Zenlytic's AI Integration: Revolutionizing Business Intelligence

0:00The wait is over. Dive into Audible's most anticipated collection, The Best of 2025. featuring top audiobooks, podcasts, and originals across all genres. Our editors have carefully curated this year's must-listens from brilliant hidden gems to the buzziest new releases. Every title in this collection has earned its spot. This is your go-to for the absolute best in 2025 audio entertainment. Whether you love thrillers, romance, or nonfiction, your next favorite listen awaits. Discover why there's more to imagine when you listen at audible.com slash best of the year. Welcome to the AI Chat Podcast.

0:42I'm your host, Jaden Schaefer. Every week and every day, we interview AI experts and bring you the latest in AI news. Today on the podcast with us, we have Paul Blankley from Zenlytic and Ryan Jansen from Zenlytic. Ryan is the CEO and Paul is the CTO of this incredible company that has raised millions of dollars to make your business intelligence better and easier. So today on the podcast, we're going to be diving into what exactly they're doing. Welcome to the show, both of you. Maybe we can kick it off to Ryan first to tell us a little bit about your background, what got you into this space in general.

1:17Yeah, for sure. So my background is I am kind of a zigzaggy background. You know, like in venture capital, they talk about crossing the table, like going back and forth from VC to founder, I guess. I'm a table crosser. So I started my career as engineer in my native Canada, but pretty quickly I went to sort of join a VC fund. I went to move to the UK actually. MBA, stayed there as kind of like an early employee of a venture fund. And I was there as a VC for a long time. That was good fun. But then I crossed the table over to Founds Analytics. And the precipitating event for that was actually just seeing how fast the tech was moving.

1:58And And this was actually in like, you know, a long time ago. This is 2017 or something. Way back in the Stone Age. In the Stone Age. And it's like dog years, man. AI is just moving so fast. But it's funny because the data point for me at the time was like, wow, things have come a long way since I was an engineer. I couldn't believe how far. I mean, the hard road improved, of course, but that just unlocked all these new software capabilities. And it was clear that data technology was going places, basically. but the thing that really stood out was not the size of the change but the pace and it was clear that it was accelerating and that has stayed clear all the way through to today right so like yeah even now it feels like we're speeding up and the acceleration never stops exactly right and like it feels like we're on a raft and you can feel the water's getting a little choppy just before the waterfall kind of vibes so uh that's what got me excited then and that's what keeps me excited now and that's how i ended up getting involved with ai very very cool and then what about yourself Paul?

2:55Yeah, so I'm like a nerd's nerd. I've been math and CS my whole way through. And Ryan and I actually met in technical grad school at Harvard. Ryan was coming back to brush up on his technicals, and I was just finishing up grad school there. And then right after that, we actually started consulting together. So that was like setting up these data stacks for companies. So that's like setting up Snowflake, BigQuery, all the sort of major cloud data warehouses and setting up analytical systems. And the thing that really got us started with Zanlytic was seeing the difference in what we were doing in grad school, the capabilities of the AI tech there versus all the old 20 year old business intelligence tools that we were setting up.

3:39So really, there's fundamentally going to be a change in how people actually use data because AI has advanced to the point that a lot of the things you want, you can just ask for as opposed to, you know, have to figure out some old interface. Okay. That makes a lot of sense. And then I guess over to you, Ryan, how would you say like, well, what's the main problem that you guys see you're solving the main issue that your customers had that they're, you know, coming over to you. So what is the main thing you're delivering for them? Yeah. Well, the main problem is really what I would call self-serve BI, which is like using business intelligence without being a technical person.

4:18If you're a big nerd like Paul or I and you know, you know, SQL or Python or something, you know, language like that, it's incredibly powerful what you can do with, you know, a cloud data warehouse these days. For the vast majority of people, they don't know and they don't want to know how to write SQL queries and things like that. So they're kind of stuck using, I guess, dashboards is the name of the game, right? So like dashboards are fine. But you know, the problem is, is that people, start there. And if they have to go any deeper, if they have to double click on anything, or they want to drill down into something or ask a follow-up question, it's an email to the data team.

