Connecting AI Agents to Live Enterprise Data
Episode Summary
Enterprises adopted AI fastest where the data was unstructured: HR, legal, support. Deepti Srivastava, founder of Snow Leopard and the founding product lead for Google Spanner, argues the harder and more valuable problem sits on the other side, connecting a probabilistic model to the structured business data a decision actually depends on. She explains why the blocker is not intelligence but plumbing, why she thinks dumping everything into one place is the wrong answer to data silos, and what a data control plane between agents and databases does instead.
Key takeaways
The unglamorous work is what stops agents reaching production. Building pipelines, dumping data, transforming it: these consume 80 percent of budget and time, and they are why so many agents never leave proof of concept
Structured and unstructured data used to serve different purposes, which was fine. AI removed that separation, and most of the visible wins so far have come from the unstructured side
Srivastava went to VPs of engineering, CTOs and CIOs and heard the same thing: they knew how to use AI for HR and support, but not for revolutionary workloads, because leadership decisions need real answers from live data
Her answer is a data control plane sitting between the agents and the data, agnostic to both layers, so a developer never has to reason about where the data comes from or whether it is current
Data silos are permanent, and consolidating everything into one place is the wrong response. There is always a source some team holds that you will never know about
Build agents intentionally rather than porting old workflows across. Lifting a legacy process into AI is the mistake digital transformation and SaaS already made
About Deepti Srivastava
Deepti Srivastava is the founder of Snow Leopard, which builds a data control plane connecting AI agents to live structured enterprise data. She has spent two decades in enterprise data: founding product lead for Google Spanner, and before that a distributed systems kernel engineer at Oracle working on RAC locking and high availability infrastructure. Her argument is that data are the crown jewels of any enterprise, and that the gap between probabilistic models and dependable structured data is where most enterprise AI stalls.
In this episode
| 00:42 | Welcome and guest introduction |
| 01:29 | Two decades in enterprise data, from Oracle RAC to Spanner |
| 04:26 | From deterministic databases to good enough |
| 05:10 | Where structured and deterministic have to meet |
| 06:12 | Why everything you knew has to be thrown away |
| 08:55 | The conversation that led to Snow Leopard |
| 09:11 | What CTOs said they could not do with AI |
| 10:25 | What a data control plane actually is |
| 11:16 | Who it is for |
| 12:30 | Staying agnostic to the agent layer and the data layer |
| 14:00 | The boring problems that eat 80 percent of the budget |
| 14:41 | Why middleware needs to simplify, not complicate |
| 15:56 | Silos were built for reliability, and the chance to rethink that |
| 16:39 | Why consolidating everything is the wrong answer |
| 18:02 | Whose work she is watching |
| 20:32 | Leadership skill: adaptability |
| 22:12 | Wrap-up |
In Deepti’s words
“Data are the crown jewels of any enterprise.”
— Deepti Srivastava (01:29)
“In this new world, anything you knew you have to throw away.”
— Deepti Srivastava (06:12)
“The boring problems take 80 percent of your budget, 80 percent of your time, and are a huge reason agents don’t make it from POC to production.”
— Deepti Srivastava (14:00)
“We do actually need a middleware that is simplification oriented, not complexity oriented.”
— Deepti Srivastava (14:41)
“Anybody that’s making predictions on what the world is gonna be like, in my opinion, is wrong, because as the words leave my mouth, things have already changed.”
— Deepti Srivastava (14:41)
Resources
Deepti Srivastava and Snow Leopard
• Snow Leopard: snowleopard.ai. The data control plane between AI agents and structured enterprise data
• Snow Leopard docs: snowleopard.ai. She points listeners to the docs page for live examples of building an agent against it
• Snow Leopard blog: blog.snowleopard.ai. Her writing on agents and enterprise data
• Deepti Srivastava on LinkedIn: linkedin.com/in/thedeepti
Systems and frameworks discussed
• Google Spanner: cloud.google.com/spanner. She was its founding product lead
• Oracle RAC: oracle.com. Where she started, building locking and high availability infrastructure
• MCP and FastMCP: Two of the agent integration paths Snow Leopard supports
• Pydantic: pydantic.dev. Named among the supported agent frameworks
• Vercel: vercel.com. Named among the supported agent frameworks
• Agentuity and AG-UI: A live working example she cites, with AG-UI as the front end for agents
• Shopify and Salesforce: shopify.com and salesforce.com. Named as examples of the structured sources agents need to reach
Related AI Realized episodes and events
• Artifact-Scoped Agents: Stop Mimicking Job Titles: Chris Butler of GitHub on scoping agents to the artifacts they produce, and where the human belongs in the loop.
