This podcast episode is sponsored by ChiefAIOfficer.com, offering training and certification through the International Association of Chief AI Officers. Interested in a new career or leveling up your value in the marketplace? ChiefAIOfficer.com can help. Welcome to the podcast. I'm your host, Chris Daigle, and the following is an interview with Avi Perez, co-founder and CTO of Pyramid Analytics. Pyramid is leading the way on data privacy, data security, and business intelligence all applied to the lens of generative AI. In this episode, he unpacks what generative BI really means, why it's changing the way companies are interacting with their data, and how this technology can empower business leaders to make smarter, data-driven decisions. Welcome to Using AI at Work. I'm your host, Chris Daigle. Each week we'll be learning how today's business owners, entrepreneurs, and ambitious professionals are getting more done with smart use of tomorrow's tech. Let's get started. Welcome to the episode, everybody. I'm really excited to have the discussion that we're about to have today with Avi Perez, the co-founder and CTO of Pyramid, which is known for something that we haven't covered a lot on our podcast, and that is the data privacy, data security, the business intelligence side of leveraging generative AI in your business. So, Avi, with that brief introduction, I'd love it if you would share with the audience more about what Pyramid is known for.
SPEAKER_01First of all, nice to be with you today, Chris. Thank you for having me on your podcast. So Pyramid is broadly speaking a business and decision intelligence platform. We typically earmark ourselves towards the mid-market and enterprise scale customers, but we have plenty on the small end as well. At the moment, in terms of where the product and the company sits, we are at the forefront of what we call generative BI or Gen BI, which is the grand idea of taking the power and frankly the brilliance of generative AI and blending it with classic BI functionality. That's what Gen BI is all about. And Pyramid sits at the forefront of that particular um sort of ecosystem. And to put it in a nutshell, I mean, it's not something new. It is a strange thing. People often ask us, oh, is this what you've been doing for the last years? No, we've we've been in the business of blending AI into pyramid for years. I mean, we're talking a decade. And it's a very natural fit for AI to be to do with data and data analysis. If you think about it, those things go very well together. But with generative AI coming to the fore in the last, let's say, two-ish years, maybe a little bit longer than that, um, it became very obvious that I'm a great home for it. And um, you know, bringing natural language, anything, to the complicated business of data analysis, which really means data-driven decisions is a very natural fit, and that's exactly where we play. And it's one of the things we excel at in particular.
SPEAKER_00So the audience of this podcast runs the spectrum of uh very capable, you know, business professionals who are already uh early adopters on Gen AI to those who were just getting started. And for those who haven't had a lot of exposure to the concept of business intelligence, how would you define it?
SPEAKER_01So, you know, business intelligence has been around for years. I want to say easily 30 plus years. And the whole idea of it was the idea that we can take corporate data, expose it to users in a format that is easily digestible, what is classically known as visualizations. You've seen them before, eye charts, bar charts, grids, and use that to drive um smarter business decisions, business decisions that are based on data rather than gut feelings. Um, you know, back in the day, it was very cool to have, and big companies had it because big companies could afford it. But in the current era, everybody understands, with the amount of data being collected, that you can make a lot of smart decisions if you knew the data behind it. And with big data and the vast amounts of data collections that's now for granted, taken for granted, of course, you're going to make a decision based on data. So, business intelligence is the entire ecosystem of taking data and using it to drive better decisions for your business. It doesn't have to be a profit business, it can be a for-profit, it can be a non-profit, it can be whatever you want, it can be governmental, but that's the broad idea. And uh, and again, to bring it back to AI, AI is the business of making that exercise much simpler. Um, the biggest headache in the space all along has always been that to do really clever things with data requires a degree of sophistication and complexity. And lots of users are non-technical, they're they're they're excellent at making business decisions, but they don't know how to play with all the knobs and turn the buttons and so on and so forth. So if you think about it, the whole effort for the last at least 20, 25 years is to make it as easy as possible and to give more power into the hands of less technical people. And that's what the AI component is effectively trying to do in the BI component to simplify that entire exercise.
SPEAKER_00That makes sense. And thank you for sharing that. Um so I would imagine that within the the lifespan of Pyramid, that AI has been in use, maybe not generative AI, obviously, but were you guys early in with machine learning or data science and that sort of thing?
SPEAKER_01Absolutely. It's a um pyramid is based on three disciplines, blending three disciplines together. One is what I just spoke about, the classic visualization of data and you know, the pie charts and reading the numbers and how you get there and so on and so forth. That's one. Another part of it is what we call data preparation and data fabric operations. It sounds very esoteric, but it's basically taking your you know, your database together with your spreadsheets and finding a way to actually pull the data out so you can use it in the visualizations. And then the third sort of leg of the stool um is the machine learning. And the machine learning is clever because you're going to use a machine to somehow read your data and teach you something about it. It's less about, I call it machine teaching, because the machine is going to learn your data and then teach you something.
SPEAKER_00Nice.
