Sponsored by: Tech Mahindra Limited | From Data to Decisions - How We Gave Our AI a Business Brain

Summary

Less than 10% of AI implementations scale at enterprise level, and the missing link is an ontology layer that gives agents semantic understanding of business entities, metrics, rules, and governance.
Tech Mahindra and Dish Network built a joint platform architecture on the Databricks Data and AI platform that grounds agentic workflows in an ontology, enabling deterministic autonomous business decisions where RAG alone failed.
The implementation delivered a 300% increase in developer productivity and a 50% reduction in costs, with a live use case in real-time network telemetry and agentic remediation at Dish Network.

Chapters

FAQs

Why does RAG fail to scale in large enterprise AI systems?

RAG relies on nearest-neighbor vector search, which destroys the relationships between entities and cannot maintain deterministic, semantic understanding of complex business rules. In large enterprises with many interconnected domains, this leads to accuracy problems, semantic drift, and unreliable agent behavior at scale.

What is an ontology layer and why does enterprise AI need one?

An ontology layer is a structured representation of business entities, metrics, rules, and governance policies that gives AI agents a shared, deterministic understanding of what data means. Without it, agents operating on complex enterprise data cannot reliably resolve ambiguous terms, apply business logic, or make autonomous decisions that comply with governance requirements.

What results did Dish Network achieve by grounding AI agents in an ontology?

According to this video, the joint implementation delivered a 300% increase in developer productivity and a 50% reduction in costs. A live use case demonstrated real-time network telemetry analysis and agentic remediation, with agents autonomously responding to network events using governed business logic.

How does the Tech Mahindra and Databricks joint platform architecture work?

The architecture layers an ontology on top of the Databricks Data and AI platform, enabling governed agentic workflows where agents resolve business terminology and apply rules deterministically. Existing Power BI logic was repurposed as the foundation for the ontology, allowing teams to build on what they already had rather than starting from scratch.