• AI Governance as Code: From PDF Policies to Pipelines: Ken Johnston and Bob Rapp on making governance executable inside the deployment pipeline.
• Agentic AI and Revenue Work: What Actually Pays Off: Christopher Penn of Trust Insights on agentic AI, measurement, and what has actually produced revenue.
Frequently Asked Questions
-
Deepti Srivastava puts it down to the unglamorous engineering rather than the model. Building pipelines, dumping data, transforming it: she calls these the boring problems, and says they consume 80 percent of budget and 80 percent of time. Her point is that the friction is not intelligence, it is complex plumbing, and that is precisely what stalls agents before production.
-
Srivastava describes Snow Leopard as sitting between AI agents and structured data: SQL databases, APIs, systems like Shopify or Salesforce. It becomes responsible for fetching what the agent needs, so the developer never has to reason about where the data should come from, whether it is accurate, or whether it is current. That frees them to concentrate on building the agent correctly.
-
Because the expectations are different. Srivastava spent two decades in a world where databases were deterministic, predictable, and expected to be exactly right, then watched AI arrive steeped in unstructured data where good enough was good enough. The visible wins came fast in HR, legal, and support. Decision-making workloads need a probabilistic model to produce a real answer from live structured data, and that marriage requires both AI knowledge and deep systems knowledge.
-
Srivastava went to VPs of engineering, CTOs, and CIOs after ChatGPT and Anthropic broke through and asked how they were adopting AI. The answers were consistent: HR systems, some support workloads, and then a caveat that they did not know how to use it for genuinely revolutionary work. The reason was that leadership needed decisions made on real data, and what was available were dumps and transformed warehouse extracts.
-
Srivastava disagrees, respectfully but clearly. Data silos have always existed and always will, and there is invariably a source some team holds that nobody else knows about. Her alternative is a control plane that fetches from wherever data lives, agnostic to protocols, APIs, and languages, so an agent can reach what it needs without anyone first relocating it.
-
It means not repeating what digital transformation and SaaS did. Srivastava warns against taking an old-world workflow and translating it into AI, and argues that in a market moving this fast the only durable anchors are ROI and doing the right work to get the right outcomes for the business and its customers. She is also blunt that anyone predicting where this ends up is wrong, because conditions change faster than the prediction can be spoken.
-
Adaptability, and an open mind. Srivastava reaches for the distinction between wartime and peacetime generals, then sets it aside: her point is that experience shapes how you see the world and stops you repeating mistakes, but it also carries baggage that has to be put down. She acknowledges this is a confusing and fearful time for many people, and describes herself as optimistic on the grounds that any platform shift can go either way and which way is up to the people leading it.
-
[00:42] Christina Ellwood: Welcome to AI Realized, the podcast for enterprise executives leading AI deployments. From tackling security, data, and operational challenges to navigating organizational transformation, AI deployment offers a unique opportunity to redesign organizations from the inside out. I'm Christina Elwood, your host for today's episode, and we're talking today with Deepti Srivastava, the CEO of Snow Leopard AI. Thank you so much for joining us on the show today.
[01:13] Deepti Srivastava: My pleasure. Thank you, Christina.
[01:15] Christina Ellwood: We, you-- we're really glad to have you. I know you went from building Google Spanner, one of the most foundational databases in the world, to founding Snow Leopard AI. What thread connects these two chapters of your career?
