Building Trustworthy AI at Scale: Takeda's Self-Service AI Foundation

Summary

Takeda, a top-three global pharmaceutical company operating in 80 countries, built a self-service AI foundation on the Databricks Data and AI platform that enables 10,000-plus employees to develop, deploy, and operate AI agents with enterprise governance.
Infrastructure provisioning was reduced from 60–70 support tickets to a single one-click approval process completed in 20 minutes, eliminating the bottlenecks that previously slowed AI development teams.
The platform rests on four pillars: self-service AI engineering with Mosaic AI and Agent Bricks, a control tower for FinOps and Responsible AI monitoring, an automated policy enforcement agent, and a marketplace for secure agent discovery and reuse.

Building Trustworthy AI at Scale: Takeda's Self-Service AI Foundation

Takeda, a top-3 global pharmaceutical company, built a unified AI foundation on Databricks that enables 10,000-plus employees to develop, deploy, and operate agents confidently with enterprise governance. Learn how they moved from siloed pilot projects to an enterprise-driven agentic journey, implementing Responsible AI principles at every stage.
Discover the technical architecture behind their four-pillar approach: self-service AI engineering with Mosaic AI and Agent Bricks, control tower dashboards for FinOps and RAI monitoring, policy enforcement via an automated police agent, and a marketplace for secure agent discovery and reuse. See how Takeda reduced infrastructure provisioning from 60-70 tickets to one-click approval in 20 minutes, and how they integrated identity management with Immuta MCP, automated ontology generation with Genie, and connected operationalsystems via native AMQP and MQTT connectors.
🤝

Chapters

FAQs

How did Takeda reduce infrastructure provisioning time for AI development?

Takeda reduced the process from 60 to 70 support tickets to a one-click approval workflow that completes in 20 minutes by building automated self-service provisioning on the Databricks Data and AI platform. This eliminated the friction that previously prevented AI teams from starting new projects quickly.

What are the four pillars of Takeda's AI foundation on Databricks?

This video describes four pillars: self-service AI engineering using Mosaic AI and Agent Bricks, a control tower for FinOps and Responsible AI dashboards, policy enforcement through an automated agent, and an agent marketplace for secure discovery and reuse. Together these pillars allow Takeda's 10,000-plus employees to build and operate AI agents with governance built in from the start.

How does Takeda implement Responsible AI in its agent development process?

Takeda embeds Responsible AI principles at every stage of the agent lifecycle, from development through deployment and monitoring. Their control tower dashboard tracks Responsible AI metrics alongside FinOps costs, and an automated policy agent enforces compliance rules to ensure agents meet governance requirements before reaching production.

What is the agent marketplace that Takeda built on Databricks?

Takeda created an internal agent marketplace as one of the four pillars of their AI foundation, allowing teams to discover, share, and reuse agents securely across the enterprise. This video shows a demo of the marketplace portal interface, which surfaces metadata so consumers can find relevant agents without duplicating work already done by other teams.