Full transcript

[00:09] Okay, good evening everyone. I think if I'm not wrong, this is the last session of the day. But um I think we have seen this flavor today morning uh in the keynote that you know Ali presented earlier today and which is the I would
[00:22] say the the topic of the day and every almost every session that I've been going to we have been hearing about this and which is uh ontologies right and it seems that that is the missing link
[00:35] between where we are from an AI capability standpoint and where we need to in order to actually create autonomous business workflows. So what we are going to do here is uh so very quickly I'll
[00:48] just introduce myself and my co-speaker. So I am Sorab Ja. I look after the data and AI business globally for tech Mahindra and my co-speaker is uh Iben Albatine who is the CTIO at Dish
[01:01] Networks uh Dish Ecoar. Right. So uh I'll quickly cover the first maybe first 15 20 minutes and then Iben will take over and and then we can go for a Q&A you know if if we do have time. All
[01:14] right, so let's quickly start. All right, so here I mean this these are just statistics I think all of us know. So I don't want to spend too much time here which is basically that AI being as powerful as it is
[01:27] still not playing the kind of role especially in critical business functions uh and showing the kind of ROI that is expected from you know agentic AI and the reason for that is what we are going to discuss here but as for
[01:39] McKenzie for example less than 10% of you know AI implementations have scaled at an enterprise level right and similarly if you look at uh the you know the other from Gart partner for example
[01:51] like 60% of agentic analytics projects they rely only on MCP and they also fail and we will look here as to the kind of technologies that we have been using for agentic in the last two to three years
[02:03] which is primarily we started with rag and then MCP server and so on and so forth is has been missing some critical element which is required to actually make them very effective at scale in
[02:16] large and complex enterprises and I'm going to you know The reason I'm emphasizing on the word complex is because that is where ontologies actually play the most important role. So let's move ahead.
[02:41] right searching more relevant and it it did that pretty well. However there are certain problems with rag which are here. So first is it destroys the relationship between entities because
[02:53] it's it's basically using using the nearest neighbor vector search mechanism and it does not maintain the entity relationships which are again required for large large scale enterprises.
[03:05] Second is you know statistic it's a statistical engine you know we need deterministic rules and that is not what rag is right. So rag can be probabilistic even though it is accurate in terms of searching but it is not
[03:18] exactly deterministic when it comes to taking decisions again critical decisions at an enterprise level where you cannot have even 1% uh you know discrepancy in your in in your uh
[03:30] accuracy of decision-m and third is where we say that semantic drift you know which occurs when when you only use a rag mechanism for your agentic context so to say right so these
[03:43] are the three reasons which because of which rag is not sufficient and I'm sure most of you would have used rag for one or two a few use cases in your enterprises but I'm you know as at least
[03:57] as far as my experience goes most enterprises have not been able to use it for complete business workflows which are agentified right what are we so what we're saying here is
[04:11] an ontology I mean so there is a way to look at the semantic layer. So there are three parts when we when we talk about a context layer. The first one is what we know as the traditional semantic layer which a lot of data professionals all of us have
[04:23] been have been creating it from time to time. Second is metrics which is nothing but all your KPI calculations and third and the most important one is entity relationships. So the point that we are
[04:35] trying to make here today is that you need all the three in a kind of a context fabric to actually ground the agents in your business context. So it's not one or the other. is going to be a
[04:48] combination of your metrics driven KPIs, your semantic layer that you know that we have been building forever with the metadata and so on and finally ontology which is entity relationship uh you know that can actually ground the agent uh
[05:02] you know in the direction that we want it to be right so here I mean again it's a very simple uh differentiation that we have tried to show here so for example when we talk
[05:14] about without ontology What what is really happening here is uh and this is an example by the way. So this is an example where you know it is let's say it has to make a certain decision in the
[05:26] enterprise. It is retrieving a master contract say you know which could be a policy document which could be a regulatory document. It validates the terms against the supplier record. So this is a workflow that we are going
[05:38] through here. But what happens is it it does its you know basic searching and retrieving and so on. But when it goes but there was an amendment to that particular policy document which it
[05:50] never touched right because the entities were not clearly correlated that this particular decision has to be taken when and you know in context of all the latest amendments which are related to a
[06:02] particular policy document and that is where you know without ontology the kind of decision-m that would happen would have missed it. Whereas with an ontology it will traverse the entire path the
[06:14] entire relationship and when I'm saying ontology it also includes knowledge knowledge graph by the way and we'll see why that is very very important. So in that case in the same example it will not just traverse the contract that is required it will also go through all the