SPEAKER_01And AI is a component of that, not exclusively ML, by the way. Um it's not only about ML, but AI is heavily involved in that. Generative AI is very machine learning driven, right? Um based on deep neural nets and large language models, all that exotic stuff. Um, but it's a big part of it. And the trick, of course, again, like everything else, is how do I take something super complicated like machine learning, which I think is one of the hardest disciplines to learn as a non-technical, non-data science, non-computer science uh person and use it. Because if you could use it, wow, imagine what you could learn. Like, don't tell me how much money I made yesterday, tell me how much money I might make tomorrow. Yes. That's interesting. Or the bigger one that I always tell people is tell me what I need to do tomorrow to make more money. That's even more interesting. Not how much I might make, what do I need to do to make more money tomorrow? And both of those forward-looking, um, what's called predictive and prescriptive um analytical disciplines require the magic of machine learning, these advanced capabilities, and that stuff is magical, but it's really, really difficult to use. And so the grand idea is could we use AI technologies to help a user use it? You know, the best way to explain it is is with my Star Trek analogy. You know, you never saw Captain Turk on Starship Enterprise saying, okay, we're flying to beta 1, 5, 3, 2, 4, and then spend the next six months programming how to do that. So he he just spoke to the computer and said, fly me to beta 1, 3, 2, 5, or 1.4, whatever the number is. And off it went. Computer understood where you meant, what you were saying, how to get there, which planets to fly past, and which solar system to fly through, or not solar system, it's galaxies, it just knew what to do. We kind of want the same idea in AI. I want to wake up and say, you know, computer, look at my corporate data for the last X number of months and tell me what I need to do tomorrow to make to be more profitable or to be more successful or to make my patients healthier or to invent the next product. I don't really care about the pie charts and I don't care how you get there, just tell me the ending. And that's basically the grand epiphany of AI in general, but specifically in BI, that's where everyone would like to go. If you think about it, yeah, that is what all of us want to do.
SPEAKER_00Absolutely. And this concept of being able to interact with my data in a natural language environment or interaction is extremely interesting, uh, especially for those, the audience who is the non-technical business user. What does that look like? Now I understand that you guys have obviously your own frameworks and software and that sort of thing, but for somebody who wanted to perhaps uh do it, demonstrate it in their own business, kind of what would that what would those steps look like? We don't have to get into the nitty-gritty, but what does this involve from that user?
SPEAKER_01So look, it it's a it's a it is actually as as we stand right now in the generic software space, trying to implement genitive AI, that is the big question. How do I take this and implement? Um I'll give you the generic answer for software in general, and I'll explain the pyramid story. So the generic idea is something like this. Um, you know, we've all played with ChatGPT and the chatting tools where you type a question in and it goes and gets an answer. This is the most generic um implementation of generative AI. It's very cool, very clever, but it's really quite detached from business applications, if you think about it. It doesn't answer my question on my problem, my data, my website. So the big movement at the moment is how do you make that all come together? And the answer is you're going to use the APIs offered by the vendors of those LLMs, um, and you're going to program into that API what's called a prompt. The prompt is anything from the question that the user asked all the way through to some, let's call it, background information to instruct the LLM how to respond. So imagine I wrote a question, you know, um, I'm on a baking site and I wrote, you know, bake me a chocolate cake. Um, I could just ask that question today to ChatGPT, and it might come up with a random recipe. But maybe the baking site is so clever, it knows what ingredients I have in my uh my pantry. And so it says that, you know, I've he asked for how to bake a chocolate cake. Just so you know, he happens to have, you know, um, you know, a pound of chocolate, two pounds of butter, no milk, um, he's got three cherries, and he happens to have an oven, but it can only go to 150 degrees centigrade. I won't say Fahrenheit, I don't know if I'm right about. But so if you knew all of those ingredients or those parameters plus the question, the chatbot would respond back, and it's not the really chatbot, it's the large language, would respond back with something far more intelligent to just the basic question, how to make a chocolate. And if you're going to implement generative AI in a generic sense in whatever it is you're doing, you're going to have to figure out how to use the API, seed it with those ingredients, and then the chatbot, which is the veneer for the LLM, will then come back with a far more intelligent response to the user, and everybody's happy. Now, when I drag that into the BI space, it gets even more complicated because the ingredients are, well, how can I put it? You've got lots of things going on. First, I've got the the kind of data that I'm using, I have the data itself. Um I have the context of the questions I've asked in the past. And um, and then the the the the large language model, the LLM, has to respond back to me with an intelligent answer. And in the pyramid storyline, it's extremely problematic because no one wants to send their data to the LLM in the first place, you know, and second of all, it would never work. Imagine you're sitting on, I don't know, uh a million rows of data in a spreadsheet. It's a classic workbook in Excel, yeah, it's possible to upload an Excel workbook. What happens if you're sitting on 10 million rows or 100 million rows? Or even I say a billion rows, how would that ever work? I mean, you'd be spending three hours waiting to upload the data just to get an answer to your question. It doesn't practically work. And so the first headache that a piece of software like Pyramid has to fix is that headache that the data actually can't move to the LLM. We have to pivot the whole thing on its head. We have to describe your data to the LLM, which is explaining what ingredients you have in the parentry. And we actually ask the LLM to come back to us, not with the chocolate cake, but the recipe for the chocolate cake, the recipe for the answer to your question. And then Pyramid, which is the software that obviously we've built, so this is how we think of it. Pyramid is going to take the recipe using your data and your ingredients and go and bake the chocolate cake for you and hand it to you. And because the data doesn't move to the LLM, it stays local and stays within the pyramid context, it's entirely doable. It doesn't matter how big the data is, it's much faster, much cheaper. Um, and more importantly, it means you don't share the data with the um with the LLM, which a lot of companies have a big issue with. It's it's effectively sharing proprietary, confidential, dare even suggest top secret information. People are horrified by all of those ideas. That's probably, by the way, one of the big um sort of issues at the moment in the marketplace is how do I use this clever stuff without being exposed to the risks of that. But that's that's a side issue, just it practically doesn't work. On top of that, LLMs are notoriously weak at doing analytics, which is a surprise to most people. They're brilliant at answering questions and doing very clever things. But if I handed it a relatively big set of data and asked it to do mathematics on it, it does a pretty lousy job, which most people don't really understand. It's not designed for mathematics. If you ask me what's the future of them, it's it's that mathematical, deterministic, um mathematical activities. That's the next sort of frontier in the deep neural net AI space. That's even the right box. So, what Pyramid does is it solves that by flipping the whole thing on its head. Like I said, we asked the LLM for the recipe, but we are the ones baking the cake without moving the data. And that's why it works particularly well for us. And I would argue that's how it might work for most other generative AI applications where you're talking a vast amount of data. You literally cannot upload it to the LLM for it to read it and understand how to answer your question, especially if the data is, as I said, private, top secret, or something that's not publicly available. Hopefully that makes sense.