Full transcript

[00:09] If we are here on third day, that means we are really here to learn something new. What's happening within the data bricks ecosystem and then like how we can implement back to our company and help our company to grow. So, thanks for staying on third day. It's It's really like a nice crowd which will
[00:24] survey the third day. So, we'll have a survey. It's in our data bricks app, so any any feedback or comments about the session, you can go and fill the survey.
[00:42] So, today we are here to say how we use data bricks and agent bricks and data bricks foundations to make it a self-service way of bringing the data and AI infrastructure up and running in Takeda.
[00:57] So, why it is important is also like talking about what does a trustworthy means, what does a foundation means, all those things. So, before we jump into the topic, I just want to quickly like introduce our company. So, I'm from Takeda. My name is Dilip Shankar. I'm
[01:13] the head of data and AI engineering. So, Takeda is is been like there for 245 years and we recently celebrated our birthday on June 12th. So, it's I'm proud to say that part of the 245
[01:30] years, last 5 years I was at Takeda. So, it's it's been like long journey and then like if you see here, Takeda is one of the few companies where they they are like everywhere from the therapeutic side of things, right? Neuroscience, rare disease, plasma. So, within the entire
[01:47] world like plasma, very few companies you can count on. So, Takeda is one of the few companies, I would say like top three company, right? So, we have a big PDT business And then like if you see the achievements out of the 80 country we
[02:02] operate we are almost like in top in 25 countries we are the top employers. Right? So, there are some open positions so feel free to apply. Good. So, like any other company we are also like starting with our data foundation
[02:19] transforming to the AI foundation. When when we started our journey we started with our data foundation that's called the enterprise data management foundation and we started with all the good things about data product MDM and then like scaling the data engineering pipelines
[02:35] to have the right metadata, right data steward, all those things have been as a base that I think this is a very similar journey that every company will go through. But what we started to do in last 1 1/2 years with AI, GenAI, all those things now with agent AI coming
[02:52] in. So, we thought like let's start implementing into the real use cases. I'm proud to say that we are like one of the few companies where we are submitting the regulatory documents to the valid dossier that we submitted to the FDA for approvals is now agentic and
[03:09] we are in production. So, now like whatever we are now is focusing on the use cases and then like what we did in last 1 1/2 years is what we see in the now and we have been like delivering the use cases one by one with starting from
[03:25] GenAI now all the way to the agent AI. So, we'll cover like what is AI, GenAI, agent AI, all those things in the upcoming slides. So, where we are going is what we are talking about the scale.
[03:40] Right? How we can make this as a full full proof ecosystem where we can adopt this use case driven agentic journey to the enterprise driven agentic journey. So, if you see right, any like any other
[03:55] company, we we also have like a lot of ecosystem partners, hyperscalers that we work with right from ServiceNow, Salesforce, SAP, and everyone is having like their own agents. So, how we can tackle this
[04:10] entire ecosystem with agent way of doing it is what we are trying to build the bottom layer. If you see the bottom layer there, so we are starting with our uh FinOps, and FinOps is what the then then need of the hour, right? So,
[04:26] agent we have we have been like spending a lot. How we are able to tell the story from the FinOps angle is where we wanted to measure the success of the agent. So, we have been talking about the cloud FinOps and uh AWS uh cloudability and a lot of things that we
[04:42] have an ecosystem, but how to measure the agent spend and agent cost, that's where the double down on the FinOps may going to be. We'll be like showing a live demo like how we are tracking that agent cost because it's easy to spin up a agent and make that up and running,
[04:59] but really like how you tag it, how you make the consumption to be like released by the value, right? That's what That's the challenge that we'll be like showcasing part of our demo. So, another thing is uh identity management. It's very easy to say in the after world uh intra world, identity
[05:15] management can be taken care, but now we are talking about agent to agent uh identity and authentication and authorization mechanism, how we can make that happen. So, we use uh a tool called Immuta for uh fine-grained access. So, we have a Immuta MCP is hooked up into
[05:31] that data and agent uh authentication mechanism. So, you might be like seeing the booths in the the floor hall for the Immuta. So, it's it's pretty much like working good for us because now we are able to impersonate by the agent, and the agent is able to
[05:46] get that uh authorization or authentication done by the MCPs with the emitter, right? And then like we were talking about the control tower. So, control tower is another interesting part. We are all talking about data AI, data governance, data policies, all those things. It will be like very good
[06:03] to see in the paper and SOPs, but how do we implement is something which we have done. And then like we'll also like showcase part of the work demo. And last but not be least like we saw the Genie ontology, data ontology, all those things is been also been a a prime
