The Data Stack Show - Re-Air: Context is King: Building Intelligent AI Analytics Platforms with Paul Blankley of Zenlytic

Episode Date: November 19, 2025

This episode is a re-air of one of our most popular conversations from this year, featuring insights worth revisiting. Thank you for being part of the Data Stack community. Stay up to date with the latest episodes at datastackshow.com. This week on The Data Stack Show, John chats with Paul Blankley, Founder and CTO of Zenlytic, live from Denver! Paul and John discuss the rapid evolution of AI in business intelligence, highlighting how AI is transforming data analysis and decision-making. Paul also explores the potential of AI as an "employee" that can handle complex analytical tasks, from unstructured data processing to proactive monitoring. Key insights include the increasing capabilities of AI in symbolic tasks like coding, the importance of providing business context to AI models, and the future of BI tools that can flexibly interact with both structured and unstructured data. Paul emphasizes that the next generation of AI tools will move beyond traditional dashboards, offering more intelligent, context-aware insights that can help businesses make more informed decisions. It’s an exciting conversation you won’t want to miss.Highlights from this week’s conversation include:Welcoming Paul Back and Industry Changes (1:03)AI Model Progress and Superhuman Domains (2:01)AI as an Employee: Context and Capabilities (4:04)Model Selection and User Experience (7:37)AI as a McKinsey Consultant: Decision-Making (10:18)Structured vs. Unstructured Data Platforms (12:55)MCP Servers and the Future of BI Interfaces (16:00)Value of UI and Multimodal BI Experiences (18:38)Pitfalls of DIY Data Pipelines and Governance (22:14)Text-to-SQL, Semantic Layers, and Trust (28:10)Democratizing Semantic Models and Personalization (33:22)Inefficiency in Analytics and Analyst Workflows (35:07)Reasoning and Intelligence in Monitoring (37:20)Roadmap: Proactive AI by 2026 (39:53)Limitations of BI Incumbents, Future Outlooks and Parting Thoughts (41:15)The Data Stack Show is a weekly podcast powered by RudderStack, customer data infrastructure that enables you to deliver real-time customer event data everywhere it’s needed to power smarter decisions and better customer experiences. Each week, we’ll talk to data engineers, analysts, and data scientists about their experience around building and maintaining data infrastructure, delivering data and data products, and driving better outcomes across their businesses with data.RudderStack helps businesses make the most out of their customer data while ensuring data privacy and security. To learn more about RudderStack visit rudderstack.com. Hosted by Simplecast, an AdsWizz company. See pcm.adswizz.com for information about our collection and use of personal data for advertising.

Hey everyone, before we dive in, we wanted to take a moment to thank you for listening and being part of our community. Today, we're revisiting one of our most popular episodes in the archives, a conversation full of insights worth hearing again. We hope you enjoy it and remember you can stay up to date with the latest content and subscribe to the show at datastackshow.com. Hi, I'm Eric Dots. And I'm John Wessel. Welcome to The Datastack Show. The Datastack Show is a podcast where we talk about the technical, business, and human challenges involved in data work. Join our casual conversations with innovators and data professionals

to learn about new data technologies and how data teams are run at top companies. Before we dig into today's episode, we want to give a huge thanks to our presenting sponsor, Rutter Sack. They give us the equipment and time to do this show week in, week out, and provide you the valuable content. RudderSack provides customer data infrastructure and is used by the world's most innovative companies

to collect, transform, and deliver their event data wherever it's needed all in real time. You can learn more at rudderstack.com. Oh, welcome back to the Data Stack Show. Yeah, exactly. We're here live from Denver. Yeah, from my house, actually. So I got to catch up with Ben Rogojon

last week in person here at Denver and now we get to do this. Yeah, the Denver 18th. I mean, it's not a good week, you know? Yeah, exactly. Awesome. Yeah, catch us up from when we last talked. Yeah, there's a lot of exciting things going on on the Zenlit side.

We've been getting a bunch of great new logos like Jay Cruz, Stanley Black and Decker, some of these just fantastic companies to work with. Awesome, yeah. And we've just been seeing AI and just go gangbusters in terms of overall capabilities of the models. And it changes a lot of how generally this stuff, needs to work in the future. And it just, it changes so fast. So it's like nothing I've ever seen in terms of rate of change of the industry overall.

Yeah. So I'm always curious to these two questions. One, is it going faster than you would have thought? And then like the follow up to that is like, what is something, let's pick a six month or year time frame that you're like, wow, this is like, I did not expect this. So I think, I think definitely faster than I expected. And I see a lot of really. bad takes for people like, oh, well, the model's only getting sort of incrementally better.