SPEAKER_00It it does, and it's it's very helpful because I I get that question a lot because most of the clients that we work with are the individuals who we're training. Chat GPT is kind of the default, of course. People have their preferences with Gemini or Claude or anything like that. But um, they are uh on day to day, they're concerned. There's a couple of concerns. We want to bring it into the company, but we're concerned how our employees would use it, right? And obviously, the number one concern is our uh exposing our data to something that would put us at legal risk, compliance risk, anything like that. So, but then you you go to you know Chat GPT, especially if you have teams uh or an enterprise, and at the very bottom it says we do not use your data to train the models. How do you reconcile your perspective on the risks with the claims being made by the LLMs?
SPEAKER_01The first thing, and this is exactly why pyramid's architecture is the way it is. In fact, in the diagram I'm going to show with you now, you can see what we call the horseshoe diagram. If you look carefully at it, you'll see that um the user would interact with the LLM. Uh, they would send in their request. The LLM, again, what we call produces the recipe.
unknownYeah.
SPEAKER_01Pyramid takes the recipe, what we call baked the cake, which includes uh querying the underlying data, resolving the response, and then sending the response back to the user. And the only part that we might expose back to the LLM is where we ask the LLM to re-articulate the results into plain language, which is optional and in many respects is very, very minor. But the raw granular data is never given back to the LLM. It can't, it functionally doesn't work anyway, and and even if it did work, you would be horrified, exactly like you said. And because of that, the data is in theory never exposed to the LLM. So the LLM can't really use your data to retrain itself and it can't use the data to optimize, or and no one can learn from it because that they can't see it. Both practically wouldn't work, and both um from a security perspective, it wouldn't work. Regardless of that, the biggest concern that people have um because they misunderstand how it how well from the pyramid side it doesn't work. What people then misunderstand is um, yeah is my data going to be used by an NLM? And the answer is there is a possibility from the question itself that the LLM could learn from you. Oh, well, look at that. You know, this guy is a male aged, you know, 40, um, and he you know, he lives in Osman, Texas, and so therefore, he probably, you know, loves to go to, I don't know, he loves burgers. So we're gonna answer the question this way. So they can learn from you in like any other solution, like Google or Facebook, and use that to manipulate the results. I mean, that's an issue you have in general with all of those public domain applications. But as long as you don't send, you know, 500,000 rows or 500 rows of corporate data to the LLM, it really can't grab much from you and going, oh, well, look at that. This is the profitability of ACME Corporation. And good, let's tell everybody when they ask what's the next most profitable startup tomorrow, and poof, it's ACME because they learn from you. So this is comment number one. Um, it's an issue. There is a bit of a leak in the fact that the user can ask a question and share something secretive in the question. Uh but you know, these are far and few between. And then again, with pyramids approach, regardless of how you sort of you know peel the onion, the real corporate data is never sent ever to the LLM, except for a few minor fragments that could slip into the question backwards and forwards. But generally speaking, it can't happen. What's more important about um our approach again, and I don't think it's unique to us, by the way, I think other people are working on the same idea. It has to work like that. It functionally doesn't work any other way today, is the amount of data we can address is dramatic. Um, you know, in the billions of rows, again, because we're not moving the data, but also um in the case of pyramid, our engines are very adept at understanding how to query the data and produce a good result at speed, which, for example, the LLM itself might not know how to do. So, you know, we often ask, well, what's the difference between pyramid querying my database and me asking uh an LM to produce a Python script, um, which we then connect to my data and extract the data and build me a visualization. But what's the difference? Um, first of all, the Python is going to be built each time from scratch. You have to figure out how you're gonna run it. You're gonna have to figure out how you plumb it, you know, let's put all the connection details in. And and even more problematic than that, every time you ask the next question, each Python script is completely agnostic of the previous one. Whereas in the pyramid cycle, which again, if you look at the diagram that I've got on the screen, uh, which looks like a horseshoe, by the way, uh, with a bit of a loop in it, because it can keep going around and around, we're able to keep what we call state. We're allowed to remember the session. So we can look at your question from three steps ago and understand the context to the second, to the first, and then to this question, let's call it zero. So three, two, one, zero, and everything will make sense because we keep dragging the session along with us in that loop. Whereas, you know, standalone questions don't work like that. So all of that is a plus sign. But let's ignore that for a second. You still have the issue of how did I get my data to the LLM? So, in summary, there's a lot of things going on, there are a lot to unpack. But in summary, moving your data to the LLM doesn't work. If you could move it to the LLM, you have the problem of sharing that secretive data, which is at an issue. And these are both problems. We understand, we agree with them, and I think that people need to understand that it it's it's a non-it's a non-starter of an idea anyway. So it's something to be less worried about. However, if your users are given access to ChatGPT and they can upload a spreadsheet, even a spreadsheet, to chat GPT and say, hey, analyze this for me and tell me what I need to learn from it. Yeah, you are at risk. Not because I don't think you know someone's taking that spreadsheet and necessarily going to copy and paste it somewhere, but you're exposing yourself to sort of who knows what can happen with it and where it could go on what those models can do with it. So there's lots of answers there to to sort of clean from and but hopefully it gives you some idea of how we think about it and what could happen.