[06:20] investment from our side from Takeda side. So, before the Genie ontology was been like announced yesterday, we had created ontology bot where we can create ontology for any any line of business. For example, it will then be like FinOps means the finance domain or like
[06:37] within the value chain of pharma in the commercial domain. Or you take the R&D, right? So, we have like public ontologies and we have like for example data contracts. So, we have a data contracts and procurement ontology that we built from scratch. If I want to say like
[06:52] procurement team, go and build ontology, they'll be like saying, "Okay, bye bye for you if you're not going to use this agent." So, what we did is we make this ontology bot to help that entire data ontology to be like built for the procurement as a
[07:07] team as a function. And then like we make those SMEs to just come and validate if this ontology makes sense for them. So, they just come here and then like edit it. So, then that's where we'll be we have been very much excited to see this ontology coming from Databricks as a investment and it's been like
[07:24] going to be a really like a game-changer for the people who doesn't understand the ontology. Because ontology is not a new word that we are talking to the industry and then like it's it's been there for few decades, right? But why it is not in the mainstream? It's hard to maintain. It's hard to generate. There'll be like
[07:39] SMEs that we we have to even like validate the ontology. All these things are getting automated in this ontology board and whatever the data bricks is also bringing to the table is going to be like fascinating for us.
[07:54] So this is another myth I wanted to break. So here we always talk about agent reuse. Can we reuse my our agent? Hey, did you build some agent? Why don't you share that agent? I can start reusing that agent. It does not a API. I want to reiterate agent is not an API.
[08:11] If you want to use an agent, you need to go through all these layers. Right from memory, identity and identity and authentication, security management, infrastructure, and then like the data data connectors. Like whatever you did I you are connecting to IQVIA. There may be like another another vendor coming
[08:27] and giving the same claims data from the Commodo, right? So all those things are going to be the delta which you need to see. And I'm I'm calling coining myself as a agent package. Maybe like industry will come with some other tool when it is going to be like getting mature.
[08:43] So this agent package is what I wanted to get as a takeaway from this slide. If you if someone is saying that they can reuse your agent, then you need to think about all those parameters. You can't go on it is not LLM. It is not something which you can go on and then like say I'll just go and reuse it. Yes,
[09:00] you can do it for the agents like creating a PowerPoint, the common skills, right? So that's where we are we are segregating this language new language segregating all the skills. Right? Those are the things which we we call we say as a common skills. But when you build the agent, that means
[09:16] it's a specific goal that you built for. If you want to make that autonomy and act for the purpose which you built, that means you need to do all the steps what have been like described here. So if someone is saying coming and saying to you that go and use my agent,
[09:31] think back of your mind these are the things I need to go and reconfigure or configure or customize for my use case. So, we in Data we are making something called golden path. So, if if the large footprint within our
[09:48] ecosystem is Data Bricks and then there will be like some people using the Salesforce or some people using the SAP dual agents and agent force. So, the golden path what we are calling here is for Data Bricks the agent package which I showed, right?
[10:04] It's like It's a little bit of a different way to illustrate. But Mosaic AI and then like agent bricks and other technologies which have been helping us to build the agent, this is what the golden path for the Data Bricks as a work stream. And I would probably say that 60 to 70%
[10:20] of our data footprint is in Data Bricks. Other footprints are still like in transaction footprints and all those things. So, slowly we'll see how the lake base is gaining impact.
[10:36] So, in the initial slide I was talking about the control tower and RAI, right? So, RAI people think, "Okay, it's it's a fancy thing which I can go and implement. I I I make the rules for the entire company. I want to implement the RAI." You can document like how we have tons of SOPs in our ecosystem within
[10:53] pharma industry. Similarly, RAI principles you can create keep on creating some documents. But if you want to really implement the RAI, it goes from the infrastructure and then like it goes in the build phase of the agent and also you need to have a continuous way to monitor the RAI rules.
[11:09] So, RAI rules people think like, "Okay, if I want to say something about the topic safety." That's a simple example of RAI rule. The topic safety within the HR world, a topic safety within the claims is going to be totally different. So, if you if you want to really like make the topic safety to be like handled
[11:25] well in the monitoring phase, then you need to part make it as a build phase itself. You need to configure that in a proper way that if there is a topic safety if you want to implement for a particular domain, that should be like really coded for it. Right? People think like, "Okay, data bricks will solve for
[11:40] me. SAP will solve for me." Everyone will solve for you as a technology, but where you want to build and how you want to build is what this layering is showing to you.
[11:56] With that, I'll give to my partner here. So, Hardeep, I'm introducing Hardeep here. Hardeep is our vendor partner from the US. And the SSBI helping us to our building this data and AI foundation for us. With that, Hardeep, I'll give it to you. Thanks, Dilip. Hello, everyone.