It's like, you just use these things often enough to realize, like, the range they're improving. So it's like definitely fast-man anticipated. And I think it's also the domains in which they are getting dramatically better is what's maybe most interesting to me. So it's like things that they continue to be sort of approximately human or like sub-human or a lot of sort of mind vendors are just sort of understanding, communicating, these kind of things that humans do a lot. And then if you look at things that they are already superhuman at and increasingly getting more dramatically superhuman at,

it's like coding. Like, you can throw an AI agent and it will win or come in, you know, within the top five, the best programmers in the entire world who have been training on doing these, you know, past. Same with mathematics. Same with any symbolic,

sort of task. And the reason for that is that you can generate a massive amount of training data on these symbolic tasks that you can verify or correct. So it's like you could say, hey, this test case needs to pass if this code is written correctly, and then you can just reinforce and learn that process at a just truly massive scale, way more so than you can sort of soft problems.

Right. So while the models have certainly improved and how they handle things in sort of softer domains, how they've improved in terms of code generation or math or physics or other hard or symbolic domains, that has been maybe one of the most interesting things to me. It's also shaped a lot about how I think about the future

because I think within six to 12 months, we're going to be superhuman in every respect in symbolic domains. So it's like there will be basically no human coder in 12 months that is a better programmer than a language model. That doesn't mean that the human won't be writing better code than the language model, just like someone who has more,

business context, but it's not as good a coding is going to do a way better job in terms of the actual impact on stuff matters than a language model. But a lot of our job is thinking about how we build an AI employee is how do we give the AI the right context to work with to be able to use the stuff that is superhuman at to really help the human make their decisions. So interesting. So another question along the AI models, what's your take on the long running models? Because I've seen some stuff come out where people like these stats are like, oh, these long running models, they don't actually perform that well, you know.

And then other people who are like, it's amazing. Like, you know, the longer it can run, like, you know, the more it could do. Yeah. No, so I think this is something that divides a lot of people, even they used it pretty often. So I love them. I think it's amazing. Like you just have to do a good job sort of

pointing it at the research problem and the method approach you want to take things like that. So for instance, if I wanted to go and do an analysis, of like, hey, what are the different data sources that people most often intervene with Salesforce. It's like I would have an idea of how I want to approach that problem and, you know, what other stuff fits with Salesforce in this way? And I can go and tell the agent, hey, this is generally the approach I want you to take. And it can go and look at a thousand different sites

or a thousand different references and sort of aggregate all that up. And then as I look at the report, I can click and see the citations and make deeper in anything I want to take deeper on. So I find it to be incredibly powerful. I know a lot of people, though, like, just don't engage with that in the same way. Yeah. And I don't really use it for, like, quick and dirty things. Right. So I think it's some of its difference in usage patterns as well.

Yeah. Yeah. I've seen some deep research, seen really neat things. Manus. I've seen some really neat things. So, yeah, I, the thing that I've wanted, and this actually brings us kind of into the B.I. territory, the thing that I've really wanted to see, which I'm interested in your take is the long-running analyst task where you're, where you know, there's a lot

lot of like there's a spectrum here there's like the text is equal essentially or the equivalent versus like I have this really deep problem with data from like a bunch of different places and I want you to like solve the problem I think with your kind of AI employee concept you guys are really trying to span both so I'm interested I want digging a little bit on the harder problem and like is like long I assume like the long running task is part of that but like what are some other components to solving for those like deep McKinsey or consultant sell Yeah. I think a lot of it is you first got to give the model an interface in which it can really cook, like which it can really work in and work well. That interface is still got to be

governed. Most importantly, it's got to be, it's got to correspond to your actual business context because if you aren't understanding the environment you're working in, just like an employee who's very technically capable but doesn't actually connect to the business, it's not going to be that valuable at the end of the day because it's just speaking a different language from what people are going to need to make actual decisions. I refer to that is the data science problem. Yeah. Yeah, it really is.

Because that's what companies 10, 15 years ago, they all hired data scientists. And then there was, honestly, I think a lot of it was a communication gap. Like, he's remarkably smart, PhD people with all the, like, maybe stats background and computer science background and a lot of them struggling with the business value. Yeah. Yeah.

And I think with AI, if you don't do a good job or praying the business context, you get kind of the same thing where you get these answers that are maybe generically good or good on apps. but not actually good for helping you make to say side of your business specifically. So that's a lot of the stuff that I think is really important. The other thing is

picking the right kind of model for the job. Like right now we have a model picker, which is not something I'm very excited about. I would love to see us have something that's more of a delineation between kind of fast and slow. Okay. Where instead, but right now you have model picker.

So you pick GPT 4.1, it's going to be really fast. It's going to answer the question quickly. But if you wanted to go to a sort of comprehensive look of what's going on, it's not going to be so great at that. Do you find at this point in time, are customers, like,

you've got to have probably a subset of customers that are really into like, oh, like this is my model. It's like a branding thing almost. Like, I like this kind of car, right? Even though maybe they're like fairly equivalent for the task. Do you find that to be the case?