[06:27] amendments because they are again related from an entity standpoint and then it will actually go and identify what is the spend cap or whatever is the policy that it has to follow before taking the uh before taking a decision
[06:40] for the workflow that it is that it has been implemented for. So the whole idea here is there is a risk to using uh you know agentic workflows without the right grounding. Whereas on the right hand
[06:53] side what you see is uh it's very well governed. The agent can actually be autonomous and as we will see on one of the slides in terms of benefits with ontology and knowledge graph a very
[07:05] important thing that you achieve is 100% explanability of agentic decisions and that is key in any complex or regulated industry right
[07:20] so this is the key message uh that that we are trying to give here so what we are doing here I mean we all saw the announcement today morning that now we have Genieontology which will be native to data bricks and so on which is great but at the same time and this is a
[07:33] joint IP that we have developed along with data bricks and here the scope is a little larger than than what Jinny is bringing to the table so what we are doing here is two or three things one
[07:45] whatever data you have not only within your lakehouse it'll be data in your lakehouse the relationships that exist there the documents the you know the business process documents that you have the SOPs any kind of thing and anywhere
[07:59] in your organization can all be taken as an input into this IP that we have developed and you can you will create and you'll be able to create the first draft ontology for your enterprise. Now
[08:12] a few key benefits here and which we'll come to very quickly is you don't have to go and do this for the entire enterprise at one go in this in the IP again this IP Samara that we have launched here today uh and
[08:26] you can see the demo in our booth which is booth number 558 the most important part is let's say for example you take a use case a particular use case for a particular domain and you create the ontology to implement that
[08:38] use case now tomorrow you have a different use case in a different domain within the organization and you create an ontology for that. What Smare helps you do is actually combine these ontologies. So the ontology keeps on
[08:50] growing as your number of domains and number of use cases grow. So you don't have to redo the entire activity all over again. Right? So that is one of the key differentiators and these are what we have actually tested within our labs
[09:03] in terms of the benefits of uh simara uh and creating an ontology and then running agentic process using the ontology that has been developed. Now as I mentioned earlier one of the biggest
[09:15] advantages is explanability coverage. As you can see using the ontology and the knowledge graph every single decision that the agent tr you know traverses in its journey of you know making a
[09:27] decision can be traced back properly as to why did it choose a particular entity why did it take a particular decision and so on and so forth and I'm pretty sure that is something which a lot of businesses need uh especially from a
[09:40] regulatory and explanability standpoint uh in in complex organizations even when you look at some of the other factors here. So we talked about rag earlier, right? So graph rag versus vector rag baseline. So hallucination reduction is
[09:53] up to 85%. From a business and that is you know that is increasing the pred the determinism of or deterministic value of the agendic workflow dramatically. Again uh the last one if you see on the
[10:06] top right side the policy grounded responses that was the example I was trying to you know talk about earlier where it goes looks through a document and then it misses a few documents if there is no ontology but in with ontology and knowledge graph it'll
[10:19] ensure that every document all its amendments and whatever are the latest policies are picked up before it makes a decision and again there for example just with ontology rule enforcement and audit trails there's a 90% improvement
[10:32] in in policy grounded uh responses is and which is why the like I said the key theme of this presentation is grounding your agents into understanding your business and one more thing which again I'm just repeating it again ontology is
[10:46] not the only thing that you need from a context standpoint you need the three layers which I spoke about ontology and knowledge graph is together your metrics KPIs are very very important because that is the mathematics behind your
[10:59] business and third is uh the semantic layer which again something which all of us are familiar with right now let's see see the point is at least for last few years and last couple of years every organization knows that
[11:12] we need to do this and I think one of the major blockers why people have not started with ontology creation and deployment is because there was a whole notion of very high cost very high time
[11:24] that it takes to create an ontology for the entire enterprise and what we are trying to do here is we want you Instead of creating a 18-month
[11:37] multi-million dollar plans to create an ontology, we just want to pick up certain use cases and within a few weeks create a draft ontology and implement that agentic workflow so that business can see immediate ROI that yes, ontology
[11:51] can help ground our agents. We can create deterministic autonomous business workflows and and from there we can start scaling it. So again not wait for boiling the ocean like we did you know
[12:03] when we talked about data modernization in the past everybody got on to two three four year five year plans that we have to modernize the entire ecosystem here the idea is not to do that right so because of cost and time people haven't