[12:12] So, thanks, Dilip, for walking through what we have built from a strategy standpoint. But if you take it to a technical implementation standpoint, so this is how we built it. Right? So, the agent package definition which Dilip showed, right? We have this agent package. Now, if you if you want to,
[12:29] you know, set up this agent package from an infrastructure standpoint, so there are two different flows to it, right? There is a producer flow, and then there is a consumer flow. The consumer flow, we are talking about agent reuse. So, the consumer flow always starts with an agent marketplace.
[12:46] So, the marketplace which we have becomes your starting point to see if you're building a use case, do you have, you know, the agents already built or would you be creating it from scratch? So, that's that's your starting point to see if your agent exists. But if not, then you are a producer. Now, in in case
[13:02] of a producer, what we did is we created a self-service um form or a or a user interface where you can simply request for your own infrastructure for your use case. And then then there are there are some prerequisites to it as well. So, not
[13:18] everyone can come and fill in the form, right? So, there is a AI governance committee. So, you, you know, within Takeda, you would have to you through that committee. You have to get approval on on the type of use cases you're building for your you know for your specific areas. Once approved, once you have a use case
[13:35] ID registered, only then you would come to the self-service form. So, there are there are some guardrails we have applied. Once you come to the self-service form, you will give your metadata like you know, if you're building building an use case, you you would have your you know, some cost center associated to it, which department it belongs to.
[13:52] Um you know, what type of infrastructure you would need. Would you need a cluster? Would you need vector databases? Would you be connecting to MCP? Is it a simple to solve a problem? So, in that case we're using you know, very intuitive offering from Databricks where you know,
[14:08] you can come in and use agent bricks. But if it's a complex scenario to solve, now we are also using Mosaic AI where you know, if you want to code your agent, if you want to integrate your cloud CLI or Genie code now, right? You can code your agent in LangGraph or
[14:24] any other tool of choice. So, so that becomes you know, so you select all those information when you're creating the infrastructure. It goes through approvals. It goes to your use case owner who who approved your use case. So, it goes to couple of approvals and once it's
[14:40] approved, all set to go, all set to launch, we use you know, we have created microservices in the back. We have we are using Terraform for provisioning all your infrastructure and everything is driven through a product called as harness.io. So, it it creates your you know, it runs
[14:56] your pipeline. It it provisions the agent infrastructure for you. And it also provisions a Databricks notebook which have boilerplate code where you can go, it's ready to use, ready to code your own you know, let's say LangGraph agent. Okay?
[15:12] So, that's that's it. Now, the the beauty of giving you a boilerplate code is it already has RAI you know, integrated in your code itself. So, what we do is uh we have used Databricks, you know,
[15:28] workbook permissions. So, there are few pieces of code which cannot be changed, which which has all your, you know, Takeda specific RAI guidelines in it. Like, it could be um you know, your PII information. How do you check PII or fairness, if your
[15:44] prompts What are your prompts? Is your output grounded? All those information goes as a non-editable, Takeda specific, hardened rules. And then there are some other rules where the user can play around. They can create their agents. Now, this all happens in the dev environment.
[16:01] We do have a separate flow when um if you're ready to if you have tested, you're okay with your uh agent, you take it to the higher environment. So, we have a promotion promotion flow where we have you know, your LLM judges involved. So,
[16:17] in your dev environment, you you you know, you can run your own code. When you're promoting, we have LLM judges. So, um RAI is is not just in development. It follows through the life cycle of what we have built.
[16:35] Now, as I mentioned, there is AI engineering. So, so there are there are four components to what we have built. It's already live. We deployed it last quarter. It it it has great adoption and and as the topic of our the title of our session, right? Making
[16:51] you know, trustworthy and scaling trustworthy AI. This is how we do it. Okay? So, we have AI engineering where, you know, you have self-service, you know, form where you can request for your agents. Um you know, it has your agent package been built. It has, you
[17:08] know, it has inbuilt prompts you can use. You have your your memory managed there. Um you have, you know, governed tools which you can use. Um you have governed MCP. You know, there are some rules which we have business rules which we have applied of when do you
[17:24] can have an MCP server versus when you cannot. So, all of these rules are inbuilt here. And then we scale it through serving endpoints. Um I know with this latest announcements of AI gateway we would, you know, we would uh have to take it to uh you know, the the Unity AI gateway. But now we we
[17:41] at least for this one we scale it through uh serving endpoints. We have your agent registry and everything orchestrates through Harness which is, you know, Kubernetes based um and I mean it it's able to call Kubernetes for your microservices and and Terraform. Uh it has Terraform connectors to provision