You find those people are like, just tell me whatever works best. I don't care. So one of the things that was interesting, as soon as we launched the model picker, People love to just go in and play around because I didn't find it. Yeah.

And for different companies, because there is this tradeoff, right, between latency and sort of comprehensiveness. Right. So, DPD 4 is going to be really fast. Right. Claude Sonic 4 are like 03 are going to be incredible in terms of the debt and stuff that they can do.

But they take forever to run. So you've kind of got to consider what experience do I want my users to have by default. Do I, by default, want it to be more thoughtful and slow Or by default, do I want it to be faster, but maybe less comprehensive. Right. So one of the ways I like using Zoe to go and explore some data is that I'll just go and I'll say like, hey, maybe a comprehensive dashboard of everything that's going on in the

and just like let it crank. And it'll just crank and put all this stuff and put it together. And I've got an overall view of everything that's going on and can kind of dive into any areas that I don't want to focus on. And I imagine too as we progress, there's going to be a mix here where if I'm sitting down, Like, all right, I'm going to spawn five background tasks to go explore these things. Those run. And then, like, while those run, I'm going to be running 4.1 or something and just, like,

kind of exploring around. So you can kind of do both. Oh, totally. I mean, not that you couldn't already do that, but I think I forget that's the thing. I forget that, you know, that you could actually kind of multi-threat this stuff. I just forget. Yeah, no, it's easy to do because it's as easy as, like, you open a bunch of different tabs and you just start different friends in the different news.

And it's crazy, the amount of stuff you can parallelize just to that right now. Right. Yeah, I mean, you know, just like a bunch of tabs, there's the whole focus thing that like, you know, like you have to be careful with because you do want, you know, you do still have circumstances where like deep thought is the right answer. And it can definitely be a distraction. Yeah, absolutely. But also kind of plays into where I think where we're going in the long run.

So I think about us in the long run as, as. as an employee, basically, as another person. So it's like the, you know, quote is like, don't buy software, like hire talent. You just be thinking about AI agents is almost something that you're hiring your business. And I think about our place in that as effectively an AI McKinsey consultant.

That's especially good at work from the data, knows your business context, and can immediately start launching and helping you answer some of these questions that actually matter. So a lot of the problem that I view that solving is being, B.I. as it's existed for a long time is primarily, like, a collection of facts. And that's fine

if you have someone who's extremely analytically minded minded and wants to, like, hop in and look at all the facts and kind of build those into a narrative and go with that. But a lot of the problem that analysts have had to solve, and a lot of this isn't really rocket science. You know, if you're referred to as like data literacy, it's like, how do you go from this collection of facts and turn that into a decision about which products you're promoting campaigns? Bring that into decisions about when you're retailing inventory for these different skews. So that is one of the real fundamental problems

that AI enables that we're especially excited. That's all. So I've got to ask about the AI employee thing because this has come up more than once. Do you think that in the future companies will pay for training for their AI employees? I think they have, and they already do. Okay.

And in what way? I'm imagining like we have an AI employee and we like send them to a conference to a big person. I know it's not going to be quite that way, but yeah. I think it's actually going to be called data engineering. It's going to be because a lot of the job that you've got to think about is how do you broker business context to a model.

And a lot of that is like doing actual data modeling, data engineering the way we about it now. A lot of it's doing similar work to data engineering but applied to text. Because like how do you get the right text in the right spot where this model is able to act on in a way that matters? And now we have incredibly capable models that, given the right context, can perform a lot of road tasks, like, really well and pretty reliably now. Right. So it's like the problem and the thing that we have to be very cognizant of is it's like you need basically context engineers.

Sure. Make sure of the right stuff shows up at the right time. Okay. So context engineers. Is that a job posting? Like, are people hiring? Not a job posting, I'm not yet, but in a bit of years? In a year?

In a year, it could be. Yeah. Yeah. Give me like prompt engineering. Yeah. Man, that's fun. Okay. So speaking of context, structured data, kind of where we've been on structured data.

We had some runs with some, you know, database is devoted to them, you know, back 10 years ago. Do you think, well, two parts of this question, do you think the existing structured data platforms are going to actually be able to kind of retool to support when everybody needs out of an unstructured data stores or anything that's? That's a whole new like greenfield thing where people are going to have maybe a structured data platform and unstructured and hook them both up maybe to it, to an AI or BI tool. Yeah, great, great question. I think it's, it's a hard question to answer because the water is going to be really bloody. Like you have kind of an enterprise search, which is previously done, the sort of unstructured things.