[12:16] been doing it we are saying let's start let's start with a particular domain and then ultimately quickly speed speed it up you look at the first two three use cases and then we scale it from there
[12:28] right And this is the point which I was trying to make that don't take it big take it small and and one property which I said which at least we are building in in Samara right now is you can merge
[12:40] ontologies so for example today you take finance domain you create an ontology you implement five use cases tomorrow you pick up marketing domain you pick up another five use cases and those ontologies can be merged so that is a
[12:52] very key point because that will help in scaling as new domains and new use cases start coming up with that I'll stop here. I'll invite uh Iben to please come and take over from here.
[13:17] So moving on from anthologies to some of the challenges that we faced uh as we moved past proof of concepts and how do we do AI and how do we want to do data and and some of those practical concerns uh I want to share some of the things
[13:30] that we went through at Echoar as we're trying to do these in our our four uh business business areas that we're looking at. So
[13:43] are we actually trying to do in 2026. The first one is that proof of concepts are over. We're done with those. We did a bunch of them in 2025. Probably half of them weren't necessary
[13:56] because we probably knew the technology was going to work in any case. However, it did point out to us that the challenge that we're facing is a lot bigger than we initially thought.
[14:08] The problem is not just about data. The problem is not just about AI. And the problem is not just about tools. It is significantly bigger than what it seemed to be in the beginning. And that's what we're starting to see right now. But the
[14:21] key purpose that we're trying to achieve in 2026 is to industrialize the use of AI. So it's not about point solutions of having a proof of concept or a small AI
[14:33] use case working inside one of the business units. It's about industrializing it with scale across the entire business and making sure that they can continue to operate autonomously or whatever method that
[14:45] they've been implemented with. And so for there we're using the term internally of adding value sustainably. So not something that we can prove to ourselves has worked for a month or two
[14:58] months or five months. This is something that needs to stay there persistently and we want to have it there over time going forward. Security is not negotiable. So as part of how we're doing data transformation, as part of
[15:10] how we're implementing tools, as part of how we're developing AI agent, security by design sits in everything that we do. And security itself is being done with AI as we progress forward.
[15:24] One of the other things that we have to do, don't know how many of you are facing that problem, but we are having to do a significant amount of work on the actual underlying data that we need to work on to make sure that the data itself is actually consistently usable
[15:38] by AI. We need to move some of our data. We need to clean almost all of it. Uh we need to figure out how we're going to govern it. And the governance is not only done centrally. We've also decentralized the governance into the
[15:51] business areas. We need to catalog it which is largely what Surup was was actually talking about it. We want to democratize that data as we move forward and then at the end we need to make sure all of it is secured. So we want to get
[16:04] all of that done this year for about 80% uh of the data that we have. We're a 45 year old company so it's quite a bit of work uh to get through it but it's 80% of the total data in the company 100% of
[16:16] our use cases. And then the last one um is people and architecture. The essence of what we want to do in 2026 is we want to be able to use the
[16:28] tools that we have implement AI agents in terms of the the components that people can access and make the data available to AI. So we can connect human creativity whether it's marketing,
[16:40] finance, HR, legal but we can connect human creativity to a nonlinear expression mechanism where people can do things which previously wasn't possible. So it's not about proof of concepts
[16:52] about it's about actually industrializing value. How can we have 12,000 professional staff in Ecostar actually do things that was not possible 6 months ago? and how can we have those
[17:04] people's creativity access those tools in a productive way going forward. So there's a large amount of actual change management that we have to deal with. So that was the what. Let's look a little
[17:17] bit at the how. So how are we doing this? As hard as the tech might be from time to time, it turns out that was the easy part. There are a number of other things that I alluded to earlier that
[17:30] are actually proving to be a lot more complicated and a lot more nuanced in terms of how we need to do this. So we want to move past for example assisted coding. So assisted coding is useful. We
[17:43] want to move to a fully automated end-to-end AI SDLC chain. That is the objective we want to end up with by the end of the year. We started in January with SDLC being done the old way. By
[17:56] Christmas, we want the whole thing to be automated all the way from feature ingress at the beginning through to testing, coding, project management, security, test data, uh, sit, test, and
[18:08] dev all the way into prod. And if there are errors in prod, we want that to be fed back into the beginning and all loops. That's the idea that we want to that we want to do. And then we want to scale that across the business. So all
[18:20] of the people are actually doing SDLC in that fashion. If we look at how we work, we want to move people past using these tools merely as a fancy way of searching for
[18:33] something that they were going to find. Anyway, we want to move past that to a point where useful conflict injection can actually have humans interact with digital agents and actually do
