Making Enterprise Data Ready for AI With The Modern Data Company

[00:00:00] The leading issue of agentic AI in businesses right now is ensuring agents act with compliance guidelines. And Denodo applies guardrails across your entire data estate. By aligning your company's data infrastructure under one system, these guardrails perform consistently across your platform. So start scaling your business and start with Denodo. Simply visit denodo.com to learn more.

[00:00:27] What happens when an AI agent finds 14 fields called date? And each one of those fields means something completely different. Yes, a human analyst will spot that ambiguity quite early on. A machine, however, it might deliver a very polished answer that is completely wrong. So today I'm joined by the CEO of The Modern Data Company.

[00:00:56] And we're going to discuss why many enterprises are analytics ready but unprepared for AI. He will talk about everything from why trusted data, business context, governance, usage signals and machine readable meaning. How they all must travel together. And we'll also discuss model agnostic data foundations. Why centralised data projects can delay results.

[00:01:24] And how teams, though, can work backward from a business outcome rather than moving every available record into another repository. So if your AI pilots keep circling production, the blocker could be upstream of the model. And this conversation today hopefully will help give you a few pointers on how you can understand where to look. But enough from me. Let me introduce you to my guest now.

[00:01:55] So thank you for joining me on the show today. Can you tell everyone listening a little about who you are and what you do? Very nice meeting you, Neil. I'm Saurabh Gupta. I'm the CEO of The Modern Data Company. What we do is something very, very interesting and unique. Like, you know, everyone is talking about AI, AI adoption, a lot of pilots. And then you see here multiple folks saying, we are not able to scale. 80% of the pilots are not scaling.

[00:02:24] 90% are failing. So, and you know why they are failing? They are failing because most of the organizations don't have a strong data foundation. And the expectation that AI will fix the problems in data is wrong. And that's the part which we tackle. We tackle the foundation layer of data and enable all consumption use cases, AI, BI, Gen AI being some of them.

[00:02:55] And we are able to do it because we have a platform or a product called DataOS, which addresses some of these very, very fundamental needs around data to enable better AI, better control. And I'm looking forward to hearing more about that.

[00:03:13] And one of the reasons I invited you on the podcast to join me today is you set off my tech spidey sensors when I read online that you made this interesting distinction between being analytics ready and AI ready. So what does the difference look like inside an average enterprise in corporate America or indeed anywhere in the world? And how can a leader tell which side of their, or which side of that line their organization is on?

[00:03:40] See, traditionally, right, most of the folks, when they started looking into data, they were always targeted for some sort of a report, some sort of an output that could be consumed. And that means the focus was only on data and metadata. On the contrary, now when we say what AI needs is not just the data metadata, but also a lot of context around data. And that context could be controlled or has to be controlled.

[00:04:09] What I mean by that is the context could be who the users are, how often this data is getting used, what's the quality behind it? Are there any transformations done? What is the meaning of this data? So all that together is what is fed into the AI. And most of the organizations are not AI, or the data in most of the organizations is not AI ready because all these pieces are an afterthought.

[00:04:36] It's data sitting one place, metadata sitting one place, usage sitting somewhere else, transformation happening somewhere else, governance being managed somewhere else. All these are all disparate. And a lot of effort will go on to bring everything together. And that's where the struggle is.

[00:04:56] So with all this data spread out and the data about data and metadata about data spread out has to be brought in together to make the data really AI. One of the other problems, I think, if we go back in time a few years before everyone got maybe too excited over the arrival of AI, most enterprise data stacks were designed around humans that were interpreting dashboards and resolving ambiguity there.

[00:05:22] So what breaks when we now throw AI agents into the mix and they're consuming that very same data directly and making decisions without humans there to spot some of those inconsistencies? So I would say, Neil, there are two areas where there's a problem with the AI world and the Egyptic world. One is expecting that AI will sort of find you an answer. Actually, AI has to be bounded.