4:52And they're called quick data polls. That's kind of a joke in the industry because they're never quick. The average person, the average organization is like maybe 60 people for every data team member. So that person is meeting the needs of 60 people. These questions after a few back and forth and you want to get the results daily instead of weekly and you know you realize there's an error somewhere and there's you know you you're iterating on this uh we're talking you know days or weeks you know so our goal is to uh make that faster uh make that instantaneous frankly uh by incorporating this sort of always on instant you know ai chatbot that can analyze the data and field, you know, a solid majority of those questions.

5:39Okay. So Paul, talk to me a little bit about like, you know, Ryan's mentioned this AI chatbot. Talk to me a little bit about how AI is being integrated into your platform, what that really looks like. Is this an API? Is this a, you know, a wrapper on an API to something like open AI? Is this your own in-house AI? Like what's the, what does it look like for you guys having AI integrated? Totally. Again, And this is a really important point, actually, because there's a lot of people trying to do, you know, OpenAI or Anthropic or something like that to SQL. So you ask it a question, it'll generate some SQL, and then it'll go and run that on a warehouse.

6:18That just does not work outside of some super trivially simple situation where you've got like two tables. It's just in any organization, the definitions themselves are super complex. Visa has a team of almost 100 people just to calculate what revenue is. These things just get really, really complicated. And you can't just let an LLM kind of guess at some SQL that might or might not be the right answer, especially when you have an executive who doesn't know SQL and can't verify that it's doing the right thing. And that's just not a situation that can happen, not for a public company, not for even a mid-market company that would make a really bad decision with a wrong definition there.

6:59So the definitions have to be correct. And the way we solve that problem is that we have a semantic layer. That's where the data team, the people who have the context, can go in and define exactly what those definitions are. And then the LLM, instead of having to guess its sequel and figure out all these complicated definitions, just basically picks from a menu of definitions. And then just says, hey, I want this metric over this time, sliced by this thing, filtered down for this. we can then compile the SQL deterministically to make sure that everything is always right. So the way that works on the LLM side is that we're using GPT-4.

7:37That's really the only LLM out there right now that's capable of doing this really well. We're using GPT-4, but the really crucial point is that we have that semantic layer. That's a different architecture from every other company that's doing this. And it's a really critical difference because you just cannot trust the LLMs to guess it's SQL right every time. Right. No, yeah, that makes a lot of sense. I think the hybrid is definitely the right approach here. You know, you mentioned using GPT-4. I've heard from a number of founders working on different projects, complaints with just the state.

8:13Well, I mean, first off, how long have you had access to that? Do you have early access or is that relatively new that you guys got on? Early access. We've had access since they launched the API. Oh, awesome. Okay, that's awesome. So I've heard a number of people complain about, they say from going from GPT 3.5 to GPT 4's API, like they have like timeout issues and some things like that. How have you guys experienced that? And if so, is there something you've done to get around it? Or has it been fairly similar to 3.5? No, so the latency is much worse. And that's just kind of the price you have to pay.

8:46You have to architect the application to be able to support latencies that are higher like that, which isn't fun for anybody. Like no one likes that. but you know you when you're when you're looking at like you know eight seconds versus three seconds or waiting three days for something eight seconds is still the no-brainer um but but it's still not fun um we had a lot of problems with latency with them probably two-ish months ago but to their credit they have sorted a lot of that out um you know average request times are down probably like a third of what they were two months ago so they're they're overall doing a really good job every now and then they'll have a bad day and things will take longer but uh luckily our systems are architected to be able to like handle those bad situations okay that's awesome yeah because i heard complaints about like essentially if you send in a bunch of requests all at the same time there's issues and stuff but so i guess there's work workarounds in the architecture that's awesome ryan talk to me about like who who your ideal customer using your platform right now is like walk me through a use case.