[01:29] Deepti Srivastava: Yeah. First of all, I'm really excited to be here. I think this community is really actually at the forefront of, of adopting AI and realizing it, so it's really great to talk to people who are not just talking about it, but like, doing it. Yeah, that's a really good question. I have been in data for about two decades, so that is the thread for me and always has been. I have always believed that data are the crown jewels of any enterprise. I've also been in enterprise data the whole time, right? So I have seen all these systems as building, with building Spanner, and before that I was a distributed systems kernel engineer. That's where I started my career at Oracle, in the Oracle RAC database, building their locking and high availability infrastructure. So I've really seen stuff from the guts up, if that makes sense. If you think about it, like 20 years ago or 15 years ago especially, even now, but 15 years ago, Oracle RAC ran all enterprise databases, right? Like, all of the enterprise data was on, on those systems. And so I saw it from the bottom up, and then when I moved to Google, I was the founding product lead for Spanner, which literally changed the game in how we think about data and data in the era of scale, which was what the internet in the last 10, 12, 10 years of like digital transformation have been. And as the data proliferates, what you figure out is, oh, governance is really important, and access is really important, and like the liveness, the freshness of the data is really important. The correctness of data is really important. So cut to '23-ish, '24, when ChatGPT first came out, and AI was no longer just like a side note or a bubble. Like I, for me at least, at, even at that point, I was like, "Oh, this is real this time." And as a side note, like, you know- Transformers were built at Google, and I happened to be there to see the original inception of them, because some of that metadata was also on, on my product. And again, I saw that from the bottom up. But when you come out of it and look at the world top down, you're like, "Oh, there's gonna be a new world." We now call the, that world the agentic world, but 18 months ago it was just AI applications and AI, and what is that gonna be? And in that world, I was like, "Oh, it's fun to talk about the weather with a chatbot," or, "It's fun to look up what's interesting out there," right? It's fun to do chatbot stuff, but really, in order to change the way people work, the way people interact with the world, AI has to be adopted into enterprises. And how do you do that? You have to connect the technology. I think of, by the way, I think of AI as a technology, right? It, it's part of the tech stack. I don't think of it as some thing that's gonna take over and I'm less worried about Terminator. I'm more worried about, as I say, where my order is for the thing I ordered. And so AI has to be adopted in the enterprise and has to connect to critical workflows, and that requires connecting to the crown jewels, which is structured operational data.
[04:26] Christina Ellwood: Yeah. I couldn't agree with you more. And Deepti, you remind me of, oh, one of the other evolutions that's happened in this process. When you started in the database world, everything was deterministic. It was all transactional. Right. And for a sort of novel use of technology or data at the time was messaging. Messaging between systems and being able to... So we went from this very deterministic-
[04:49] Deepti Srivastava: Predictable ...
[04:49] Christina Ellwood: reliable, accurate, critical, all of the things that are associated with it has to be perfect world, to a world where it, good enough was good enough, right? This sort of vague world of what is the status of the data? At what state is the data? How old is it? Is it, has it been written yet to the database? Things like that.
[05:10] Deepti Srivastava: Yeah.
[05:10] Christina Ellwood: And that changed fundamentally how we use data at the very moment when unstructured data really started to explode. So when AI came onto the scene using, being steeped really i- in unstructured data, the databases that we were using in the enterprise for our transactional systems, which are 100% structured- And
[05:33] Deepti Srivastava: deterministic, yeah
[05:33] Christina Ellwood: and deterministic need to meet in this world. And I know that's a place where you've spent a lot of time in this sort of AI meets structured business data, while still embracing the unstructured, because a lot of our data now is in unstructured form in the enterprise. It's not all structured. Yeah. Some of that data has its own quality parameters to it. And just like in the days, old days of, of your early career, you're still connecting data to workflows to people.
[06:01] Deepti Srivastava: That's right.
[06:02] Christina Ellwood: So that triangle is still something that needs to be dealt with. So what did you see at Google that an, that most people in AI still haven't fully reckoned with?