You have all the BI players, like legacy VI players doing like everything just on top of SQL. I think it's inevitable that in the long run these two things come together. One of the things language models did is they changed what people think about when they think about data. So data used to mean like a table, a CSB, something in a database that you have to ask somebody about. Data now also corresponds to contracts and, you know, internal knowledge docs and all this other stuff that is immensely valuable, but previously wasn't thought about as data. And it's time for business intelligence to actually like make good on the business intelligence. knowledge of the sequel work. Contracts is such a good one

because I can't tell you the number of companies I've worked with, all of the companies I've ever worked for of all sorts of valuable knowledge in DocuSine, essentially, and PDFs where they, you know, they customize the contract for that enterprise deal

and they keep doing it and they keep doing it. And nobody has any idea what the deals are. Somebody has to dig up the PDF and the person left and you have to go get into their evening. You know, it's like a whole thing. And then there's an initiative like, hey guys, get all this in Salesforce.

So it would be so interesting once like all of that. is accessible in the same way, probably even a better way than like a structured data would be. Yeah. Well, I think the other thing that you alluded to is that part of the benefit and problem with BI, the way it's done now, is that it relies on very structured data that sits inside of a SQL warehouse that requires a person, set up a connector, transform a bunch of data, clean it to the point that it actually corresponds to what the business is referring to. And that's great. That's all of a lot of very important metrics and,

KPI's and how are you actually performing at a business at this high sort of aggregate level. But it is just so much work that no one would ever bring all PDS of all their contract. Right. And like you would need a five-trand to connect to AccuSine. That probably doesn't like this. And if you don't have to read, even if you don't have to write that yourself, you still got to like somehow manage all these PDS and that sort of engineering work. Some companies will go down that road and do it. Other ones, it's they need a more plug-and-play away to get unstructured data out of these systems and into effectively search indexes.

So the agents can start work. Yeah. Interesting. So kind of another topic I would think about a lot, or MCP servers, because there's, so now as we're talking unstructured data, structured data, and then we've got like MCP servers here. We're like, okay, that's an interesting interface. Like, how do you think the MCP servers play it into the BI space? I think NCP servers are actually really interesting.

And maybe hot take is that an NCP server, if you are building one as the BI product, you are eventually going to get in place. And the hot take there is that BI systems aren't actually systems of record. They don't have the end of their just processing layers. True. And that processing layer can be shown as email code. That processing layer can be shown as SQL, depending on which VI products you're using and how it's structured internally.

But it's like if you're basically outsourcing your entire UI to Claude or to open AI and MCP server, eventually you're going to be in trouble. Yeah. Because you're just a processing layer. You don't have the actual data like Salesforce does. Like that's what you guys. I think it's only a defensible position if you have proprietary data for some reason. Like I know of a couple of companies would have a unique proprietary source of data

and them putting MCP on top of it. It's the same model they have now. they're charting for data via an API. Like, cool, put an MCP. Like, the model. But if it's not like proprietary or unique in any way, I totally agree. Yeah. And I think it's like for how people work with enterprise data to succeed, it's like,

you do have to own the interface. And like, if you can't add enough value in your interface and how people understand what you've done and the attractability piece, how it integrates with dashboards, how it integrates with other parts of the system, it's like you'll kind of get heat. And it's like a lot, you, the thing that people want, and this kind of goes back to your MCP question, is a lot of people will come and ask us, like, hey, can you be an MCP server? I want to build one ager. And the question to go back to your unstructured comment, it's like, well, why do you want to do that? The reason people almost universally want to do that is because they want the one system, they can talk to their PDFs over here, and that can talk to their structure data up here. And if BI, like, BI must evolve to do that. Otherwise, it's not going to actually. be business intelligence, it's going to be SQL intelligence. Right, right. Yeah.

Well, and the other thing, too, is MCPs are still just a layer over APIs, and we'll have these same problems that APIs have until a different abstraction is formed. Because I think people actually have too high of an expectation on MCPs, when it's like, look, there's a reason we put all this effort in centralizing these things in databases is because it's better for analytic care clinics. It's not that we didn't know how to interface with APIs. We knew how to do that. We did this on purpose.

Yeah, exactly. It's like, and that's what I mean about the value out of the UI. It's a lot of people think you just kind of magically post something into an MCP server. Everything's like great. Right. But it's like there is value in UI. Yeah. Like the final UI is probably not one universal chat system that literally everything.