9:49Paul mentioned Visa using tools similar to like calculate revenue and stuff. What's an ideal use case? Like I'm Coca-Cola or I'm some business and I'm going to use your platform. What am I going to do with it? Yeah. So our users are actually sort of like there's two groups in an organization that touch the product. The first is the data teams that sort of deploy. They're also the teams that are setting up data pipelines and the data warehouse. they're also the people that used to be doing those quick data polls. So that's side one. Side two is the end user, the self-serve user. And those are the non-technical folks who are just trying to find the answers.

10:28And we get usage from both sides. It's designed for the non-technical user, but that means the technical user can also use it to get faster answers. So, you know, quite often, even as someone who writes SQL, I'll use Analytics to, like, generate a SQL query and just iterate from that instead of starting one from scratch, for instance. Yeah. But the name of the game is the non-technical user. And to give you an example of, you know, what that looks like, if you say you were a performance marketing manager, you know, for some sort of, you know, leading DTC brand, for instance, you'll probably have a whole bunch of dashboards that you're checking every single day, monitoring campaign effectiveness, monitoring, you know, user traffic flows throughout the app, monitoring funnel conversion, checkout rates.

11:12You just, you'll have a giant dashboard of stuff that's monitoring that. And maybe say one day, you'll notice that, for instance, your ROAS is going down. Your return on ad spend on one of your dashboards. You'll see a bit of a dip there or something like that. That, under most circumstances, you'd actually be asking the data team for help at that point in time. And you'd say, all right, I need to understand what's going on with this situation so we can rectify the problem. I'll email the nerds and be like, all right, can you give me the breakouts of these by X, Y, and Z? I want this by first product purchase, last product purchase, whether or not they're a premium subscriber, all these slices at once these combinations.

11:49I want this trended versus last month at this time is this a real problem? Trended versus last year at this time and you'll ask all these questions to really sort of get your head around the problem and get the shape of the problem. That's a slow process. Now it can be handed handled automatically and instantaneously and you can actually just fire those away and iterate on those live just by updating those charts and updating that data and say, all right, tell me more about this SKU. Is this a problem? Or even use some of the advanced functionality in the analytics to break the problem apart. And you can actually just click and over drag over and say, all right, explain to me why this dip is happening.

12:26And it will build hypotheses for you. It'll give you half a dozen ideas. Oh, it turns out this SKU or this combination of this SKU and this particular campaign have been underperforming their historical ROAS, you know, also look at this particular promotion code. This is also not doing well. You know, did you know there was a stock out here? So like this particular, you know, SKU is also underperforming. And it'll actually break that down into ideas for you that you can action. So basically, you know, your objective with any sort of, you know, data research project like that is to take a big, vague problem and break it down into a bunch of smaller actionable problems with solutions and and we just accelerate that process and i guess we you know the build measure learn cycle uh we're the we're the measure part okay very very cool paul tell me a little bit about um you know like what are some of the differentiators between zenlytic and other business intelligence tools that are currently in the market how are you guys setting yourselves apart and being unique in in kind of in the space totally so i would say the first thing is our one One of our taglines is that we're the world's first self-serve BI tool.

13:35And that's a bold claim because BI tools have been claiming to be self-serve for a really long time, like since before I was born. But the reason that that's actually true is that, you know, we can dramatically alleviate this problem that Ryan's talking about, which is that quick data pool. Data teams spend, you know, 60, 70 % of their time just answering these ad hoc requests. And that's a huge amount of time just doing very basic, like filtering, slicing things that are just a little bit too complicated for end users, but really just mundane and sort of mind numbing for the data team. So we can actually do that because we've got the full BI suite of functionality.