[00:05:52] If you let AI go unconstrained, it will hallucinate. It'll try to find an answer for you. And I think that's a very, very hard problem in a structured data environment. And that is one issue which exists and which we definitely have managed it and we have found a solution for that.

[00:06:13] So in context of our product, DataOS, we have a very, very unique definition or a very, very complete definition of something called a data product. And that data product defines the guardrails within which a Gen AI interaction or AI interaction should be constrained. So that's one part of it. The second part, which is in the agentic world, how do you manage these agents?

[00:06:41] Someone has to, like we are talking about agents will do the work for you. But is there a way to orchestrate these agents? Someone manage these agents so that they are also not going out of form. Those are the two issues which I feel are very, very big. Yeah, completely agree with you. And another reason I was excited to get you on to join me today is I was looking through your AI-ready data playbook, which essentially identifies four properties that data needs to have.

[00:07:10] Trusted, context-aware, productized, and actionable. So can you tell me a bit more about what each means in practice and where enterprises are typically weakest today from what you're seeing? So context is one which actually you want to know everything around data. Yeah. What it is. Like whether you can trust the data or not. How often this data is getting used.

[00:07:35] In the data space, Neil, one big, big, big challenge is too many copies of the same data. Too many copies of the same data with a little bit of change here. So what to trust? I'll give you an example from my past in one of the organizations where I work. There is a system of record which is producing the report. That report is taken in Excel, adjustments are made, and it is taken further for regulatory compliance.

[00:08:05] Now what has happened is what has gone to for the compliance reporting is very different than what exists in the system of records. In the next reporting cycle, if you don't make those adjustments again, wrong data is going to go into that. So how do you make that awareness, governance around data? I think it's very, very hard. And it's very critical, I would say. So that's one issue. And AI cannot solve that. Because if when you start having multiple copies of these Excel sheets sitting,

[00:08:34] which one do take AI won't know it. So that awareness around, like how do you win that trust on data? I think it's very, very, very important. And this is the place where, like if you think about the traditional way of doing things, before any report goes to anyone, there is a human who's checking the data. So how do we incorporate that into the mix? The second part, which you mentioned, is trust. I think it's so important.

[00:09:04] Like when humans are putting trust in the mix, validating data is easy. Now machines are consuming the data. How do you bring that trust into the mix along with data so that it is machine consumable? And that's where there's a need for having a proper governance. And something we call analytical workflow, governance in the analytical workflow. So typically when you think about governance, what happens is from source, as data moves,

[00:09:34] they keep applying these governance checks and balances. What I strongly believe is governance should be seen differently. You should see from the consumption point. And we call it applying governance into the analytical workflow. So if Neil is looking at data a certain way, everything that has to be limited as per Neil's profile should be applied at the consumption and it should flow all the way to the back.

[00:10:01] So example will be like if you are looking at a record and you are able to see the name, address, date of birth, because that's your level of approval you have. When I look at the same data, it should be masked for me. And can this be done from the consumption side rather than bringing it and replicating those policies over and over again? And if we do it in the consumption side, automatically those contexts,

[00:10:28] those trusts will automatically pass to the AI layer. And that's how we apply it using our platform. And there's enormous attention right now on maybe, some people will argue, maybe too much attention on choosing models, building agents and buying the latest shiny AI platform that everyone's talking about. And refreshingly, I was reading that you've argued that

[00:10:56] many failures actually begin much, much further upstream. So what would you say are the warning signs? An apparent AI problem is possibly really more of a data architecture problem. So I would say like, the point which you're bringing up is actually very, very valid. And this is just going to get more and more visible. So today I want to use Cloud. Tomorrow I want to use OpenAI. I may want to use Gemini another time.

[00:11:23] Now, the challenge what is happening is each model operate a certain way. Now, and these are going to get very much more specialized. Like financial, there might be models which will be operating only on financial side, on healthcare side. We are seeing that already. Now, the way we should think about our foundation layer of data, it should be model agnostic.