It's like, cursor is very successful. Cursor is not the same thing as chat GPT. Like it has UI. Yeah, exactly. That is really beneficial in coding. So. Yeah. And one of the, you know, from this,

and letting product and fairly unique, actually, like the ability to switch back and forth, I think it's going to be more commonplace, at least I hope, where you can start with a gooey element of a chart and say, hey, I want to essentially talk to this chart or dashboard and know more about this thing. Or I can start from chat and then like, hey, I want, you know,

just say what you're looking to learn more about and create charts and graphs. I think the ability to go back and forth is a big deal from user experience. Totally. And I think that's something that cursor got really well, something Andre Carpath was talking about in his recent demo day talk where it's like a lot of AI applications that succeed are going to find this right balance when giving the model

flexibility to do what it wants by where you can kind of control the autonomy of the agent. And I think that's a pretty good way to think about it where it's like if you in Zenlinic, right, if you want to be in full control, like agent is not really involved. You can be in the UI clicking on things and same way you did before. You can also ask Zoe to go and pull a piece of data for you. It's like barely agenting at all. And then it's like you could also ask it to do deep research on what's the tariff impact

on your gross margin going to be of all these different skew. It's like take this CSV of tariff on raw materials and blow it through to my final margin impacts. And that's a really big ask and a big task that you need to do. So it's like there's this slider of autonomy or our own product right now.

And I'd expect that for products that do well, they master the kind of transitions between that. So it's really easy and feels really natural to go from I'm kind of running everything to the agents running some stuff to the agents running a lot of stuff. And you can sort of go back and forth as needed to solve your problem. Yeah, that's awesome. So I want to go back a little bit to the unstructured data conversation. What's hot there right now? Because originally it was like, okay, everything in vector databases and obviously still some of that.

But say I've got hundreds of thousands of PDFs, like what are you seeing people do? with it, want to get that epitata into a business intelligence, you know, AI business intelligence layer. So I think there's a few different components. There's like the deterministic filtering that's going to happen in whatever query language on whatever database you're using. You probably do need some component of embedding or vectors. There's a lot of complicated stuff that you can be there.

Junking the documents, which embeddings do you use based on the domain where the text is going to show up most often to make sure your embeddings are most fitted to the problem you're trying to solve. You're also going to need some hybrid component to the search as well. It's probably not going to work that well if you just use the embedding. You don't have, you don't have also a keyword search as well. So it's like there's, it's going to be a pretty sophisticated system. Yeah, okay.

Yeah, that's what I have sense. I think a lot of people that desire, and I've seen a lot of people start on that road and kind of, never mind. It's just because of the complexity, not because they don't want it. Yeah, yeah. It's hard. It's hard in a different way from data. I think data is kind of like deceivingly easy. One of our best, one of our best sort of pipeline motions for customers is customers

that have tried to build up themselves. Sure. Because it's like, their own AIBI agent. Yeah, exactly. Because it's like, it's trivial to start. You can literally use an MCP server. Yeah, and it's like, boom, it can go and it can answer any question in any table in your whole CIP warehouse.

But it will inevitably pick the wrong table a lot. It will not have the business context on how you define gross margin. And you won't have the governance for anyone who's not an analyst, like SQL level permissions that you want to use. It won't have interoperability with dashboards or other things that are actually useful assets for data. So it's like it's really easy to start. But as you kind of go down the road of not only do I not have this integration, do I not have this governance? Most importantly, the business people don't actually know what happened. Because you can't show someone a giant 200-line SQL statement and say, hey, this is what we did.

You know, check our work. I don't know what's given on there. Yeah, it's so funny. Back to the like, like the, essentially the search problem. That's been one of the things that I have wondered for years why nobody has just self the basic search problem for BI. I can't think of a single BI tool that has great search. Oh, yeah.

Because, like it's, but the reason it matters is I see two things. One, you would obviously not be surprised business users all the time ask for things that already exist all the time. For sure. But worse, is analysts a lot of the time rebuild things that already exist too? So it really goes both ways. No, it totally does. And I think a lot of the problem is due to hierarchies that get introduced and don't get searched over.

Yes. Because it just makes the engineering a little bit easier. So like a good example is like, Looker has explores. If you just need to see revenue and you see sales, the marketing, you know, pipeline, revenue, explore, like, what do you click on? Yeah. I don't know. And it's like that, that usually stops a person. And they're okay, well, I don't know.

And it's in the Boulder called leadership. Yeah, exactly. You know, who knows? So it's like all these hierarchies that don't really make sense are what caused a lot of problems with data discoverability. Yeah, it's super interesting. We're going to take a quick break from the episode to talk about our sponsor, Rutterstack. Now, I could say a bunch of nice things as if I found a fancy new tool.

But John has been implementing rudder stack for over half a decade. John, you work with customer event data every day and you know how hard it can be to make sure that data is clean and then to stream it everywhere it needs to go. Yeah, Eric, as you know, customer data can get messy. And if you've ever seen a tag manager, you know how messy it can get. So rudder stack has really been one of my team's secret weapons. We can collect and standardize data from anywhere, web, mobile, even server side, and then send it to our downstream tools.

Now, rumor has it that you have implemented the longest running production instance of rudderstack at six years and going. Yes, I can confirm that. And one of the reasons we picked Rudderstack was that it does not store the data and we can live stream data to our downstream tools. One of the things about the implementation that has been so common over all the years and with so many rudder stack customers is that it wasn't a wholesale replacement of your stack. it fit right into your existing tool set.