[06:12] Deepti Srivastava: I think it's not just about Google anymore. I will be honest with you, in this new world, there is anything you knew you have to throw away. I think that's my insight number one, right? But going back to a couple of things you said, number one, not only did we move to an unstructured sort of not as deterministic world with the sort of data lakes and those kinds of things, right? With the proliferation data, like unstructured data had to be there in order to think about forms and about a bunch of HR text and things like that. So we definitely moved to a world of unstructured information, and that was okay because structured and unstructured were used for different purposes, and th- that was fine, right? With the AI revolution, what's really interesting is that AI can touch anything or should touch, like the way to use it is the technology itself is not limited to just unstructured. Although obviously the most gains, the most aha moments were seen with unstructured information as with chatbots and HR and legal, where it has been adopted really immensely, right? And, and really causing change in the way people behave and work and act, and bottom lines and top lines are changing there, right? But again- It has not, and I continue to say this, like it has not been an- adopted in the central core of enterprises because that core has decision-making in it, and decision-making is done deterministically. So the question really is, how do you take this LLM based, you know, non-deterministic probabilistic world and connect it to a very structured, very deterministic set of things and workflows that require deterministic answers, right? That's the gap that Snow Leopard is supposed to bridge. That's the gap that I want to make sure is bridged properly. Because to give you a very quick example, you could do HR systems, but if you're trying to hire somebody, you need to know what the... You do a background check, and Christina with a K, and Christina with a CH, and Christina with a C sort of can all, in a rag type world, in a probabilistic world, right, map to the same thing. So you cannot have an answer of maybe Christina has a criminal record. That is not acceptable, and this is one of the reasons why, you know, this type of information and this like live fresh information, you, we can't rely on yesterday's dump of criminal records, right? We have to have today, right now, what does Christina, what is her status, right? Background check. And for that, like you need access to the correct systems at the correct time, right, with the correct data, right? This is the gap. When,
[08:55] Christina Ellwood: when did, when did you have the realization that you just have to build this solution that is now Snow Leopard AI that solves this very problem that you're describing, that it connects the, the AI to the business data?
[09:11] Deepti Srivastava: Yeah. Being an enterprise product person, the first thing that happened when I saw ChatGPT break free in Anthropic and stuff is I went and talked to my enterprise friends, right? That run like VPs of engineering or CTOs or CIOs, and I was like, "Hey, how are you adopting AI?" And they all talked about, oh, HR systems and maybe support for certain types of workloads, et cetera. But every time there was like this, "But we don't know how to use it for like truly like revolutionary workloads yet." And I was like, "Why not?" And then they were like, "How..." Going back to how do you connect this probabilistic LLM to I need real answers right now, like my leadership needs decision-making capability on- perfect data to your point earlier, and that's just not available. It's dumps. It's transformed stuff into warehouses. I cannot make decisions based on that. And that was when I was like, "Oh, this requires not just LLMs and AI knowledge, and not just systems knowledge and how to build these predictable, deterministic, strong systems. It requires a marriage of both."
[10:14] Christina Ellwood: So are you building a layer on top of the data that is doing the stru- the, the, the solving the gap? Is that what, what- Yeah ... you think about Snow Leopard AI?
[10:25] Deepti Srivastava: That's a great way to put it. Yeah, so the idea here is that Snow Leopard sits in between your AI agents and your data, like specifically your structured data. So think SQL databases, APIs, Shopify or Salesforce or things like that, and, you know, we are responsible, we become, Snow Leopard becomes the data control plane where it'll fetch the data for you based on what the agent is looking for, right? So the idea becomes that you do not have to think about, "Where should the data come from? How do I make sure that it's accurate? How do I make sure that it's timely?" Like that stuff is left to Snow Leopard, and you build your agent and think about how to build the agent correctly, right? Because we have to be intentional about how we build these agents in this new world.
[11:10] Christina Ellwood: I see. Okay. And so who's a, what's a sort of perfect use case for your technology?
[11:16] Deepti Srivastava: Anybody who's building an AI agent that requires SQL databases or structured information that is, you know, critical for decision-making, so live data from the right source to make the right decision.
[11:29] Christina Ellwood: Sounds like- That's the story. Great ... um, things like transactions. It sounds like things like purchases or-
[11:38] Deepti Srivastava: Yeah,
[11:38] Christina Ellwood: that's
[11:38] Deepti Srivastava: a good- If, if- Yeah. Yeah. If you're building a support, like a truly live support agent or just-in-time resource planning or things from there all the way to I need to make purchasing, hiring, those kinds of decisions that I have to be based on what's happening in my CRM system right now, what's happening in my order management system right now, what's happening in my customer system right now, and most importantly, what's happening in my revenue system right now. So fintech and AI support tech are, are the sort of top adopters of this technology right now because they seem to have the most need for, as you said, transactional or live real-time data, right? But I think it applies to everyone.
[12:18] Christina Ellwood: Yeah, it sounds like data-wise, it's completely agnostic in terms of the structured database it sits on top of. Yeah, I know what you
[12:23] Deepti Srivastava: mean. Yeah.