[17:56] your package. Uh this is the second part of your uh you know, the the four uh pillars which we have. Um so, this is two parts. You have agent control tower and then you have RAI. So, we have discussed about
[18:13] RAI. Now, where do you track this RAI? So, if your agent is is violating any rules or if your agent is, you know, hallucinating, if you know, if if it's giving you if it's giving you unfair responses, where do you track this? So, we built a control tower dashboard where
[18:28] you would be able to, you know, look at all your agent parameters on how it is performing. Plus, we have built a FinOps dashboard at an executive level to see uh you know, overall how many agents you have in your ecosystem and how much is the spend per department.
[18:44] So, we have per department per use case. So, there are there are beautiful uh you know, drill down capabilities which we have built uh as part of this this control tower dashboard. And it also shows you your RAI uh constructs. One interesting thing which we did as part
[18:59] of uh implementing RAI is So, we want everyone to use this umbrella across enterprise. It's an enterprise offering, right? But, you know, if if someone goes and, you know, they does not follow the traditional this
[19:14] this traditional route of getting your use case approved and implemented through this self-service you know, whatever we are building. If they build a if they build something in silos, if they are using models which are not approved by Takeda. So, what we did is
[19:29] we built something called as an enforcement agent or police agent. So, what that police agent does, it runs continuously. It scans all your notebooks. It scans I mean, there is they're scanning at your infrastructure level like if you have right policies,
[19:46] right cost controls, budget policies, right tags and everything. And then it also scans if you're making rogue API calls, if you're using your own tokens to connect to OpenAI or or to Anthropic which are not approved by Takeda. So, it it captures all of that.
[20:02] And and the the the beauty of it is it it will send you it will create incidents. And then there are some mandate which Takeda has in terms of you know, what is allowed versus what is not. So, we also you know, introduced a concept called as
[20:19] kill switch. So, you know, if if there is something which is not followed based on Takeda guidelines, we kill the agent and we you know, we we send a notification that you know, the agent you're running is is not doing what it is intended to do. You can
[20:35] make modifications and you know, you redeploy. So, that's that's two important concepts we did as part of RAI. And then is the marketplace. So, this is your consumer journey. So, this is where you'll start
[20:50] you know, consuming your your agents. You'll start searching for it. You you'll start discovering your agents. Now, you know, we can take one example uh So, let's So, so let's a simple simple PubMed example, right? So, we created a
[21:06] PubMed vector database for market analysis. Now, 2 months down the line, no one would have recreated it. With marketplace in play in in in in there, you would end up searching for it and end up reusing it. So, 2 months down the line, it could be used for research
[21:22] analysis, your uh you know, other other use cases which uses your same vector database. So, you don't have to reinvent the wheel. So, it's it's so at Takeda, the uh reuse is numerous slogan, it's it's measurable.
[21:43] All right, let me let me walk you through a quick demo on how did we implement this?
[22:00] Yes. So, this is your the home page of the other the uh product we built uh within Takeda. So, this is the agent marketplace and just a disclaimer for this for this demo, we you know, created a sandbox environment. We don't wanted to show all the agents
[22:15] which we have in Takeda. So, it's a small representation of what we have so far. Uh so, it you know, we have a lot of features we built in this. It's it's a React bit based portal. You have uh multiple agents. You can see the uh status of each agent if it's upcoming,
[22:31] if it's live. You also have an option to uh you know, make your agent private. Let Let's say if it's, you know, uh dealing with some sensitive data, you don't have to you cannot reuse that agent. We have all those features built in here.
[22:49] Within each uh you know, agent, you can also look at, you know, what was the agent, what is the intent of the agent, what is the definition, what's description. You can see that the agent details. It's a summarization agent. You can look at the info of the agent, when it was created. And you know we have some unique IDs uh we we're using
[23:04] for the agent. Uh you also have RAI scores. This is interesting. It shows you all the you know all your agent behavior. It gives you a confidence if you're able to use you know reuse this agent or not, right? So it tells you if it's breaking
[23:20] any PII, if if if you know there are any jailbreak uh incidents happening on the agent, if it's hallucinating, you know if it's biased, if it's fair. All of those things are you know are are uh you know year on uh the marketplace itself. Plus you can
[23:36] look at all of these uh in control tower, too. So this is your marketplace. You know if I want to reuse this agent, I'll simply you know proceed and and fill in some metadata. Post approval, I'll be able to use this agent. Now let me take you to
[23:53] the uh the control tower. So at an executive summary level, you would be able to see all the cost which my you know all the agents, all the use cases have spent so far.
[24:09] And since it's a it's a representation, it's it's it's a small cost, it'll show up here. And by the way, all this data uh all this registration is happening through your MLflow. Your models are registered through MLflow. Traces and everything is through MLflow.