Yeah, and even with technical tools, Eric, things like Kafka or PubSub, but you don't have to have all that complicated customer data infrastructure. Well, if you need to stream clean customer data to your entire stack, including your data infrastructure tools, head over to rudderstack.com to learn more.

Let's talk products. So Zinn-Lenik, we're talking about the product maybe over six months ago now. Yeah, maybe give us a little bit of a roundup but what you guys have been working on, so things you guys have rolled out. Yeah. So a lot of cool stuff on our side.

So workflows are maybe the biggest feature that we launch. Workflows are a way to take intelligent analytics, like intelligent sort of processes. So some of our customers will use this for a weekly business review. They're a little good group, pull a bunch of data, be able to take that to impact. It's like, hey, you know, this you went down the most. We walked that hierarchy and went into the, you know,

class that went down the most. department that would doubt in the mode, the individual skew levels and you're able to give this very comprehensive analysis that can go step by step and incorporate some intelligence the same way, a human would. Other use cases where people are taking this feature and using it for almost like dispute resolution, where you can take, you know, I've got a screenshot or a CSV from SAP and I need to take all the items in there and marry that with something in my warehouse and figure out, you know, based on this, what sort of message should I be sending to each of the suppliers that come through in that CSV I got from SAP.

So a lot of different, really interesting use cases when you can kind of take some of the processes that you do for analytics and turn that into something repeatable that you can just schedule it to run every week on Monday. You can just incrementally remove this work from your play. That's one of the big ones. The other stuff is a lot around making our interface more flexible for the language models. So I think to get a little history, of how these interfaces work. There's kind of text-a-sexual interface. Like Snowblake Cortex is maybe the biggest one,

but a lot of YC companies are doing this as well. And it's really flexible. It can do anything, right? It's just whatever you can do in a warehouse it can do. The problem, though, is that governance level, where it's like, how do you govern this to someone who doesn't have a Snowblake account? And then most importantly, the interpretability.

How do you make clear what you did in the same way a talented analyst would and in a really high fidelity way? to a business user. So they can trust the result that they're actually getting. That trust piece is really important and that's what's missing in Texas SQL.

Then there's Texas semantic layer. We kind of led to charge in this and for a long time we were saying like, hey, this is how you've got to do data. And I think we've realized that also doesn't. Interesting. It's too restrictive. And it ends up being

where you end up just pulling mostly facts that already exist on your dashboards and just instead of finding a dashboard and clicking on something, you're just pulling that. not flexible enough to give you the power you need to actually answer

questions. So the pro, of course, is Trump. Yeah, it's gone right. But it doesn't have the... Which if you had to pick, then if you had to pick, then you would pick the types of semantic layer. But the unstructured data part, like how do you put a semantic layer unstructured data, for example? So I think this is actually the exact distinction that I would want to get

in. This is what we're building on the structuring side is it's like the way I think about the problem is that you have to interpret what you did to the user. So to take one step back before we hop into that, the unstructured question. Why does it work with unstructured data? Why does deep research work? Because humans just naturally understand the, okay, you said you got this thing from here. There's a link where I got it from. You can click on the link.

It's literally a bit of thing. We can do it. We can do it that for a really hard time. So you can go and check the sources and any human can do that. If you see something that looks questionable, you can say, I don't know, check the source. Maybe that's not quite a fair summary of that. And you can dig in really easily. That's one of the reasons that deep research is so successful. The research just came in and said, hey, I came up with this, and you had no way to dig in or question any of the things that came up with. It just wouldn't work as well. Yeah, definitely. And that's exactly how I think about the problem on the data side. You need to give the model the most flexibility. You need to free to sequel, like let it write

sequel. Like language models, like we were just saying, it had it become perfect at coding. I think that's inevitable. And you need to let them write SQL. You need to let them write SQL. SQL, let them do with their good app. But then the hard job of the application is how do you take that SQL that the model generated and have a truly trustable way for the business person to be able to look at that and know what it did with absolute certainty, the same way they can have certainty if they click on a citation link in a text document. So does the semantic layer become a QA layer? Is that part of quality control? I think you have to think about the semantic layer almost in this inverse way where it's like the whole time up until now,

We've been taking the semantic layer and saying, hey, we've got all these building blocks that humans understand. And we then have to, the semantic player's job is to compile that down into a bunch of SQL that we brought on the warehouse. I think the real job of it in this new age is the inverse. You have a bunch of SQL that the language model wrote. Given the context that's in your semantic layer, like the semantic player still owns the context.