[12:24] Christina Ellwood: Yeah. But from an agent point of view, are there certain stacks that you support and certain-
[12:30] Deepti Srivastava: Yeah, so listen, I, I feel we have to be the place where developers don't have to worry about what is supported or what is not. We are in the middle, right? So we will allow you... We're agnostic to that too. So we're agnostic to the agent layer, and we're agnostic to the data layer. In fact, we have partnerships and live examples even on our website if you go today. Like, on the docs page, we have examples of how you build an agent with Snow Leopard, with ThinkChain, with MCP, Fast MCP, with Pydantic, with Vercel. We have a live working example with Agentuity and this thing called AGUI, which is a UI front end for agents. So we want to reduce the friction for developers right now, AI developers right now, from building agents. Because they have, in the previous world, in the SaaS world, right, like it was the same thing where you have a bunch of tools and so you have to choose which tool is right for you. Then you have a bunch of like things around data pipelines, data lake houses, data like oceans that you have to create and maintain and continuously spend 80% of your engineering time trying to build and maintain these things. What if you just erased all of that and said, "Build your agent, talk to Snow Leopard." Snow Leopard takes the headache of routing the data intelligently in real time, like making sure that you get the right data that is accurate out of the box. Like accuracy out of the box is a huge challenge today for agents, agent builders, even with MCP. MCP is just a connector layer.
[13:57] Christina Ellwood: Yeah, it doesn't do anything with the accuracy.
[14:00] Deepti Srivastava: That's right. And so complex engineering, those kinds of things are actually the real friction point. The boring problems, if I were to say it that way, because these are boring problems. Build the pipeline, dump data, transform it. And these, by the way, boring problems take 80% of your budget, 80% of your time, and are a huge problem of agents don't make it from POC to production. I wonder why.
[14:22] Christina Ellwood: I wonder why. This is, of course, the story of middleware through every era of technology. So you're like the new middleware, which is fantastic. If we're looking forward to the next couple years, what do you think are gonna be the, the innovations in the use of agents, and therefore, what does the middleware layer need to be able to do?
[14:41] Deepti Srivastava: Yeah, I ha- in a way, I hate the term middleware because everybody has these like negative connotations around it. No. But I think, but I think that is exactly right. Like, we do actually need a middleware that is simplification oriented, not complexity oriented. I think that's really important. And in the... I think, again, in the age of AI, um, and the way that things are moving so quickly- First of all, anybody that's making predictions on what the world is gonna be like, in my opinion, is wrong, 'cause as the words leave my mouth, like, things have already changed, right? Mm-hmm. And so in that kind of lightning speed world, the only thing you know is you need sort of two things. You need to make sure that you have ROI, and you have the right set of work done in order to get the right outcomes for your business and for your customers. That is what needs to be focused on, which means you have to build your agents intentionally. I think this is the time where you have to think through not just let's take the old world workflow and transform it into AI, as we did with digital transformation and SaaS. That was the old world. In the new world, I think it is a great opportunity for enterprises to rethink What workflows do we really need to even have anymore?
[15:56] Christina Ellwood: Yeah, there's certainly that for sure. But I also think it's an opportunity to rethink how we're operating because we built silos for reliability and accountability, and now we have the opportunity to build systems that are cross those silos, cross-functional, and are able to pull our silos closer together and create a, uh, smoother, instead of a flattened out organization, a, a more cross- Always both vertical and horizontally oriented organization. And I think that's an exciting thing to think about how we would instantiate that technically and how we will look organizationally when that's-
[16:39] Deepti Srivastava: Thank you for bringing that up actually. That's one of the key things that I care about with Snow leopard, right? This is why middleware, like the way you describe it, is actually really important in this new world to be thought of in the new way because data silos have been a part of life and a fact of life, and the way people are trying to say you should solve it is dumping everything into one place, and I don't think that's the right way respectfully, right? I think those data silos can exist where they are 'cause they will always exist. There's always a data source of some variety that some team has that you're never going to even know about. So what if the middle agent, right, the middleware allowed you to, um, especially the data control plane as I call it, allowed you to fetch data from anywhere and for you to connect to the plane, control plane from anywhere, right? You're agnostic to the protocols, the like, all of that stuff, right? APIs, languages, doesn't matter. And your agent can always, through the control plane, access the data that it needs to make the right forward progress. That is a world I want to be in, right? Where your data silos are no longer a topic of discussion, but what are you doing with the data is the topic of discussion. How are you using it to make forward progress for the business, for the individual, for the whatever it is top... That's the key to unlocking the value that we all think AI can provide. I'm excited
[17:59] Christina Ellwood: about that.
