the user base of a data set or we think of as data democratization. um and just enabling them to be able to go get their own uh analyst grade insights from the data um without any assistance. And so there's some challenges that we face in order to get to that point. Um and one of those challenges, for example, is just empowering the user.
How do you get it to where you're not just giving them access to something, but they actually feel comfortable. Um they feel confident in the results and they're able to, you know, explore unassisted um on their own. Um also this data set is extremely nuanced after you know the telecom uh industry has gone over or has gone through tons of M&A and uh consolidation over the years. Uh that results in mergers of data sets uh system migrations and all sorts of things that cause nuances in the data that have to be accounted for.
And so that's part of what the data governance really helps out with uh to kind of smooth that out for the user. Um also just building trust in AI. So there's still a lot of folks that are skeptical or afraid to use it or um so you have to kind of show them the way I guess you so to so to speak. Um you also we also need to just basically uh provide data that is AI ready. Um and that's where is going to get you know a lot deeper into that here in the next coming up slides.
And all of this is just to kind of fall in line with our corporate vision which is to um uh make datadriven decisions across the company no matter who you are. And it takes you know improved data skills and comfort with the data and the ability to access it directly. So Arvin's going to get a little further into oh I'm sorry and so the uh the uh our solution to actually get this to uh the users uh based on the gold level data set is the least speech genie room.
And I don't know how many of you are familiar with the genie space uh in alter or in um data bricks in data bricks. Yeah. Uh but what it basically is it's a natural uh language uh it's like a chatbot but it takes your natural language and it writes the SQL for you. It does a great job of it. Um basically it's just the user is able to go in there with no code using natural language. It interprets it. It uses the embedded um data expertise that you train it on and it determines the query, writes a generates a whole new query, executes it, brings it back and uh and it does all that kind of stuff and it really uh makes it a very efficient way for somebody who knows zero SQL, zero coding to to grab the data themselves.
So, you know, they can take they they something comes in and they've got a list of IDs they need to research or um site IDs, contract IDs, whatever. they can take that list, paste it right into to the chat room and say give me these fields or or give me uh this type of data and it will go write the query, bring it back. Um it helps you explore the data because uh it even can do calculated fields, simple fields like summing up expense or uh it'll even do pretty complicated things like if you have two tables um you need to find the distance between coordinates, it actually will will do some of that geospatial stuff. basically anything that SQL will do.
You can you can get this to write that type of code. Um and also like uh it really helps people explore the data because you can chain prompts together. So you ask an initial question and it brings up another question and another question, another question eventually leading you to that kind of insight what you're after and that answer that you're looking for. Um and a lot of times um you know uh an actionable list of things to research or do.
And then um uh also it produces pretty nice visuals and the data bricks team continues to work on that and uh that function just keeps getting better but you can make a little bar chart line chart and all that kind of kind of stuff take it with you uh to do other things. Uh so for the next slide we actually I don't know if there are people here have or haven't seen the genie space uh in action uh but we've got a little demo.
We're going to mouse through it on the screen and I'll kind of talk over uh see if I can see it here. I'll kind of talk over what it's doing. So, this is what the Genie space looks like and go over some of the configurations. These are the tables based on the lease beast uh that talked about. Uh you can see the table comments. You can see the uh field comments. These all help guide the L&M. They're like breadcrumbs uh to help it write the correct query.
These are the general instructions. If you've ever uh designed a custom GPT or anything like that, it's very similar. You give it a persona. You you put a lot of the business logic in there. These are example SQL queries that um the LLM will find like uh questions and kind of pattern and build those queries based off of those. At the bottom there was the SQL functions. We'll go over to those here in a little bit. Uh that's just showing that the compute which is what gives you your permissions uh and runs the job.
And then now we're going to have a few questions just a demo of you know asking some real world questions people might ask um and getting some results. So here I'm asking how many leases uh expire in the next 5 years. And this is obviously demo data. So if you're from another carrier, don't get too excited. So you can see it's it's telling you what it's doing. It's thinking it's it's it's trying to determine uh the query, right?