[18:46] productive work. Not just by themselves, but productive work in teams, which means that there's collaboration between humans, but there's also collaboration between the assistance of humans with
[18:59] humans and other agents. But that requires for us to be able to solve ontology problems. Otherwise, getting the harnesses to work well at an enterprise scale,
[19:12] I don't think it can be done. It is it is too hard to get that many people to use these harnesses in a good way if we do not give them the ability to support that infrastructure.
[19:24] Then what we use so 2025 we basically said to people do whatever you like. Use any tool, use any LLM, experiment, get excited about the technology, do proof
[19:37] of concepts, let your mind run wild. Let's get excited about how we're going to change the organization. 2026, we're being a little bit more control freaks in terms of how we're going to help
[19:50] people to do that because at the end of the day, the objective for us is not tool selection. The end of the day, what we want to do is we want to connect human creativity to the ability to
[20:03] perform work and the tool is merely an instrument. The output, the value creation is what we're actually focused on. How we own an important principle that
[20:16] we've carried into this year is that the lines of business, the P&L owners own the initiatives. They own the value creation targets. They own the initiatives. They own the accountability
[20:28] for the outcomes. And the IT teams are the people that support, that enable, we help them to find the tools and the data that they need to be able to do that. And we do rapid ideation with them to be
[20:41] able to get them on track to start to move. But the business owns the outcomes. So there's a healthy amount of push and pull happening in the organization from that perspective. But that's a critical outcome that we need to achieve is that the ownership of how
[20:55] value is created is actually owned by the business line themselves. I've spoken about support. So we track, we provide visibility, we enable, we make sure that the data is there, we
[21:08] guide them, we help them with rapid ideiation, we make sure that they have the tools, the resources, the people available to them to be able to do this. Some of this is happening automatically in some of the business units that are
[21:20] further ahead. In some other areas, we're having to embed teams in those business areas to help them out. So very unccreatively, we're calling them AI business partners, but they are technologists that are embedded inside
[21:34] the business teams that first of all help the executives of the business teams understand where they are in the transformation process. Where are you with your data transformation? Where are you with your use case definitions? Are your use cases on track? Are they doing
[21:48] what they're supposed to be doing? And then on the other hand, they're spending time with the business to understand how this unit functions. What are the key business processes that are important? What are the key data sources that are important? What are the insights that
[22:01] they need? What are the competitive metrics that they're up against either in the marketplace or elsewhere? and translating that information back to the IT teams so we can help support transforming those business processes
[22:15] for them and with them so we can accelerate that forward. And we've moved away from centralized governance. We do have components of very strong centralized governance. Enterprise
[22:29] architecture for for example sits centrally. Cyber security sits centrally. But in terms of how we govern data that is very much distributed and federated across the organization as a
[22:41] whole. How do we ideulate and how do we connect AI use cases to business cases and the corporate strategy that is also federated across the organization. So
[22:53] we've moved away in terms of value creation from a centralized control body to a body where we support and the business owns. Now, if we look at casting our eyes
[23:07] forward into the future, there are several things that we're trying to achieve uh as a as an operator, as a as a as an entity. The first one um is that we would like to
[23:22] have realtime telemetry coming from every component of all of the networks and the systems that we own. all the satellites, all the towers, all the customers, every single entity that is
[23:35] active with any one of our consumer customers. We want all of those that information to be fed through real time. We want it processed real time. And at some point in time, we would like all of
[23:47] that to, as you can see in the middle, we want to have anomaly detection to be done automatically. We want to have the reaction to that anomaly to be dealt with by a robot and we want to have it
[23:59] fixed by a robot. So about half of our technology business operates in that space. About half of the infrastructure that we have about 99% of every single
[24:11] thing that happens in those networks is found by a robot. It is troubleshot by a robot. It is fixed by a robot. The RCA is written by a robot and the tickets are logged by a robot. there is no human intervention. We would like to move that
[24:24] forward to all of our technology areas because it is predictable and it is an area that leans itself to machine learning uh which is something that we're very well acquainted with. And
[24:36] then lastly, we want to expand the learnings that we have from that towards customerf facing areas so that we can eventually get to a point where we can learn, we can interact, we can react and
[24:48] we can also act as we interact with customers. Whether it's a setto box, whether it's an interaction over a phone call, whether it's something else, we want to be able to have agents take