[17:59] Deepti Srivastava: Yeah. I,
[18:00] Christina Ellwood: I agree, and the ROI as well, right?
[18:02] Deepti Srivastava: Yeah.
[18:02] Christina Ellwood: So who is someone in the AI space who y- whose work you're quietly watching closely right now?
[18:11] Deepti Srivastava: There are actually multiple people. One of them is this, I think he's deputy CTO at Microsoft. He has a team that is basically building Tools around agentic coding so that you could just fully, like engineers no longer have to code, right? But it's building skills around Claude and those kinds of things. His name is Sam Chalise. Did I... Hopefully didn't butcher his names. And he has these things called Sunday Letters. He writes a blog all about where AI is today from a, from an agentic coding and software engineering perspective, and I think that's a really great one to follow. But there's a bunch of people around me that are doing this, like that, that are working on making agentic coding just fully automated, right? Like where you don't even, like where you care about outcomes versus like code and code reviews and code writing. Like they're fully adopted AI, they're building fully AI agent companies, like not... And just seeing what that looks like and what that means, and they are at the ve- they are at the very cutting edge of what AI can do. But it's very- Can you name them? I don't know if I can. There's a team at Strong DM that sort of released a blog post recently about what they have been doing. There's a bunch of people who are in stealth mode that I just haven't- Yeah, that's why
[19:26] Christina Ellwood: I, yeah, that problem too, that sometimes these companies are in stealth mode and you can't actually say who they are.
[19:32] Deepti Srivastava: Yeah, but it's fun to watch them, and I will call out one thing, which is, you know, all these people trying to like figure out what AI and the nature of work in the new world looks like, which I think is cool because they're just open-ended doing research and it's really cool to like, uh, like practical research and it's cool to see them. And I adopt some of their tried and tested methods in our team as well. So there's that on the one end. On the other end, like there's enterprises and security and governance and that's not where enterprises are, and I wanna fully acknowledge that and I wanna say I don't think you can jump from where we are today for the majority of the world to where these people are. And I think anybody that tries to tell you should do that, I think just doesn't understand the reality. So the real question for me, and what I think about a lot, is how do I take what's happening with AI today and help enterprises adopt AI more effectively throughout their organization? Everybody can't do greenfield, and so how do you do this AI transformation in a realistic way?
[20:32] Christina Ellwood: That brings another question to mind for me, Deepti. As a leader of a AI organization, but also a leader in the community, what do you find is your defining leadership skill in this particular era?
[20:52] Deepti Srivastava: Yeah. The, you know, there's a thing saying that wartime generals are different from the peacetime generals or something like that. I certainly... the metaphor may not be the most apt right now, but, but I think the key skill I feel is both necessary, and I happen to be that way personality-wise, is just adaptability. I just-- you have to have an open mind, and you have to get rid of all the strings and all the baggage and all the knowledge. Like, the knowledge is really helpful in how you tran- like, how you shape the world in your head and how you therefore shape the world through your actions, right? We don't wanna keep making the same mistakes. But at the same time, like adaptability and just open-mindedness is just, is key. Like, I think there's a lot of people who are afraid. There are a lot of people who are confused, uh, because it is a confusing and therefore fearful time. But at the same time, I think I am optimistic because any technology to me, right, any change in technology or any platform shift, um, has the potential to be very positive or otherwise, and it is on us as people, as leaders, as societies, as communities to choose how you use that technology in service, in my opinion, in service of us, in service of humanity.
[22:12] Christina Ellwood: Yeah. At AI Realized is fostering AI adoption, responsible AI adoption, so we obviously buy that, buy into that same, um, premise and are optimists ourselves. Thank you so much, Deepti, for joining us today. I really appreciate it. We have been talking with Deepti Srivastava, who is the CEO and founder of Snow Leopard AI, and thank you for joining me today on AI Realized. Thank you so much for having me. I love this community, and I really honestly
[22:38] Deepti Srivastava: enjoyed the conversation so much. Thank you so much, Christina.