Comes back. Oh, okay. This is how many leases are expiring in the next five years. It's fine. Everything that we did is still accurate. It's just the data itself isn't um correct. I just stripped it down and masked masked it. So, so now we come back and say, okay, aggregate that by region because maybe we want to see, you know, how many of these leases are expiring by region. Well, we get that answer and it's going to generate a nice little bar chart there for us.
Maybe we take that with us for a presentation or whatever. But that's not what we're really after. We're after some sort of action. Uh, so I'm going to ask yet another question that says, um, I'm asking it to basically count how many of these leases are with each vendor and then sort it descending. So maybe what I'm after is I'm making a plan on which of these groups of leases I'm going to negotiate first and maybe I want to start with the landlord that has the most leases.
Uh so I ask the question, it comes back there. There are the vendors, there are the lease counts with the biggest ones at top. So I can take that list and uh kind of get started on making my my plan. And that's just a simple example, but but really one of the coolest things about this is how you can just keep chaining. You can keep digging. people don't know SQL and they can just explore all over the data set as long as you have a good governed data set to work off of.
And that brings us to the impact. Yeah. So, I'm not going to go through all of these, but um we've definitely had some very positive impacts. Uh we we released it initially to one of the national teams um and um it's gone really well. It gives really good answers and we're getting ready to open up more to some of the regions and a lot more users and we're expecting that to be the user base that really drives a ton of usage and we plan to kind of start with kind of start small with some things that they can really trust and then start building more and more into this uh into this genie space.
Excellent. So the first principle I want to talk about is all about governing the most important asset of the space the data. How do we do this? through data contracts. Data contracts are really neat because they're agreements between the producer and the consumer. They do interesting things like guarantee the uh verif uh verify agreed upon guarantees and make sure that we have shared understanding of the expectations of the data.
While the concept of a data contract isn't anything new, the five that went to ours tonomy, lineage, DDL, operations, and business expertise um serve to quality control the 500 columns and give much needed context to the space. I'm going to talk about taxonomy in great detail and leave the rest of the as an exercise to the reader. So the taxonomy um is based off of the four different data domains that I mentioned earlier.
Those were finance, network, organization and location. When we think about each of these domains, you'll notice that there could be specialized tables within them and they uh the initial grain of those tables may not always map to the final composite entity that you might create like lease beast. For instance, there might be many parties to a lease and that's within the data organization domain. So you might have to find aggregate that to the most common.
Additionally, there could be many payments and you might have to aggregate that over many months. So in this way, a well understood taxonomy informs the choices you're making and the data contract binds them. If it doesn't, we can also make sure that uh we can iterate on the taxonomy so that the genie space selects the right columns that we need for those for that data set. Here's how everything fits together. So, as I mentioned, we're going to declare all of the metadata that we've had for our data contracts from uh definitions to stewards in a data contract and then we're going to deploy that in an automative way through something like data bricks asset bundles.
Data Bricks asset bundles are wonderful because they basically orchestrate the management of different objects within data bricks through jobs. That way we can determine what runs with what dependencies at what cadence at a very granular level which is super helpful when we're creating dozens of composite entities like lease beast. In this way we can make sure that only the approved changes make it to the table that the genie space uses as a source.
So I mentioned that we use um data bricks uh we use data contracts to govern this data and um the primary way that we do that is through YAML. Although you can use any data serialization format we prefer YAML for two reasons. One is that it's very easy to read and the second is that it integrates very effectively with tools like dbt and dabs. I already mentioned dabs but dbt or data build tool is where we take our business logic convert it into code write tests on that code in an ongoing manner to catch anomalies early and make sure that the outputs are trustworthy.
So um you could think of dbt assembling the finished product and dabs deliver them. This interoperability that uh YAML provides reduces the amount of context switching our engineers go through. Additionally, the uh further metadata from the data contracts like the lineage and stewards go into tags in uni catalog. The other place that um we find uh information from data contracts going to is column comments. All of which as you know gives the room much needed context on what our data means.