[00:11:48] So the whole data layer should offer the data, the metadata, the context package together and let the users bring the model of their choice. And if tomorrow a new model comes in, it should automatically connect. They should be able to bring in without any integration, without any changing the design of the data model behind.

[00:12:12] So that level of transparency should be kicked in when you surface your data for consumption so that they become model agnostic. And I strongly believe like end users should have the freedom and a choice to make on which models they want to use with their data. And the reason there's another shift, and I strongly believe there's a directional shift happening, is if you think about it, you have data engineers,

[00:12:40] you have data analysts, and you have consumers. With these models coming in, the role of analysts is gone. And maybe in a couple of years, we might not even find any analysts. And why those analysts layers are going to go away is because when I go to a model and say, give me what is the revenue for last quarter? And it goes in tracks and give me a well-presented report. Then I say, give me the trend over last four quarters, and it should be able to draw me a graph.

[00:13:11] Now, all these activities were highly analyst dependent. I want to get, I strongly see that we are going to get to a place where a question is asked for which a model doesn't have answer. It should say this data is missing. So that the hop is in place of to the analyst. It should straight go to the engineer saying, add this data to this data product so that I'm able to get my answer. And another massive topic this year, and a word in particular, is context.

[00:13:40] And you've mentioned it a few times already. And the reason it's so important is because an AI system can receive technically accurate data and still end up misunderstanding exactly what it means. So on the subject of context, how can a leader listening attach business meaning definitions, ownership permissions, and lineage to data in a way that machines can reliably understand and grasp that context?

[00:14:08] So on this, like, there are two parts of the context I believe in. One is the technical context, which comes in attached with the data, which is the lineage, the quality, the governance, all those. So those, for me, are straightforward. If you have a strong data foundation, all these will come package. Where the challenge comes in is the business understanding of data, like net revenue versus gross revenue.

[00:14:37] If both are tagged the same way, you have wrong answer. With one of our retail furniture manufacturers we were working with, they had 14 date fields, all tagged the same, but meaning different things. So shipping date for a furniture, for a warehouse guys, when it has gone out of the warehouse, for an e-commerce guys, when it is payment is made, and for a delivery guys, when it is received.

[00:15:06] And they have 14 these flavors. How do you bring it together? And this is the place where, again, I think a very defined ontology, semantic layer, bringing common definitions together is needed. So context, which is technical, semantics, which is understanding of business, together is something which is going to make the full context. There's another piece of context which no one talks much about it. That's about usage.

[00:15:37] Now, if I have similar or same data in five different places, I think that the one which is getting more used probably is the one either is visible easily, or it is the right data. And that context also should be curated as a part and fed into the AI. So that next time when the AI interacts and they're looking for, say, net revenue, it goes to the right data. At the very beginning of our conversation today,

[00:16:03] you did mention DataOS from the modern data company there, which I believe can turn fragmented enterprise data into AI-ready data products. But, and here's the kicker here, in weeks rather than requiring another major transformation program and at a time where the speed of technological change is going so quickly, that'll be music to the ears of many people listening. But tell me more about it and how this approach works and how it coexists with the warehouses, lake houses,

[00:16:33] APIs and other investments companies might already have. Tell me about how this works. So, Neil, one of the challenges in the data industry is centralization. I want to bring, I bring Databricks. Everyone, all the data should come through Databricks to make it easy to use. I'm using Snowflake. But the reality is the volumes of data are going to continuously increase. They are increasing. They have increased in the past. They are continuing.

[00:17:00] The type of data, the veracity, the variety formats are going to continue to change. Now, for us to use this data, do we need to centralize it? Do we need to bring it into a lake house warehouse before we can even start using this, I think, is flawed approach. And this is more defined by the technology underlying or the tools which are built around it.

[00:17:26] What we have done is, we have, DataOS, we call it a platform for data products. What it does is, it creates these AI-ready data products irrespective of where the source is. So, we bring multiple disparate data sources together. And we do it in a way that they are tied to an end outcome. So, a typical use case which takes several weeks to do it, we are able to do it in a couple of days.