SPEAKER_00It does. And I'm thinking about this from the from the user perspective. So they they rolled out Chat GPT usage, they did some basic training, let's say, for Company A, and now Um, you know, somebody from a department is ready to use it. They watched a YouTube video on how to do data analysis with a you know, with a spreadsheet and Chat GPT. They're like, oh, I'm gonna try this. A couple of points that that I hadn't really considered before. With or without the data, the question itself, the models are smart enough to infer perhaps proprietary operations or something that was unintentional, had nothing to do with the data. It had to do with the way that the question was asked to the LLM. That's very interesting. And that's something I hadn't considered before. Um, and then secondly, for I know that you guys, ideal client for you would probably have those billions of records that, I mean, obviously more compute, more higher fees, and all that kind of stuff. But for those individuals who have concern, that they don't have a big business, maybe they're doing uh, you know, seven figures a year, but uh haven't broken through to eight figures, maybe they're doing less than that, but they still want to benefit from um that, you know, uh big business opportunity of business intelligence from their data. How do how does somebody like that benefit from something like what Pyramids structured, um, but may not necessarily uh like what size clients, what would be too small for Pyramid?
SPEAKER_01Well, look, I mean, Pyramid itself is a is a vast platform and there's lots and lots and lots of things. And this entire conversation around generative BI, believe me, is just one piece of a very big suite of tools. The product wasn't designed though, that it was exclusively for big companies, it just happens to fit that story well because big companies struggle trying to build a scalable BI solution that can cover lots of users with lots of different needs and you know, intersecting lines. It's designed for scale, but you can quite frankly scale it down. The question is, is it is it worth going to the effort? What do I mean by that is if you're a small company running your entire office out of a spreadsheet, um, you know, we're overkill for that, to be totally honest. But there are plenty of companies who are small but are not running out of spreadsheet, actually running out of still sizable data. And and I would argue even two million rows of data, which is you know, spreadsheet today is one million rows. So two spreadsheets worth of data, corporate data, whatever that happens to be, you know, you're starting to like, okay, well, you know, spreadsheet doesn't work. How am I gonna do it? So, regardless of the data footprint size, the way it works is you have your data in some database technology, a pyramid, uniquely in this regard, doesn't care where the data lives. Um, it can be anywhere from you know really fancy data lake technologies like Snowflake and Redshift and BigQuery, all the way through to some you know, SQL Server database have sitting underneath your your desk on a server somewhere. We don't really care, it all looks the same to us, it all works. You literally would install Pyramid or uh you know uh have pyramid running in the cloud and choose either way you want to go. You literally point it at the database um five minutes later, without moving the data, by the way, it's very important. We don't we don't copy the data to pyramid. Um we have reverse-engineered your data structures, and then you're asking your first question. It's literally a 20-minute exercise without doing anything complicated. That's how Pyramid rolls. And there are other competitors of ours in the market that um give or take, something like that. I think they take more effort to get up and running. And and to be totally honest with you, most of them require you to move the data into a a data layer that they either they prefer or their own proprietary data layer. And pyramid is a little different here. We don't expect it, mainly because we're targeting the enterprise market where you can't move the data. It's just too big. You know, you've you've built yourself a five-petabyte data lake in Snowflake. Yeah, you can't say, well, to get my first pie chart, I need to move it out of Snowflake into pyramid. You know, the tail cannot wag the dot. I've always said it's clumsy. So we we go to you and we don't need to move the data. But it extends to the AI component too. So the AI component, again, following that entire arc that I've been describing, you ask your question in pyramid. Um after analyzing the structure of your data for a minute or so, that's all it takes. Um, you ask your question, um, we go straight to the LLM, ask the LLM what's the recipe for answering this question given the data structure. Those are the ingredients, which is very descriptive. There's nothing, there's no data in it. And then it hands it back to us. We then ask the question on your database, whether it's a million rows, two million rows, or two billion rows, or two hundred billion rows. Pyramid runs the art and then hands back the result to the user. And that, believe it or not, is the same amount of time. It almost doesn't matter how big the data footprint is. It's that simple. Um, is it is that a unique story to pyramid? Is it the only way to do it? It's if you slice the market up carefully enough, you will find a very, very few tools that work that way, but conceptually, we're all heading in that direction because that's ultimately where it is to be. So the real question going back to specifically on generative AI, how quickly can I get there? Well, I can speak for ourselves very specifically, obviously, is very quick. We can measure it in under half an hour without doing crazy gymnastics. And I think there are other tools out there who probably uh could also do it in a short amount of time. It's just a question of you know the hoops you need to go through to get there. And like I said, I think the majority of tools are working on the basis you need to move your data. But if you're small enough, you kind of don't care anyway. If you have to move two million rows, all right so you spend 20 minutes moving it. It's nothing terrible. Um so it's just a question of you know where you want to play on that sort of continuum. But that's pyramid story, is how we think of it. And I would say to whoever's interested in it, to think of it like that. You know, um, I can't move my data anyway, so I need something that goes to the data rather than take the data to it. Hopefully that makes sense.
SPEAKER_00For the business owners that are intrigued about this topic and are ready to explore it more, what are some questions they should be asking themselves to know if they're ready or not for uh leveraging business intelligence from the data that they've already collected?