There are three key benefits to going through this approach of data contract oriented orchestration with DBT and DABs. The first is that it um builds trust between the developers and business users. Developers can rest assured that the only time that their understanding of the column comments change is when the business users notify them. Conversely, business users can rest assured that the repeatable audit of schema transformations make sure that developers can pinpoint changes from their data.
Additionally, it also clarifies the maze of technical support roles. In the original chaotic reporting world I mentioned earlier, logic was directly built into uh reports like Tableau and PowerBI. This was terrible because we didn't know uh who to reach out to when um information was uh wrong. For instance, if like a value was out of compliance thresholds because there were different owners and authors of that report with varying skill levels, uh how how would we find out who to get technical support from?
Data contracts make sure that we know who to reach out to for what purpose. Finally, it inspires platform thinking. By building the pattern once and uh executing it hundreds of times, um we can make sure that we build several hundreds of tables like we've um done at the Elevate um center of excellence. I wanted to leave you with a practical tip. If you've added several spa uh tables to your genie space and are having trouble getting it to understand the relationships within your data, then we highly recommend writing data contracts and especially taxonomies to help it determine the overall data quality in the long term.
Another lesson that uh data contracts taught us is that a successful column comment not only defines but also contextualizes. We learned this through investigating two columns. The original commencement date and the commencement date. Sounds completely different, right? Actually, there are subtle differences between the two. The original commencement date is when the lease actually uh is when the lease planned to begin from documents like the master lease agreement.
The commencement date is when the lease actually began. And although these definitions alone weren't enough to have the genie space select the right one 100% of the time, it took adding scenarios of when to use which that um we figured out when to uh select what we wanted. Next, I want to talk about governing um the implicit tidbits and terms your business users use when referencing their business processes. That makes a great basis for instructions.
And this principle of collaborative and targeted testing can make sure that our instructions are the best that they need to be. What we really needed was a good data store to collect all of this feedback. We needed something that was easy to edit, easily integrated with data bricks and captured state effectively. Fortunately, we were able to repurpose a patent protected tool called incubator that had all of these properties.
Here was the process that we followed. I don't need to elaborate on every aspect of this, but I would like to make two quick remarks. The first is that we began to achieve collaboration between um the business users and Mayan team um by being available to translate their answers into SQL statements. Additionally, we experimented with creating a separate genie space that just loaded the evaluation um and feedback that we were collecting in our incubator tables and exploring trends in a targeted manner.
Finally, all of this feedback would be summarized into instructions. What I'm showing you here is a special feature of uh incubator which is capturing version history at the row level. Because the agent behind Genie is highly probabistic and fastchanging, it was important to tie the date of value was changed to who did it and what the previous value was. You can see that when I asked it the question, how many what was the average monthly lease rent for the Houston market, it originally returned an answer of undefined when a certain set of instructions were available to it.
After some iteration, we were able to get it to return the right answer, redacted in this case, and that was only made possible because we were able to directly tie a relationship between the spaces configuration and the feedback on whether the answer was correct or not. Additionally, another valuable feature that incubator has is rowle attachments. By attaching the before and after configurations of a space, we were able to determine what was happening to anybody, especially to data brick support during the early stages of gene.
If you find it difficult to collect all of the question answer pairs from all of your users at once, this principle teaches us that we can actually flip that and be deliberate with our target uh with our testing. Specifically, you can spread your collection efforts over time and manage each user group's concerns independently. That way, the golden data set will naturally be aggregated in phases and you can manage competing requirements.
For instance, user group A was focused on lease statuses and loser user group B was focused on escalation clauses. the amount of uh business knowledge between those groups varied and so it was totally acceptable that um certain sets of users require more rounds of testing than others. So finally going through this somewhat arduous scientific process has three benefits. The first is that it allowed us to test question variations to make sure that the room did not overfit.
This was especially useful before the benchmarks feature was eventually released. Additionally, it teaches us the importance of including creative and external ways of bringing change management to the tool that you're evaluating, even when the tool that you're you're currently using doesn't natively support it yet. And finally, this process was transparent enough to empirically determine the exact minimum set of instructions needed to answer all of the users questions.