14:12So it's like all the row-based access control, the complicated, you know, BI stuff that you need to be able to, you know, deploy this in an organization of scale. and then we can marry that with the LM interface where you're able to just ask about the things you need. You don't have to know which table something's in beforehand. You don't have to know how to join these two things. You don't have to know exactly did we include returns or not include returns in that net revenue definition. You just ask for what you need and you're able to get it in seconds. So the main differentiator is just that ability to reduce those ad hoc data requests, speed up things for both the data team who don't have to answer those and the end users who don't have to spend, you know, a week commenting on the Jira ticket to get some data that they need.

14:57Okay, very, very cool. So Ryan, I want to ask you a little bit about the early days of Zenlytic and kind of founding and bringing this together. So I understand, right, you and Paul, you meet when you're at school at some point and you start doing some consulting and whatnot, if I'm getting the story straight. What like inspired you to be like, oh my gosh, we need to start a company. What was the first steps that you kind of took in that? And I guess what made you decide to go from being VC to consultant to now all of a sudden you're starting an actual software company? It's a big step, right?

15:31Yeah. Yeah, I'd say we had an unconventional path to it. And I'd say that we, yeah, for context, after Paul and I met, when we graduated, we started this data science consultancy. that was doing everything from building data pipelines all the way through to data-driven strategy and involved a lot of sort of, you know, touched a lot of the various BI tools that existed out there. That was always sort of, there's two intentions there. The first was to sort of bootstrap a more chest to go into product mode. The second was to really understand the shape of the problem. And so the intent was always to sort of shift into product mode once we had a, you know, really good idea of how we're going to tackle this.

16:13Okay. I'd say we had a vague inbound notion that, you know, self-serve was not working, but then after having had a chance to live the lives of really both sides of the problem and a lot of different lenses, basically, I think that helped us sort of refine exactly what we had to do, what we were going after, you know, what the shape of the product was going to look like. So that was the first big unlock. The second big unlock actually came about a year after we went into product mode. So we set out to build the world's first self-serve BI tool. And I think we got a lot of the way there. Was it just the two of you working on it?

16:50Started out with the two of us. We're now up to 11 and hiring, if you're interested. So if you're listening to the listeners that are awesome, please reach out to us. Yeah, so it was the two of us at the time, but we'd sort of started building for that first year. and we made a tremendous amount of progress but we didn't even realize at the time that self-serve probably wasn't possible then and it's like it's the old vc adage where it's like all right vi tools have been trying to do this for 20 years like what's changed basically and for that first year nothing fundamental had actually changed we're doing a better job rounding out the corners and we were getting rid of a lot of the sort of blockers to effective self-serve from a UI perspective.

17:33But the second big unlock for us was the LLM revolution. We'd always had some language capabilities. In fact, Paul and I, when we were studying, that was the year that attention is all you need came out, which is like the seminal, like that's the foundation for GPT, basically. We'd always had some language capabilities, including the great, great grandparents of, you know, GPT-4 and GPT-3 or even two, right? As soon as things started to accelerate, we saw the opportunity ahead of us. And I guess we'd always expected that would happen at some point in time. I think it's happened faster than everybody expected, including us.

18:09But, you know, we realized that was the big unlock that was missing. And that let us, you know, first improve, you know, fidelity and everything, all the great stuff that comes with these models. The other big thing is it let us move from a single question search paradigm into a chat paradigm. And that's important because these quick data polls never end with a single question. It's always like a conversation. There's refinement and iteration and things like that. And by adding the ability to do that, that was the necessary unlock. And now that's what's changed. Companies have been trying to do this for 20 years.

18:43It wasn't possible because this technology just wasn't ready yet. So that was the second big unlock. As soon as we did that, and Zoe, the analyst was born, that changed everything that really allowed us to, you know, go way, way further down, like what's possible for self-serve than ever before. Okay. Paul, tell me a little bit about Zoe, but also tell me a little bit about integrating AI. So like, I understand, right, it's running on GPT-4 right now. Did you have 3.5 in there? Were you guys on like DaVinci before that? When did your like journey start in that, right? So that's the big unlock.