SPEAKER_01This is a great question. And in fact, it's I'm getting I ask this probably the number one or number two question all the time. What would it take for me? Uh what do I need to do to get there? And it's a simple answer. It's been the answer for all BI projects going back 30 years. You need to have good quality data. Primary requirement. If you come with garbage and you put garbage in, you basically get garbage out, the Gaigo principle. Um, so you must have good quality data. It doesn't have to be perfect, and which is actually the bigger the company, the more fictitious that idea is. It doesn't have to be complete as whole, but you need to have really good data for what you could get your hands on, because the AI components are gonna ask that data set for a response. And if you even you have a great question, and as they call it, the recipe that you you it generates is fantastic, and and all the processing is amazing. If all your numbers are off by 20%, then the response is gonna be off by 20%, and basically your answer is complete rubbish. By the way, that's the same problem in business intelligence in general, without even AI considered. Yeah, um, the the world's best pie charting tool is only as good as the numbers going into it. So that's it. If you've got your good you've got hands on data, good quality data, um doesn't have to be structured perfectly.
SPEAKER_00Is that a gut answer for the business? Or how does one know if their data integrity is high enough to benefit from this?
SPEAKER_01I would say on for a smaller company, it would evolve to somewhere between gut and and knowledge. And I think the smaller the company, um the operators know that what they're sitting on is good quality. They they know because it's theirs, you know, they did it. The bigger the company, ah, now you're asking a really, really, I would call it the $100 million question. Um, I think big companies inherently know they're not sitting on on the best sets of data, and they're continuously working on refining it, improving it, and working on it. But this is not a new space. It's as I said, it's it's it's been around for you for decades. So they already know the quality of their data, and they've already started to architect it for years already now. It's not a new idea that happened yesterday, and so they would would understand that the responses they get may not be perfect, but they've already baked that into the price, as I say. You know, they understand that what they're getting is not flawless and they know what they need to fix. But everybody in the marketplace has fully understood that you can't make good quality, data-driven decisions if the data itself is flawed. Structure notwithstanding. And most people today are uh educated enough to get some of the structure sorted out. Um, but if you don't have that, we have nothing to talk about. No data, we have nothing to talk about. So if you're a farmer with with you know a fantastic idea on how to improve the yields on your you know, your fields, and you're collecting no way of capturing any data points on it, we have got nothing to tell you. And I use that because farmers are all often pointed to as being the less technical least technical kind of industry. The irony, of course, today is agriculture has become very high-tech. Yeah. And you've got all these devices and and sensors out there collecting gobs of information, and yeah, you can you can completely change our entire uh sort of uh yield, I think that's the right term. Um and it's just a question of analyzing the data. And I use that because that's like a good example of going from one extreme to the other. You know, banking, finance, insurance, that's the entire business. Uh, an insurance company doesn't do anything except play with data, they don't make anything. Um, health insurance companies don't make you better. They basically take the the data coming in, work out who will be sick, slap 5%, and then charge everybody else a premium for it, and put the money in their pocket. It's a numbers game. Banking outside of the ATMs, you know, honestly, it's the same idea. So that data rich, data sophisticated. On the lower sophisticated end, not necessarily that's the wrong one, not sophisticated, but the lower data-centric industries are all catching up. And I think everybody's in the game now. It's a question of at the individual level, people getting out of bed and getting into it.
SPEAKER_00If you're enjoying this episode and want to learn more about how to start using AI at work, we've made it easy for you. For just one dollar, you can have full access to the Chief AI Officer community, which will give you additional training, custom software, daily training calls on AI tools, using AI automations, getting more from your Chat GPT sessions, and the business of being an AI consultant. Simply go to ChiefaiOfficer.com forward slash insiders to accelerate your AI journey. Now, back to the episode. So I would imagine that at all levels of business there are uh vendors who specialize in the analysis of the integrity of the data and then optimizing it in preparation for business intelligence evaluation. Now, is that something that someone should expect their BI vendor to do for them, or is that a completely completely different activity?
SPEAKER_01Um it's a good question. There's some tools in the market that have specific uh capabilities to help you fix the data. So Pyramid, for example, has a whole data preparation suite which is designed to help you fix your data, restructure it, fix it, clean it, so on and so forth. Um which is which on its own is sort of one one bag of tools to come to the table with. Yeah. Um having said that, if your data is messy enough and it's complicated enough, or it's coming from lots of uh lots of places, it does require some kind of expertise to clean it up efficiently, especially if you want to do it effectively, efficiently, and quickly. And remember, time is money. So then you might hire some kind of expert. And there are plenty of SIs in the marketplace that support SMBs to mid-market up to enterprise that are that's all they do. From the Deloitte and the Accentures of this world that do enterprise data, whatever, all the way through to mom and pop consultants, you know, um, who would come in and help you fix your data set. So and by the way, they might still use a tool like Pyramid to fix it. Then you can buy um highly specific tools for data preparation. Yeah, that's all they do, data prep, which might include AI in itself to help you fix the data. Um and they go anywhere from very basic, which would again be more appropriate for smaller operations, to super sophisticated, IT-centric, you know, tools that require five years of skill and training and learning to use. And that again is another continuum. Um, and again, it depends on what your problem is, where you're gonna land that. I would say that, like everything else, uh, you know, AI is not just a business of questioning your data to get a response. Uh, like I said, the predictive, prescriptive, or the descriptive, tell me what did happen. In pyramid, for example, is AI in the data preparation probably just as much as there is in the data visualization business, which is to help you work out how to clean it. And I think this is uh on its own a whole ecosystem. But you're probably, if your data's a mess, you're gonna need something in this world, pyramid or not pyramid. And and from our perspective, again, uh pyramid side of the house, we have customers who use pyramids. Some of them use other tools, they use combinations of them, depending on what their use cases are, who they're what their skill sets are and where they want to go with it. There's a whole gambit of different tools that they can use.