[25:00] action with customers. So that is an area that we're proceeding with very cautiously and that's probably something that's on the road for 2027. But it is something that we want to be able to do in the future to be able to get there.
[25:14] We see that there are some pretty major shifts that have to take place in how we think about how our company is going to be put together. First of which is it
[25:26] seems like for the most part all of us humans have to going to be stopped being the librarian of storing information, registering information and cataloging information.
[25:38] It seems like the machines are going to have to do that. They do it better. They do it more consistently. They don't forget. They're not lazy when they do it. And they do it the way that they've
[25:50] been told to do it. The funniest interaction that I've had with my own software engineering teams is that they say that the fun thing about doing SDLC with AI is the AI actually follow the rules that we made for ourselves. We
[26:04] ourselves have told ourselves this is the best way to write software. The point is half the time we don't do it that way because we don't feel like it. It doesn't feel like the way that I want to do software. So why is SDLC with AI
[26:16] doing so well? Well, because they follow the rules. Um, and when we ask them, "What is the problem with a given project or feature delivery?" The report we get back is more accurate than anything I've ever seen from any project
[26:30] manager ever. It is to the point. It quotes exact dates. It says exactly who said in what meeting, what feature needs to be delivered, where the scope keep screep came from. Something that I would never be able to get from any project
[26:43] management system that we had before. And it's not removing engineers from the teams. It's changing how those teams are put together. And it changes how effective those teams are. And it
[26:57] changes where those teams are working. Those teams are becoming smaller. They're becoming more effective and they are starting to embed themselves into the business.
[27:09] For those of you who were around with the big agile transformations of 20 years ago and safe, it's basically feels like what we used to do with agile and safe 20 years ago. We have smaller teams that are highly capable that are now
[27:21] sitting in the business areas delivering value right there and then. So the people aren't going away. We're changing the way that the teams work. And then lastly, we're seeing that the possibility of
[27:33] being able to have zero touch customer care is very very much in front of us on the horizon. And so that's a target that we're looking at to be able to achieve in the very near future.
[27:58] by about 300%. And reduce the cost by about 50. Um the challenge we've set ourselves is that by a point in time in 2027,
[28:10] early 2027, Q1 2027, we want to be able to deploy um a million lines of code per day that is tested, validated, cyber
[28:22] secure, and policy compliant in such a way that we have no more feature backlogs. Um, and we have no version one or version 0.1 features going into production. Everything has full error
[28:35] logging. Um, everything is fully featured off the go and we're discussing with business teams an actual production ready piece of software and interacting with the production ready piece of software on a daily basis because
[28:48] today's feedback will go into a build tonight and tomorrow we will have a complete rebuild of an entire system and we can discuss the production system again tomorrow. That's where we see the value is moving that we're having with our teams.
[29:01] The AI digital assistant is something that we want to be able to move to all of our professional staff. As I mentioned before, the AI digital assistant, whether it's co-work or
[29:13] clawbots, you know, whatever, whatever you prefer, they have some value as they look at my information. What is going on in my email and my documents?
[29:25] It has a little bit more value if it can look at the documents and the emails of my team and mine. It has a lot more value if it has the ability to connect
[29:38] to Slack and to Jira and to the project backlogs and to the code and to the knowledge base effectively the anthology we're talking about. then it has real value and then it can actually do
[29:51] something that I probably in the future won't have to check because today there's too much checking going on because you just don't know if the context was correct. Did it miss a document? Did it miss a fact? Was it not
[30:03] part of a meeting? Did we not record that information? So, it didn't have access to it. scale as I mentioned uh currently we have about a little bit over 80 AI use
[30:15] cases across 16 business units that are live in production um generating about 200 million a year of incremental uh savings and benefits to us as an organization.
[30:27] We want to keep on scaling that and we need all of these use cases to be persistent. So we cannot have use cases go into the business that have a shelf life that they run fine for 6 months and then they start to fizzle out because
[30:40] the data is bad or the use case is not relevant or it wasn't updated. Autonomy um we need to move agentic workloads into
[30:53] autonomous modes. The area that we're having the most success with as I've mentioned are very heavy machine learning based workloads as I expressed to with how we're working in the knock
[31:05] and SDLC. Now SDLC we're riding on the coattales of anthropic and open AI who are as a matter of business trying to do SDLC. They're trying that's all they're doing is trying to figure out how to
[31:17] write code with with robots. So we're riding on those coattails. But the other parts of the business, legal, HR, finance, marketing, sales, we need use cases to become persistent and
[31:31] autonomous. We need to have interactions with customers be able to be closed loop and monitored by agents by themselves. We cannot have them as human in the loop for too long because we will not be able