It's structured kind of similarly in terms of its input. But the language model has written the SQL, raw sepule based on the input here. The semantic player's job now, is how do you take the SQL language model as written and effectively invert the problem? How can you take that and take it back into business concepts that a normal human could understand

without having to write the reason. Yeah, those kind of translation problem. Yeah, exactly. Which Ellums are good at. Ellums are good at. And there's a lot of code that humans have to write to do this. It's because, again, you can't forget permission, you can't forget governance, you can't forget all the complexity

and actually being able to verify that you did stuff the way you think you did. So that's a lot of what's what we're working on and what we're replacing. I mean, speaking of SQL, then we're going to go down the flexibility route. I think they're both actually really important. The reason for SQL is that SQL is the lingua franca of data because it's what runs natively in the natives. Yeah, right.

Like if everyone had the money to just spit up massive spark clusters to sit on top of the air, you know, as everyone's moving iceberg. Yeah, right. And great. Yeah, Python would be fantastic. But it's like, as it is now, you've got to bring the, you've got to bring the compute to the data. Yep.

And that's what. It's just most effective. It's just most effective. And then you do the aggregation, you take this truly massive amount of data, aggregate down to something reasonable, and then you have Python work on it instead out of that. And that's exactly how we're architect. All right. Well, I got to ask you that roadmap stuff.

Like, what is something like kind of midterm, you don't have to give us details that you're excited about? maybe directionally. One of the biggest things that I'll give you two. One's like an improvement on our current experience. And that goes really in line with all the like flexibility stuff we've been talking about. Part of the part of our problem and every semantic layer kind of BI people's problem

is the setup process. So like if you buy Lusker, if you buy us, if you buy holistics, you buy, you know, whoever. It's, you're going to have to do a lot of work setting up this part again of how we think about the model should be able to generate SQL, and we should be able to translate that into intelligible concepts for you. That also means the model needs to be able

to build your semantic layer for you. Sure. Because as it's translating that into intelligible concepts, if you're an admin, you should be able to say, hey, that's now governed. That's verified. Whenever you show that to the user, give them the thumbs up.

Like, that is good. We know that's good. And so now we can actually help you build the semantic model as you go, as opposed to you having to build it up front to make where he is. So that's a huge difference in terms of... That's so good.

Like democratization, you know, it's overused at this point. But that's like a reality when you can do stuff like that. Yeah, because otherwise you... And this is what we saw, and I'm sure every looker and everybody else saw this too, is it's like, otherwise the semantic model becomes something

that the data person is building, trying to guess at what the user actually wants to do. And instead, it should be where the user is able to ask the questions they want, get answers, and they know which ones are sort of rubber-approved by the data team and which ones aren't. And they can either ask the data team about or take with a grain of salt. I was going to say that.

Do you imagine there are like gold standard semantic model and maybe some personalization, personalized semantic versions too? Yeah. So I think the way we think about it is actually what the teacher we're calling dynamic fields where it's like you have the governed, you know, semantic measures. Those, the model uses all the time whenever you ask you out those concepts. But if you say, hey, what does gross margin look like if we take out discounts and we take out out refunds that we add in this other like adjusting factor and it can just do that for you

yeah and it'll say like hey here's your adjusted gross margin that you asked about jumps gross margin yeah exactly it's like hey i did it for you it's not the verified one and you can see the verified one next to it it's not the same thing as that right so it's like you can't really confuse the two but it's like you do need that flexibility and that's one thing that I think looker and similar sort of bi products are going to miss yeah with this approach is that people don't want to come in and just ask about the same things on their dashboards they want to ask about new things. They want to be able to do stuff they were not able to do before. Previously, they could only do a conversation with the analyst. Right. Well, and I think that

then at least every FBI analyst role that I've been in the past, the, like, not well-kept secret is that most of what you do goes to waste. Yeah, totally. It's like people either don't look at it. They look at it once or they forget that they ask, you know? And, you know, and, you And you're doing it a week later, and they're like, oh, I forgot about that. So, like, just the inefficiency there is mind-boggling. And I think anybody that's been an analyst for longer than a couple weeks now is. Oh, yeah. And then to paint that picture, again, a little bit on, like, where we're going.

And also on that piece, too, with the inefficiencies. It's like, I view us as a, again, like, McKinsey consultant and coworker. And it's, like, the way that works now is that we have the business context via a semantic model, this analyst is able to answer questions for you, interpret what those things mean, be really flexible, you know, help you build assets that matter, like these intelligent workflows, these dashboards. And then as we get more advanced, and as the AI systems get more advanced, it's like that increasingly becomes something that you're able to get, the deeper questions in your business to be more, the harder and harder things. Like Sam Alvin said

in the stuff like, keynote, it's like he expects them next year, people not just be getting their sort of like mundane questions they ask this to give it your harder questions give it the really difficult stuff right let me see that and then the other thing is it's you want it to be proactive you don't want to have to go and ask this and everything you want it to be able to be monitoring your inventory and it's like let me know when I should be buying stuff I don't want to ask you when I should be buying stuff I want you to tell me what I should be buying stuff so yeah that's a lot of what I'm curious on the proactive stuff because I've had that come up a number of times how far out do you think