SPEAKER_00For companies who are interested in a pursuit of this, what would be some of the like typically what are the the departments or the areas where they should expect to get started and get kind of the the quickest wins or the highest impact?
SPEAKER_01Um okay, it's a great question. I would I would again thinking on the let's say the the lowest common denominators, right? Because obviously there are many many places to go here. The first natural space for business intelligence is on financial data. I mean accounting financial data. Um it's obvious, it's it's very data-centric. By the way, accounting data tends to be very well organized, it tends to be very clean. Makes sense. Um and you can't believe the difference between an accounting platform that has a great BI solution at the end of it and one that doesn't. One allows you to play with the numbers, and wow, I had no idea I was spending so much money on X, Y, Z versus one that doesn't have it. So this is an obvious place to go. And it's probably the number one use case, if I were to broadly describe the entire ecosystem. Number one. The second one, um I this would I would fork you into two different directions. If you are a uh a company that's very into digital marketing, the second use case I would go after is digital marketing data. Again, readily available. Um, you're collecting gobs of it, you know, through your your Google AdWords and your Facebook this and your LinkedIn that. Again, lots of good data. The structures behind it are pretty good. Some of it is a little complicated because it's not stored in a necessary analytically useful fashion, but still lots of things to glean from it. And if you are in the digital marketing business, um you can either cut your costs dramatically or improve the quality of those marketing initiatives if you can analyze what's working, what's not, and it's a perfect use case for it as well. Remember, the data is readily available, that's the big one. Sure. You know, you're gonna go to your Google manager and download the data and poof off you don't. And I'd say the third one, and this would be highly uh it's related to your business. If you've got your operational data, so your sales information, your your POS data, your transactions, if you're running uh a real digitized workplace, you should be able to extract that out, and that's another very obvious place to go. So these are the three obvious ones. The bigger the company, the more use cases like HR, um, you know, investment, organizational data, you you name it. There's no way into the stuff. Um, but those are the three I would go after in day one. Very, very sort of baseline. Everyone's got it. We should all be able to do it kind of stuff.
SPEAKER_00So I don't I don't know if you're able to quantify this question, but let's say two comparable companies, equal on most levels. One has a very robust uh business intelligence initiative in place. The other is using kind of the the baked-in reporting from each of the tools they use in the departments. Do you have data that would say like companies that are doing it well versus those who are just kind of it's an afterthought, the impact on revenue, margin, sales, turnover, growth, anything like that?
SPEAKER_01First of all, it's it's a phenomenal question because everybody's asking it. The bad thing is it isn't a hardened variate. It's one of those things that for whatever reason people don't like to share. No one wants to tell you, you know, how much thinking about it. Because ultimately it's the competitive advantage in the in the digital world we're living in today, it's it's it. What I can give you is anecdotal examples on reverse into this. So I'll give you an example. Um, and this goes back quite some time, but 20 years ago, Walmart um would tell you, and there are plenty of business cases to as evidence of this, you know, from Harvard Business School and you name it. Walmart would tell you that they their greatest competitive advantage predating uh Amazon was in their um IT and technological infrastructure in running the operation. That's what they would tell you. Now, obviously, it covers a lot of things from the way they digitize their warehouses and their logistics and the supply chain and blah, blah, blah, blah, blah. But all of that is predicated on data. And Walmart will tell you, if you look at those old business cases, that they were able to kill off the Kmart of this world by being technologically superior, which ultimately is data. So that's that's an example that I think it would be, you know, sort of black and white. Of course, leap forward to today, and you know, I think that kind of uh competitive advantage has disappeared because everybody understands I've got to be in the data business. I mean, it's it's it's unheard of in in a first-world economy where you don't have a data-centric organizational approach that everybody has to have. Now, if you're smaller, um you know, you're running um, I don't know, the hairdresser on, you know, the corner of fifth and whatever, um, you might not feel that because your business is not really related to volume, and you're not trying to work out the details behind it. But um, if you're trying to work out are my digital ads working on Google, or am I paying too much for my Facebook um ads and so on and so forth, then yeah, there's something to be said by actually sitting down and scrutinizing that information and seeing if you can do something better. So, very, very small companies that are very localized and their business is local, I think it's harder to make that leap. Suddenly you get bigger and bigger and bigger. Um, it's unequivocal you need it. And that's and the ROI is pretty obvious. So, for example, Pyramid is focused on what we call mission critical or operational use cases, which means we don't really we don't really come to the market with a, hey, let's build you a very cool set of pie charts so you can go home and show your you know your significant other look, honey, I built these beautiful pie charts and it's all clack and sticking on the fridge. No one cares about that. That that was something that happened 10 years ago when this was the cool idea. Today we're like, no, no, no, we actually need to show you something profound in your data so you can learn some of that and change the way you you run your business. Um, or we kind of we don't have a reason to exist. We can't we can't really convince you to buy something like Pyramid. There's still products in the market that do focus on what we call simple visualizations, but the world we live in today, people are looking for real, real insights. So if your business is somewhat opaque to you, which means it's complicated, there's some difficulty in understanding what's going on, you need to unravel it to pull the ball of wool apart to understand what is really going on, which is hard to say where that could or couldn't be, you're going to need tools to do it. That's what we do. And if it's data-centric, that's exactly what we do. And suddenly a piece of software like Pyramid would make a difference. Uh, admittedly, it gets it's more relevant the bigger you are, because the ball of wool is five times bigger and exponentially bigger than the next step below you. It's harder to unravel it. Um, and the impact could be bigger. But I would say most organizations today, even smaller ones, have that problem. Um, it's just a question of whether, as a business owner or a business manager, how small their problem is and if you can untangle it yourself in a spreadsheet with your own mind without having to do something exotic. If you need exotic, it doesn't cut it, you need something stronger.