19:17So when did that, when did that kind of begin? Yep. So like Ryan said, we originally started on the old school models, just like parsing text using BERT and other sort of progenitors to the current LLMs. We started with OpenAI with the DaVinci model before they had 3.5. And then we've basically seen our own capabilities just grow as we've used increasingly more advanced models. And then we test all the models on the market. Like, you know, we test Anthropic models, Google models, OpenAI stuff. So it's like we really, you know, look around to make sure that we're using the best thing on the market.

19:57And our architecture is very able to swap those out. So if someone else passes OpenAI, we will happily use the best foundation model available, basically. Have you seen any like features in other AI models that you're like, oh, that's kind of interesting. Like I know Anthropic does a lot with like PDF uploads and that kind of stuff. Are you, have you seen other ones that seem promising? So the really large context windows are promising. The trouble with the context windows is even if they are really large, it's difficult to get the AI to actually know what's going on in the middle. It does a very good job at the beginning, very good job at the end.

20:34With those really large context windows, a lot of the times things get lost in the middle. so you know context windows that are larger with you know good attention throughout would you know be be really game-changing so that's one of the things that's most appealing about the anthropic models okay yeah that makes sense because i guess what it's probably trying to do is grab everything and do some sort of variation of a summary which the beginning and the end but what you probably want those models to be able to do eventually is to be able to go through the whole thing and highlight every time they think there's a key finding then they like condensing it condensing it so yeah that makes sense that that'll be very interesting it's super fascinating though those context windows but you know there have been a few studies that have shown that they don't retain the information in in the middle of a long context window and i find that so so interesting because that's how humans work too right i was just thinking that i'm like that's exactly like my problem i'm like if you gave me a pdf and said like give me the data from this i'm like skim the beginning skim the end here you go yeah so it's It's just neat how they seem to be coming the same direction that we are.

21:37So that's a sign that we're on to something in terms of intelligence. Yeah, we got to get the AI to stop cloning us mere mortals and go for like the super geniuses. We got to get Einstein's brain. That's what AI needs to be based off of. But then the question is, is the input, is it an input problem or like the actual like algorithm problem? Because it also might be that all the stuff you need to predict what's going on is either in the beginning or in the end. In everything that it's been trained on. because that's how humans read right it's like you don't stick something super important in like paragraph eight of 16 right so right you know it's also maybe just math in the sense that all the important stuff is at the beginning or is at the end oh that's a good point that's a really good point um ryan tell me a little bit about uh the journey of zenlytic so you know at the beginning you said you guys decided you're going to kind of work on this project you get started you have like these unlock periods um did you you guys raised funding correct for zenalytic what what was the when did you guys do that what made you you know want to do that tell me a little bit about how that how that went for you yeah so we've raised two rounds of funding um we raised raised our uh pre-seed round pre-product led by primary ventures okay uh and then we raised our seed round uh led by bain capital ventures uh it's a total of five and a half million give or take us and um yeah i think i guess my advice for internet fundraising would be uh i mean know what you need the funds for for sure uh but also you need to feel that pull you know there has to be a you have to have the sense that you it's ready time to fundraise and like yeah i say that because i mean first fundraising uh you know is expensive and it's like you know don't fundraise if you don't have to right uh so don't do that but also i think that unless you're feeling that pull, it'll be a lot harder to successfully fundraise.

23:26I think a big part of a successful fundraise is when you're having conversations with investors, they're going to feel that pull through you and that'll actually make you more effective at pitching. So I guess in our case, the first sort of pull we got was like, it's time for us to get out of consulting mode. So that was the first thing is like, all right, we see the opportunity here. We've validated that there's a huge, huge problem in a big market. And it's kind of an interesting place because a lot of the big leading BI tools from then until now have been sort of languishing after acquisitions.