[31:44] to scale it successfully. Then if we look at genie rooms uh and what we've been speaking about here as well in terms of how genies are progressing is to be able to get the data transformed to make sure that the
[31:57] data is clean and safe so that all of our data can be accessed by all of our teams using an agentic interface. So we can start to push the limits of how
[32:09] people are interfacing with information. And then it's not about asking the same five questions that they used to get out of some old Tableau report. It's about actually having very intelligent
[32:21] interactions with the data that they have and sharing it with teams in a real-time fashion. And as I mentioned, one of the major goals that we have is that we have to actually clean all of that data underlying to the business
[32:33] from that perspective. As we do that, we are seeing areas of architecture change. We're having to question the way
[32:47] we used to do it architecture, we're having to question that constantly. the AI strategy that we wrote down for ourselves November last year for 2026,
[32:59] about three quarters of that we threw out the window by late March because so much has changed in the landscape. So, not only are we having to rethink how we think about enterprise architecture, but
[33:12] we're also having to get to a point where we actually distance ourselves with an idea that we thought was a great idea 3 months ago. And so the first one um is not so much
[33:27] the tool selection. It's not so much what LLM we're going to use and it's not so much what cloud we're going to sit in, but it's coming to the realization that the control plane and the harness
[33:41] that we apply is the area that we need to seriously focus on. These other things are peripheral questions. You like Opus, use Opus. You like JG GPT, fine. You like Grock, I don't care. You
[33:54] want open source models running on a a GPU in your office, happy days. What's more important is how are you going to industrialize the deployment of the harness to your professional staff
[34:07] so that all people have access to the same quality of professional digital assistant that's in front of them because expecting that each individual person will write the skills.md and the
[34:19] instructions.mmd files at equal quality is is completely unreasonable. So we do need to figure out how we do that at industrial scale. The second one
[34:31] is we physically cannot change our organization fast enough to keep up with the decisions we're making about data and AI. So what we're doing is we're benefiting
[34:44] from a federated model in the organization where again drawing from learnings from the past where we used to have guilds that were virtually spread across teams
[34:56] that were doing delivery. We're effectively doing the same thing here. We embed teams in the business because there's no other way we're going to learn how each other needs to interact to get the velocity that we need. And we
[35:08] need to balance centralized control with federated decision making and ownership. And those two is something that we're balancing to be able to to literally just be able to make decisions and
[35:21] implement them at the pace that we need to. The third one um is a problem that I think most of us uh have run into already um and we've seen some some good presentations about that and that's
[35:34] about how do we track the identity of of nonhuman actors inside our own networks? How do I know who this agent is? If the agent inherited my user rights, is it
[35:48] allowed to have that? Is it supposed to be doing what it's doing right now? And how do I know that an agent didn't hack the identity of another agent to try and achieve the objective? How do I even
[36:00] know that agent is who it says it is? How do I understand whether the thing that the agent is doing now is part of what it was supposed to be doing? So tracking non-human identity and behavior
[36:13] is becoming crucial because we want to very quickly get to a point where there is more work done than any human team will ever be able to monitor. But for that we need to be able to have
[36:26] trust. And so that's proving to be a critical architectural question that we need to solve. And it's not disconnected from the harness and control plane in the first problem. And then the last one
[36:38] is a new breed of observability. Previously we were worried about uptime and disk space and alarms and object monitoring and element management to try
[36:51] and understand how systems were performing. The kind of observability we need to do now is completely different. We need to understand how much work has been performed, how much value has happened.
[37:03] So, if I spend $1,000 today on tokens and a colleague of mine has spent $1,000 a day on tokens, like, am I happy? Am I sad? I don't know. What did I do? If if
[37:15] I made $1,000 worth of cat videos, but he created a security uh uh harness that protects us. Well, clearly there's more value coming from that. So, how do we measure value? How do we measure work
[37:28] inside the organization? How do we understand the velocity with which agents are moving inside the organization? How do we know whether a specific LLM is the right kind of LLM for this activity
[37:43] that's being performed? How do we know that the order decomposition or or or agent decomposition is happening in the right way? It's a completely different thinking that we need to have around observability.
[37:56] >>
[38:08] company functions, how we were thinking, and how we were interacting at teams. So we can get to a point where we can industrialize the use of AI so that overall we achieve our core objective
[38:21] which is how do we attach human creativity to the ability to express and deliver that creativity in a way that previously was never possible. That's really the goal that we've set for ourselves. Thank
[38:34] you very much.

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.