SPEAKER_00So I I want to end on this topic because it you you mentioned something that I hadn't really heard before, and that's generative BI. Can you explain to me and to the listeners like what that is exactly and the benefits of it?
SPEAKER_01So the generative BI concept came about by the following Generative AI Um uh is is basically based on AI of taking large language models and then blending it into technologies, right? From the basic thing that everyone's favored, the chat GPT thing again, and then blending it into actual applications. And in in the context of business intelligence, we have taken the marketing direction, you need to take generative AI functionality and blend it with BI, and that's where we get generative BI from. Having said that, to put a little bit more sort of meat on that bone, uh the grand concept in the pyramid context is to be able to talk to your data. So pyramid offers an ability where you can literally, using a microphone on your computer, talk to the machine and say, hey, machine, you know, explain to me why my sales are going up. Or could you show me my sales for the last 12 months and project where they're going to go for the next six months? Literally by that very simple question. As I said, we've discussed already, take that question, generate the recipe, go back to your data, work out what's going on, come up with the machine learning forecast where they present it back to you. And we call that generative BI, which is again a blending of the generative AI magic, which is profoundly useful and quite amazing, all things that are done with the power of business intelligence. And that's what it is in a nutshell. And as I said to you earlier, if you deconstruct the entire picture, it's the idea of taking very complex activities and simplifying it down to uh a user, a non-technical user just being able to talk to their data, as we say, have a conversation with the data, just ask a question and then get a response. And by the way, then following it up, it's like, why did the numbers go up? Or why did they go down? Well, who's driving my sales in September? What's the reason that they dropped from from from uh you know from August to September? What was the reason for that? Being able to ask questions like that in plain English that are very what we call train of thought as a user, like just telling what's going on, and having the LLM and an engine like Pyramid go in and dredge out the answers and then respond back to you in plain language, it doesn't have to be English, by the way, but let's call it plain English. This is profound, and that's entire construct is what we call generative PI. Yeah. If you could imagine that working and it was accurate, useful, quick, easy, why everybody could do it. And suddenly we're going to make access to that clever stuff, you know, 10 times simpler than ever before. That would be profound if you could deploy that in your organization. If you're two people, not so useful, but if you're a 50-person company, that means 40 people can now play with it rather than one person sitting at their desk in Excel. That changes the entire nature of the operation.
SPEAKER_00Or out walking the dog, interacting with the app, or on the drive to work. Hey, what if we did this? What if we did that? That's incredible.
SPEAKER_01Yeah, I'll tell you a funny story as we end this. I've been on a flight flying transatlantic, sitting there with my iPad connected to the internet, which is great. Yeah and me asking a question through the chat box, which is great, but no one wants to sit there and and type into an iPad, which is very difficult to do. So very quietly, I'm talking there. I'm hoping not to bother everybody. It's like, you know, um, could you explain to me, you know, um why our Twitter feed is um you know trending up with? See, but it comes up with fan fantastic. And the question is going digitally from the plane all over the internet to our server sitting on you know the East Coast, uh uh querying our gigantic Twitter feed coming into pyramid. We ourselves are obviously using for our own needs, it doing the analysis and then coming all the way back up um through the internet onto the plane and giving me an answer on the tablet. Yeah. And the only thing I forgot to do is it it actually talks back to you. I had to turn the sound off because I was bothering people. Um and it's unbelievable the fact that I can do that in what was maybe like 20, 30 seconds. Um it might sound like a lot, but the fact that it was doing all of that work while I'm flying over the ocean is incredible. And that is the beginning of a whole new era of how we interact with computers, data, and software. Um, obviously that's BI, but you can imagine that extending to everything you do in your life, it's unbelievable. This is you know, definitely the beginning of the next era in technology, no doubt about it.
SPEAKER_00You know, and maybe let's just one last question on what do you see happening in 2025 for all of us when it comes to generative AI, not exclusively to BI.