[00:17:55] Because we focus only on the right parts of the data that is needed. We really don't believe in bringing all the data together into a centralized lake house warehouse. We'll bring it if it is needed for latency reasons, etc. Plus, on top of that, when you bring all this data together, the quality credentials, the type of data issues, all that gets packaged together.

[00:18:22] So, if you are a consumer, from day one when it is made available to you, you are able to look into the details of data and you can start asking for more. And every time with our approach, what we call it, right to left. So, think of the outcome that a user is expecting and then start building backwards to the data rather than bringing all the data into a lake house warehouse. So, that's how we do it.

[00:18:46] And what we have seen is, typically, when the first use case is done in a couple of weeks, the subsequent use cases takes days. Because we start looking at building all the adjacencies. So, you don't need to bring all the data over and over again. And once you get to a certain level of maturity, we have co-pilots which will automatically start building the data products.

[00:19:12] And before you join me on the podcast today, I was reading how you've reported that enterprises using data OS that we're talking about here have accelerated AI adoption by 90%, while at the same time reducing total data ownership costs by as much as 50%. So, there's the ROI right there. But I'm curious, digging a little bit deeper on that, what changes actually produce those results?

[00:19:39] And what should leaders begin measuring to determine whether improving their actual data foundation could actually translate into business value? So, very interesting question, and I'm glad you brought it up. So, two statistics which we always really, really build ourselves on. One is acceleration, which is bring, solve use cases 90% faster or save time by 90%.

[00:20:06] And the reason why we are able to do this is, our approach is that right to left. We don't believe in bringing all the data into a centralized lake house warehouse, then do a quality runs on a transformation and make it accessible to that. What we say is, Neil, give us the strongest problem, the hardest problem you are struggling with. We will target that problem, solve it in no more than four weeks.

[00:20:34] Now, this is, so what typically what happens is like the several months of investment that people put in in building the centralized data repository, the whole data platform around it. We shrink it because of data is we are able to shrink it to a couple of weeks. That's number one. Second is, and I was telling you like it's hard to track what data is getting used.

[00:20:56] Now, as we, what DataOS offers is every interaction, whether it is a machine interacting with data, human interacting with data or a BI tool interacting with data, we are able to track it. Now, what it does is, over a period of time, we are able to do two things. Stale data, we are able to identify and we can make a case that these don't need to be processed anymore. So that's one step.

[00:21:22] Second is, like take for example, someone is looking into a report, an executive report. It was processed once a month. You know, most of the product companies, they'll go and say, why do you want to, why are you processing once a month? You can do it every day. Now that every day is a very good need. People love to have it every day, but it increases your compute by at least 30 percent, 30 times, 30 times.

[00:21:53] But what we do is, we say, we are able to correlate usage versus production. So if someone comes and tells me, we'll never go and say that this can be done daily. We'll say, Neil, you are reading, looking at this report only once a quarter. Why do you want to run it on a monthly basis? Let's shrink it. And that's what, over a period of time, can bring a significant shrinkage in the costs, in the compute costs.

[00:22:22] And all we are doing is making it transparent, making it visible, and making it just easy to correlate. And for any CIOs or CDOs listening who have dozens of AI pilots, but somehow still struggling to get them reliably into production, anything you'd advise them to examine inside their data estate first?

[00:22:46] And what practical steps could they start taking now to become genuinely AI ready? So, first, Neil, I think AI is inevitable. It has got into everyone's life, and we have to just figure out how we can take advantage of this. Enterprise landscapes, especially on the data side, are very complex. And they have organically grown, so we are carrying a lot of dead weight.

[00:23:14] I was on the enterprise side for a couple of decades. And the amount of dead weight we are carrying on data is very, very high. So, I strongly recommend that in place of seeing AI enablement, data foundation enablement as a technology problem, we should start thinking of a business outcome.

[00:23:39] Identify the most problematic areas or most areas in demand first. Put it together, use cases around that, and start solving those problems. So, incremental business focus implementation is something going to be important. The second is, we have all or all enterprises have accumulated a lot of tools in their data landscape.