We are really meaningful, useful, like, hey, because that, I mean, that's all another domain, too, that we haven't even touched on is, like, monitoring and alert. Like, you know, I've got a little bit of the dead office background, like a huge component to that. And some interesting, like, solves around that, like, companies like Datadog. How do you think that comes to the BI space? Because the BI space has been bad, honestly, not very good at the monitoring and, you know, anomaly detection and stuff. I think it's also just that it's not, it's because the anomaly that they are explicitly trying to, is not the anomaly they actually care about.

So it's like you don't actually care and that X, Y, Z metric is like two scanning deviation. No, yeah. Very rarely do you actually care about that. Like, what you care about is some process that's affected by it. And the ideal interface there is that you can say,

like, hey, I'm worried, like, if any of our product pages go down, it would result in a drop in, you know, conversion rate for that process. Like, apply general sort of human heuristics. And it's like, if it drops to zero because there's no visits, then it's obviously no problem.

And it's like there's a lot of different heuristics that humans have to be able to say, well, is this really a problem or not? And if you just alert the person every time the thing drops, you can't do anything from its last value, you're going to lose all of that. Yeah.

And there's so many false closets and people ignore it and there's use laws. Totally. And it's like a lot of how I think about as the agents become more ambient, like the next step in that direction is that you say, make sure nothing bad is going on with conversion rate.

And it can actually, check, hey, well, we've got all the conversion rate. These ones, the, you know, number of sessions also dropped to zero. So from like fours, that's not really a problem. And it's like, it's able to just alert you and things are actually a problem. Right. As opposed to every time some state. Right.

Yeah. And you're starting from the end and saying, like talking about conversion rate or revenue or whatever, and then talking like, well, if it impacts that, that's what I care about. You're not several layers down of like, well, monitor each page and check the traffic and So, you know, and you're going to, like you said, miss something or something like is going to get alerted to like, well, no, I didn't mean that. Because it would be so precise if you're going to monitor like kind of the old way. Totally.

And that's why a lot of the monitoring codes just as kind of brittle like sequel or ad on that gets written that doesn't find some edge case. And it's like really what you want is reasoning. You want intelligence apply the problem. It's not even a lot of intelligence. It's just like a little bit of intelligence. Previously, that was impossible. Right.

So it's like, I think that's sort of. the next step in terms of how things get kind of more ambient and more just right for you and then at some point you onboard the system and the system is just crawling around in the background and you decide how much compute do I want to allocate yeah it's just sort of figuring out for me so yeah and you can just slide hey I want you just go and find all the stuff right I'm willing to I'm willing to spend on you're like just check a little bit like yeah right don't check that's the big stuff yeah I mean how far away do you think we're from something like that I mean that's like where

That's on our roadmap for the end of 2026. Okay. So I think it's closer than you think. I think it's because the more flexibility you give the models and the better job you do a brokering the business context, they can just do this increasingly magical stuff. Yeah. Through no, like, it's not like we or anyone else or cursor builds like the,

you know, you don't have to build the model. Right. It's like your job is to orchestrate all the stuff around it. And the assumption that, you know, us and Cursor and a lot of people are making who are taking these bigger bets around AI is that they're just going to keep getting better. Right. Which means like as long as you give them the right flexibility and the right guardrails, then you're set up for them to just get dramatically better

and be able to take on these bigger tabs just naturally, as long as you get them the right tools. And the right way to think about the tools is what tools do the humans have that are really good. Sure. Yeah. So it's like, yeah, the humans do have a B.I system, which is the people have a semantic layer that you can hit around in and pick the things with the semantic layer. The humans also have a sequel editor. Right.

Yeah, yeah. And both of those, due to the human's understanding and communication skills, can become interpretable to the receiver. Yeah. And so it's like, okay, we're going to take those things and give it to the AI as well and not limit it to like one or the other. Yeah, exactly. And I think the thing that you're going to see miss from all the kind of B.I. incumbents, is it they're too fixated on

B.I. They don't think about themselves as a human, as an employee that you buy, and then think about what tools do we need to give the employee? Like, well, we should let the AI agent move things around in the B.I. That's great. That can answer a subset of problems, but that's not actually going to replace the kind of work. It's marginally better.

It's marginally better. Yeah. Interesting. Well, I think we're at time, but this has been a blast. John, thanks for coming out there. In person, round two. I think with our last one in Denver here, but yeah, thanks for doing that a show. All right. Thanks for having me. The Datastack show is brought to you by Rudderstack. Learn more at rudderstack.