SPEAKER_01So I think um there are a few things. The first problem at the moment is those models are quite expensive to run at scale. They're then it's not cheap. I mean, there's an argument that if they're answering important operational questions for you and they're doing it very effectively, um, it's actually very effective. If they're if it's just a toy, it's actually quite expensive to use. They're like, you know, oh, I just did this and it's junk, it's a waste of money. So there's there's that headache. So I think that the next um small step, these are the small incremental steps, is to find a way, the models for them to be uh cheaper to run, which means they have to shrink, they have to be more effective, they have to be a little bit faster, maybe a lot faster. And um you're definitely gonna see that the the competition is heating up, and you compare Google Gemini Pro to GPT-4 from OpenAI versus uh Mistral um and so on and so forth. And then you've got clawed all the way on the other end, which is very powerful, but also really expensive. So there's a market movements that are going to happen, super competitive. That's part one. Part two is uh more the longer term, is the entire concept of um doing more mathematical deterministic operations in something similar, analogous to an LLM. It's gonna be a different animal, but analogous to an LLM. In some respects, that would cannibalize the ideas that I've expressed to you in the GMBI space, because now the the mathematics can happen there, you still have the headache of moving the data. And I don't think that's a problem, it's not gonna go away anytime soon. So it's debatable what will happen there. That's longer, longer term. And I would say the last concept um is actually taking all of it and implementing it. Um, lots of people are talking about it, lots of people are excited about it, but a lot of people are still sitting um on the bench and trying to figure out how to implement it. I mean in specific applications. So Pyramid is already deployed, we work, it's in production, we've got customers who use it, and they effectively get to the LLM vicariously through the Pyramid application itself. Yeah. But you know, you're you're running uh, you know, a set of you know retail shops, and you want to put a chat bot on your website so people can ask questions around what you know whether to buy the blue skirt or the red skirt today. Who knows what the questions are? How do you do that? And I think that's the next other somewhat semi-incremental step at the moment is how to actually implement Genitive AI in my specific application today for me. I think that's uh also very, very much here and now. So that's that's how I would categorize it. Um and uh and and what lies coming on the pyramid side is we're really do some really sophisticated operations. I mean, crazy, crazy stuff. I can tell you now where we've already got prototypes of something even crazier for the future where we can resolve business formulations for you. Like you can also really complicated, I'll give you something crazy. Like, you know, can you do the can you do my option pricing for me on, you know, uh my Microsoft yada yada shares with this kind of duration, blah, blah, blah. Which means I'll go and work at the Blackened Scholes formula, read the data from here, add this, divide by that, multiply this crazy, crazy complex logic. Not just asking questions, but actually solving mathematical operations without the math engine. That's what we're working on at the moment, and I think that will be the next leap forward for what we do today on our side. And all of those things are moving in tandem. You'd have to have it if you're if you're gonna move to the next level. We're riding a rocket, man.
SPEAKER_00It's crazy.
SPEAKER_01Crazy stuff, crazy stuff.
SPEAKER_00So, first identify who listening to this would be an ideal client for Pyramid.
SPEAKER_01So, like I said, we tend to go on the bigger side of things. It doesn't mean if you're a small company, you're not relevant to us. Um you're definitely going to have to have some amount of data. Uh, it can't be so small that it would be overkilled. Yeah. This wouldn't work. You're gonna have to have a a decent amount of data to come and talk to us, otherwise, like it's it's overkill. And something where you think you've got something to learn from your data in the first place. If you if you don't have a good answer to that, come and talk to us. We're always happy to talk to everybody. I think it'd be less relevant. That's the honest truth. Um, and I would even recommend for smaller, simpler, cheaper solutions in the market, even that's a lot of work just for the sake of getting to the cool stuff, it's overkill. But if you've got a decent amount of data um and some really complex business use cases you want to solve for, then we have got something to talk about. And obviously, the bigger you are, the more relevant the entire platform is. But like I said, that that definition is is uh is a bit of a sliding scale, and it depends where you are in terms of your needs.
SPEAKER_00So somebody listening to this, they say, hey, that that sounds like us. What would be their next steps? Uh obviously, folks do your due diligence on your own internal stuff and all that. But based on what I've heard, I would approach Avi's team at Pyramid and say, what does that look like? So how do they get in touch? What does that next step look like for them?
SPEAKER_01So specifically with Pyramid, um, you come to our website, we have a contact form there. Uh, give us your details, uh, we'll contact you and we can set up a conversation. Um, obviously, our sales team will be thrilled to talk to you um as they would be to everybody. Um and you know, we're we're we're gonna have a real conversation with you. Like we're we're very reluctant to um to sell ice to the Eskimos. It's it's not a great idea. Um, we're gonna tell you what works, what doesn't. We can show you demonstrations. We can even use your data in a quick and dirty demonstration so you can actually see it working live, real you can see there's a lot of technicians. Um, our site, pyramidanalytics.com, nothing complicated. You should see a sort of a link at the top there to contact us. And um, yeah, and come and take a look. Um we've got a bunch of videos um on YouTube, we've got a bunch of videos on our site. Um, if you're in LinkedIn, and uh if you find me, you'll find me in LinkedIn. You should see a bunch of postings I've made over the last seven several months showing some of the stuff. It's all real, by the way. Very cool. This is not the fictitious marketing department making up very slick videos. Uh, it's actually you know, the the technology guy actually doing it. Um and um, you know, it's it's the real deal. And we can take a look and see if it you know piques your interest. You have a com if you want to have that conversation, happy to have it, and we can go from there.
SPEAKER_00Wonderful. So, for those of you listening, all this information will be in the show notes, as well as the diagram that Avi referenced earlier, the horseshoe diagram when he was using to explain Pyramid's process. Um, Avi, awesome. This again, this isn't a topic that we've uh really addressed on the podcast before, and huge oversight on our part. This was very enlightening and um just really gives me a lot more questions to ask.
SPEAKER_01So we can always have a part two. When everyone has a coffee drink, and we can have a we can have a part two at some point in the future. I'll be happy to do that. Thank you again, Chris. It was it was a great uh session. I hope I hope uh everyone viewing this uh has learned something and uh we'd love to hear from you. Awesome.
SPEAKER_00Thanks everybody. We'll see you on the next episode. Thanks for tuning in to using AI at work. Don't forget to subscribe for more conversations about how to use AI at work. And a special thank you to our sponsor, Chief AI Officer, for empowering businesses with AI education and training. Visit their website for a free AI readiness assessment and AI strategy guide to help you get started using AI at work. That's www.chiefaiofficer.com. Well, thanks to our producer, Evan Desolnier, for making this episode possible. Follow us on Twitter at the handle usingAI at work and visit www.usingai at work.com for free resources to help you harness AI in your role.