[00:24:09] Some doing security, some doing quality, some doing governance, some doing orchestration. And most of these tools are not fully utilized. And the challenge with these tools, what we are also seeing is, like you need to have different skills. They have different upgrade cycles. So, there's a trend which is picking up as convergence of these capabilities.

[00:24:36] So, start looking at a possibility of converging or compacting your data infrastructure so that you have less number of tools to manage, less number of skills to worry about. And in place of spending time in integrating these tools, shift it towards dealing with creating new use cases which can deliver better business outcomes. And DataOS is one of those platforms or products which does that.

[00:25:04] So, we have a convergence like we offer end-to-end capabilities. DataOS offers end-to-end capabilities for a data lifecycle management. So, you don't have to bring five different tools and an SI to support to glue tools together. And as we mentioned at the beginning of the podcast, as the modern day company released, I think it was the AI Ready Data Playbook from Pilots to Production.

[00:25:30] So, for anybody listening wanting to find that, maybe connect with you or your team online or dig deeper on anything that we've covered today. Where should people listening go? So, the best way for our folks to reach out to us, I think, we have done a lot of work. We have a lot of case studies. And anyone who is interested in really know more about us, happy to sort of reach out on our website.

[00:25:59] We have a demo. We have a playbook. We also have built a community, Neil, of data practitioners. Over 20,000 folks now called Modern Data 101. So, we get a lot of inspiration from them. We learn from them. So, love to sort of introduce anyone who's interested with all the resources we have and anything that can help enterprises.

[00:26:28] Lastly, we are also going to start a webinar series and would love for people interested to participate or not just attend, but also come and join us on our webinar series so that we can sort of bring their practical challenges and use cases to everyone. We also just wanted to also highlight one more resource which will be very useful. So, because of this community, we also run once a year survey.

[00:26:55] So, we learn from the industry what's happening and we publish it as a report. So, the last report, we call it the data activation gap. That is also available and which brings out a lot of details about how enterprises are, what type of challenges they are facing and how they are trying to solve the problem. So, a good resource to have in your access. Awesome.

[00:27:23] I cannot thank you enough for talking about a problem hiding in plain sight in a language that everyone can understand. And that is that the data stacks powering Fortune 500 analytics were never built for machines. They were built for humans. And there's a lot needs to change to make sure AI can understand that too and get that context. So, I will include links to everything you mentioned there. And I'd love to throw this out to everyone listening.

[00:27:51] Let me know your experiences and what you're going through. And maybe we can keep this conversation going. But big thank you to you for taking your time to sit down with me today and share everything you're seeing. Really appreciate you. Absolutely, Neil. And I really appreciate that. I think my guest's central message today is incredibly easy to overlook because the industry loves talking about models. But AI cannot interpret business meaning that the data estate has never captured.

[00:28:21] And before adding another agent, maybe leaders should ask which data product it will be. Who owns the definitions? How are you going to measure its quality? What permissions apply? And whether usage proves that the sources can be trusted. As we mentioned in the very beginning of the podcast there, those 14 date fields in that furniture example. Yes, they are amusing.

[00:28:46] Until the machine selects the wrong one and acts upon it at speed. So, a big thank you for a very timely reminder why AI-driven data requires context, trust, governance, and actionability by design. And I will be adding links to the company's playbook, community, case studies, demo, and data Activision Gap report. And, of course, the Modern Data Company website. But over to you. Is your data ready for machines?

[00:29:16] Or are your agents just relying on humans to resolve ambiguity that they can no longer see? Let me know. TechTalksNetwork.com We've got lots of stuff coming your way. We're having 60 interviews a month at the moment across eight of our podcasts. We've got a lot of sister shows. Check it out and let me know your thoughts. But that's it for today. So, thank you for listening as always. Bye for now. Bye for now. Bye.

[00:29:46] Bye for now. lenses. Movie insulin.ソン