24:02You know, it's like a bunch of them just got bought up by some of the bigger companies and they've been sort of undergoing transformations and transitions and changing focus and changing owners. And, you know, that does not lead to a fast-paced, innovative environment. So there's, you know, we figured this out. And as soon as we saw the opportunity ahead of us, that got us really excited and that said all right it's time to tackle this full on and that helped us fundraising the pre-seed motion uh i think at the seed stage uh you can we started to feel the wheels are rattling off the tracks in terms of how fast we could expand uh okay when we when we closed our seed round we were three and a quarter people uh which i think is quite undersized by most companies yeah yeah that's pretty small pretty lean yeah uh as part of our philosophy we'll always be like a you know super dense lean team uh i think that's just how we like to run things and i think we can get a lot more done a lot faster that way just by having a handful of really really great people um but anyways when we were raising the seed round it was yeah we're three and a quarter uh wheels are falling off the tracks we were totally just kind of stunned locked out between product development and speaking to new people and we said all right It's time to add more hands.

25:14We're feeling the demand for this product. People are getting excited. It's time to add more hands around the go-to-market side. You know, add more talent there, add more engineering capabilities. We could feel in almost every area of the business, we had a really good use for additional capacity. And we said, that's time. Let's go for it. Okay. Yeah, that makes a lot of sense. Paul, tell me a little bit about what you think were the hardest technical aspects of getting this platform off the ground. What was the hardest things you had to struggle with to get right on this thing? I think there's three things that I'd say here.

25:47The first one is figuring out our explain change capability. That's basically this very esoteric field of algorithms we actually had to push forward to be able to answer those why questions. The other thing is critically just like the semantic layer. We made some improvements to underlying SQL compilers that involve a lot of complicated stuff, like graph theory and other, you know, just really complicated to actually get that out. Those are sort of like individual crux moments. The thing that I think is really underrated in all software businesses is that it's just complex. You know, we've been building the ITool for several years now, and you just find all these little problems, these little issues that just come up that you have no way.

26:34There's no way to find them or anticipate them outside of just experiencing them. So the example with an established company is Salesforce. You can't just replicate Salesforce by copying basic CRM functionality. Salesforce has figured things out in two decades of building a CRM that all other CRMs just have to go through that same painful experience to be able to replicate those features effectively. So it's like going through that, that is really the thing that's hard to copy. That's difficult things we've figured out in the UI, difficult, complicated, you know, bugs that have come up, things that we've had to sort out.

27:08So the complexity itself is one of the cruxes that's often missed in software businesses. Yes, I definitely can see that. And then Ryan, question for you. You're talking about raising a pre-seed and a seed round and whatnot. What was the actual process like for you? Did you go and just cold outreach? I know that you had worked in venture capital before, so were you tapping into some of your previous contacts? How did you put that round together and what was that journey like? Yeah, for sure. One last one, because we've got to bounce back to this. But I think that it's interesting. So even though I was a VC for a long time, and I have a network of friends in VC, which I think is why a lot of people get into it, and so I think it ultimately fundraised, I actually didn't use that at all in any of our fundraisers, really.

27:54Really? Once or twice. But it always just happens so organically and kind of quickly, I guess, that I didn't have to really tap into that. And I think that the greatest lesson I learned as a VC is actually not that the network is important but I think it's important to if you do your homework and you get people informed and excited about what you're doing before you're fundraising and just maintain those contacts of people as an entrepreneur I feel like just being prepared is 90 % of the battle you know like once you if you come to a firm that you know that you've been keeping up to date you've been calling shots and making those calls and saying you know next time we meet we're doing this and then you do that, that is just super impressive.

28:38That's more impressive than telling anybody your story, just letting them live it. And if you've done that and you say, okay, it's time to fundraise, I feel like you're going to get a lot more people leaning in. So do your homework. Most of the fundraising happens before the fundraise is my big takeaway. Yes, I think that is something critical a lot of people are learning in today's field with a lot of these different AI startups happening. I know a lot of listeners definitely appreciate that advice. Paul, what's one piece of advice you feel like you could give to technical founders on, you know, perhaps people right now starting new AI startups?