In our case, that ended up being 19 textbased and 30 SQL based instructions. So now that we spent all this time talking about the data, how do we make sure customers can get their hands on it? That's where the principle of auditable access management comes in. Essentially, we should treat CRUD concerns on the data separate from compute concerns on the processing resources. Think of it this way. Someone who works with leases just needs to be able to read the most important information from a lease and track its progress.
They don't much care how that reading was delivered to them in terms of the compute type. We could use data bricks SQL warehouse which is wonderful because that is how genie is also built off of but if somebody was a domain expert and love using Azure synapse that should be an option that's available to them because our team manages the provisioning of these and other infrastructure it was important for us to automate the tracking of both of these requests independently the old process that we were following was manual slow and did not scale very well and that's why we created provisionator represented by this magenta card reader It's another patent protected tool and it works by creating security groups and then translating and syncing the permissions on where the data lives Azure blob storage in the form of delta files to where data is presented or uni catalog that's represented by this CRUD card.
Once those security groups have been created, we can then use any compute type that's compatible with Delta to uh read from that uh read from that data set represented by this compute card. But access to these data sets were highly sensitive and so only authorized users were allowed to see them. By referencing the same stewards we used in the data contracts and including them in the approval gates do data stewards were able to seamlessly audit and approve requests.
Another piece of practical advice is that if you're finding it difficult to have um active involvement in your data stewards for your data and AI products, then we recommend that you involve them in every uh step of the process to make active decisions. That way, they'll improve their lit data literacy and increase trust in the overall process. But that's not all. This added security brought a lot of new experiences to the space as well.
For instance, this project was originally scoped for the national team. That means that the audience was able to see leases from all four regions. We soon had to support a new scope which was to make sure that only um engineers from their corresponding region can see those leases. Region level CRUD security groups easily came to our rescue here because we realized that we could use membership to the group as a rowle filter.
That way we didn't have to create unnecessary genie spaces or data sets and just use the one that we um we originally created. It also helps with agile and collaborative development. While I was on PTO, was able to continue writing resources to his space that he'll elaborate in a future slide and persist them securely to his location. Because the expertise to deploy Genie Spaceac's um primary attribute, the lease beast, lay within our team, we would be the owners of that CRUD group. and his team's expertise lay in creating meaningful analysis reports and they should be the owners of that resource.
But the room was made better by the collaboration of both of our teams and the principle of auditable access management allowed that to happen. Finally, we can model the genie space as um a combination of one compute group that acts as its brain and one or more CRUD groups that determine what data the user is allowed to interact with. We found that this pattern of creating a many to many relationship within a user's given access profile is the best for creating a world where teams across your organization can come together and create meaningful, purposeful and secure genie spaces.
So typically when you go to this chat room into the D gen Genie space and you ask a question, it's going to uh interpret your natural language um uh look at the general instructions, look at all the example SQL you give it and then it's going to generate a query. Um there may be times however where you want it to return an expected result. And so that's what uh the SQL functions are for. And these are predefined code blocks that you put into publication and you can add them to the genie space and then you add to your general instructions and help it figure out when to trigger this function.
Uh once a genie space triggers the function, the function takes over. It's not generating a SQL a new SQL query anymore. It's executing that function. it will take the inputs such as the contract ids or site ids whatever kind of u unique identifier you have and put it as an array into that function and return back a table and so why would you want to do that? Um, so there could be instances where you have uh legacy reports or existing reports that have been developed uh by and contain a lot of the business knowledge.
Um, and maybe you you want those reports. They're good reports. They don't need to be redone. So you can have it trigger a function that basically writes that report back for you. Uh, or possibly you have extremely um complicated calculations that need to be done the same every time. um or extremely uh not extremely but complicated joins that you you want to make sure that they're done in a certain way. And so that's when you might want to use a function to ensure that you get the same result every time.