[24:25] Um and then all this cost data, we are pulling up from your Databricks system tables. Uh you know everything you know Databricks has been a great great partner to uh help us implement all of this.
[24:43] It's taking some while here to load. Yeah. It tells you the cost. Yeah, it tells you how many use cases you have. Uh it tells you the cost variance uh you know in the last how many days if you want to configure last 30 days, last 20 days, last 6 months, whatever. And then it also gives you how many
[24:59] agents I have across different business units within Takeda. Uh and it also tells you how many AI sessions my agent is using and what are the current live sessions. Uh at a use case level you can drill down from here. You can look at the use
[25:14] cases. You can look at the use cases. You can, you know, look at your you know, individual agent performance in terms of uh your your RAI scores. It comes in here as well. Um now if you see if if you get to an
[25:31] agent where you don't find it here, you have a use case where you will be building an agent, then comes your you know, your AI engineering which is, you know, uh for your, you know, AI engineer persona. So, you come to
[25:47] your AI engineering. So, if you if you notice we have different tabs here. One is the self-service form which I was talking about to create your uh infrastructure and your the the the agent package for your agent. And then we have different you know, features for your CICD, your your
[26:05] you know, your promotion of your infrastructure from dev test prod. So, we follow a typical I would call this uh AIDLC. I won't call it SDLC. I'll call it an AIDLC cycle which we follow here.
[26:22] So, if you go to create new project, so these are all the metadata we capture before creating the AI package. So, your cost center, APMS ID limit specific to Takeda is uh you know, your cost center, what's your use case ID, which environment you would be using, uh you know, business units. And then comes your
[26:39] you know, your model selection. So, you know, we do have uh some limitation in terms of which model. It's it's not open. It goes through a you know, a risk assessment before we approve any model. So, not all models are approved. Um so there are chat models or you know
[26:55] if you have summarization, if you want to create vectors, you have embedding models as well you can select. You have a list of models and embedding model to choose from. Uh then you come to you know um your cluster creation, SQL warehouse creation. We kept it very simple. Uh we we don't
[27:12] uh specify you know we just encapsulated everything in a small, medium, and large. We have a definition for what is small, what is medium, and what is large. We did we made it very simple uh and I depend and you can also modify your projects. If you feel your computer is less, you can increase it and so on.
[27:29] And then uh you have vector databases, uh MCP servers, you know based on the selection you can select uh you know how how do you want to use MCP? Like if you have agent, you have any uh vector database, or if you have any Unity Catalog function you want to
[27:44] expose through the MCP, you can do it right from here. And then there is vector databases. If you select yes, I need access to vector databases, you can select as many vector databases as you want. And then you know you will just give the location of where is your vector
[27:59] database, which catalog, which schema, and and what's your exact uh index you can select from here. When you submit this, goes for an approval within 20 25 minutes, boom you have your project ready to use. So you don't have to wait for you know months to get get into the system. 20 25
[28:16] minutes, you should be able to you know run your project. The why why we are building this ecosystem is to collect the metadata because agent breaks and all the other hyperscalers we we have the basic things to be like up and running. Click one of the request list.
[28:32] But what is not going to go away is your cost center, is your application owner, and the metadata that you want to collect before even you spun up a project, right? So this is where we want to make sure that we have a one-click approach to make this entire infrastructure up and running for our project which is getting
[28:48] approved in our AI governance console, right? So, when you see here creating three octa groups or develop development versus test versus production like that. Octa group is not going to go away or if you're using an intra intra group is not going to go away. Similarly, like if you see the other infrastructure if you want to
[29:04] orchestrate through airflow then we we need the infrastructure. So, without this automation we used to have like at least a 60 to 70 tickets and with multiple owners, different different platforms. All the people are going to be like in silos understanding the use case, approving
[29:20] the use case, all those things, right? So, if you if you're if the same ticket goes to a a person who is owning the octa groups, then he'll ask questions about what is this use case about? Like he need to understand more. So, everything is getting like automated in one single click and then like if you just fill the form and then like submit
[29:36] it you will be having the agent enrollment ready for you. So, that's where awarding 70 70 tickets is one but getting the dev test and prod automation is also another key. So, the takeaway from this entire thing is how much metadata you can collect
[29:51] the same thing that's what you can report it back to the control tower and the FinOps dashboard and other things, right? If you don't collect the metadata, then the entire organization will be running in silos. That's the that's the thing which we we all we wanted to make this as a takeaway from this. Yes.
[30:11] I think yeah, that's the end of the demo.

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.