And so I can show just a couple of examples of where we use functions in our in our genie space. On this diagram here, you can see um starting on the left, the user uh prompts the genie space. it goes into the orange uh decision uh di part of the diagram there and it's gonna that's where it's looking at the general instructions the SQL queries determining do I use a function do I not use a function if it is if it does isn't going to use a function it will go down it'll write the query execute the query and do everything like it normally would once it triggers that function however it's like I said it's going to go off and it's going to follow the instructions it's been given there and so a couple of examples uh that we did is uh we have one where it goes and retrieves an expense history for a lease.
And this is useful because we didn't want to put the whole gigantic payment history um table in there. Plus, we wanted to make sure it it brings back a certain view every single time. Uh so, we wrote a function that handles that one to many join and brings back what we want to show the users. Um another example is uh we have some teams where when they receive certain types of work, they have to do a certain amount of swivel swivel chairing to basically create a view for themselves. they're going uh to this system, this system, and this system.
Bring back these data points, and then they can move forward and do their job that they're trying to do. So, we wrote wrote those views into a function where now they just ask, hey, give me um you know, this view for these leases or these sites or whatever it may be, comes back and it's right there for them uh the same way that they can expect every time. Thank you for going over that final principle. Now that we've learned all four, let's take a quick peek at a road map to see if we can apply our new understanding.
The first thing I want to do is to add a new table to the Genie space, specifically this list of milestone triggers. These will include things like the construction statuses, the catalyst for those statuses, and other related data. Because it's a brand new table, I'll want to add a brand new data contract for it. And so that way, we'll be able to define the column comments and other related metadata so it understands what those relationships mean.
Additionally, I have another requirement for this room which is that it needs to support multi-turn conversations. Currently, the other available option curated functions does not yet support um including the question context in its memory and therefore data contracts would be the right place for this use case. I also want to expand the user base to other national teams because it's a brand new set of users. they'll all have their own corporate knowledge and um expectations on how the data set should perform.
Therefore, I'd want to use the principle number two, collaborative and targeted testing to make sure that we can write the best set of instructions, tweak any of our existing ones, and make sure that we prevent any regression in behavior. Finally, I want to explore the delivery of lease adjacent data sets like utilities and back haul. These include the primary consumption of what the cell tower goes through and it's the next logical step for analysts.
Because we're still in the experimental phase, we don't yet know what sorts of joins we want to write. So we can leverage the um principle number four is um balancing expected reports with AI creativity to bypass a lot of the cardality issues until we're ready to write the correct foreign key relationships in our data contracts. We also just want to enable the users to retrieve a report they're already familiar with and is guaranteed to have the right answer using natural language.
Eventually, once enough trust is built, we can then graduate this principle to principle number one. So, we can see here that these principles apply no matter what stage you're in from experimentation to expansion. There are two key takeaways I want to leave you with. The first is that this MVP succeeded due to the great collaboration between data bicks's powerful agentic assistant and the wellquality data that we were able to provide it with.
That quality was assured in large part thanks to the four principles. Those were data contracted oriented orchestration which allowed us to teach the room what our data meant, collaborative and targeted testing which gave us a scientific process to determine the best set of instructions. auditable access management which used security as a backdrop to bring new outcomes to the space and finally balancing AI creativity with expected reports which brought deterministic and familiar features into a probabilistic generative world.
These principles are what broke down the accessibility barriers between our gold data set like lease beast. The second takeaway is that although the specific governance tools have been in the works for many years, we really sit on the shoulders of giants here. The principles themselves remain highly automatable. Here's how. First, you should learn when and how to apply these principles and see how they fit into your data strategy.
Next, you can make templates and make sure that they uh repeat in an automated way. If you do so, you'll make sure to gain reporting productivity and prototype much more uh much more effectively. This is especially important because Genie is a reporting force multiplier. In our case, as mentioned, we saw as much as a 30% increase. More people were able to perform more analyses more frequently thanks to Genie's self-service nature, and uh they also didn't have to rely on a separate analyst team's intake.
Those analysts could then in turn be freed up to tackle the next layer of complexity. With that, that concludes our presentation and um please remember to complete the session survey at the end.

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.