Mark Lerner:
All right everybody. Welcome to this episode of the Revamp podcast. Very special guest today and Clemen who I’ve had the opportunity now to podcast with before we got to meet in person this year at HubSpot Inbound. He was dressed for the occasion and I’m really excited to chat with him about AI’s place in the Rev tech stack in the coming year. Before we jump into all of that stuff, why don’t you go ahead and tell the folks at home a little bit about yourself and your background professionally and how you got interested in this kind of stuff.
Klemen Hrvat:
Definitely. Thanks Mark for hosting me. I’m really excited. The previous podcast hosting with you was fun, so I’m looking forward to it. Long story short, I’m a pharmacist by training, so I’m not a developer, but I’ve been, as a founder , involved in sales and business development for the past 13 years. Every new company is further from pharmacy and more into sales. I’m not sure why, but that’s how it goes. So yeah, now I’m a co-founder at Sellestial where we’re building an AI supported system for writing personalized messages for HubSpot users.
So I’m learning, hearing, thinking a lot about AI from various aspects. I’m not a technical guy, but I do think a lot about how AI can help marketing and sales for sure.
Mark Lerner:
Yeah, and I think you and I are similar in some ways because we don’t have a technical background, but certainly over the last two years since the first iteration of chat GPT became available, that’s changed. I have become technical. I’ve been writing code or at least working with an AI assistant to write code and really kind of stretching myself to go beyond what’s expected of someone in a quite non-technical role. But I’m interested, obviously the jump from being a pharmacist to being a founder of an AI sales tech company is quite a jump. How did you get into tech generally and then kind of the AI world?
Klemen Hrvat:
So the first step was after I realized I used a really good use of labor at my first startup where it happened to pivot from being a tech company, a bioinformatics software development company, into being a data science for pharma companies. I was the note taker before AI note taker was a thing on every call. So I realized, okay, I’m not scientifically skilled enough to really help there, but at that company I developed with my co-founder a system for lead generation into pharma. So we realized, okay, why don’t we build a company that is helping do what we’ve been doing there for other vendors and we decided on purpose to build it in a way that it’s really product focused more than a service. So from there, for five years we’ve been building a lead-gen agency. It was really, we developed our own system for sending out emails.
We developed our own scrappers for data, for updating, for managing responses. We now have a small team of a few people serving 80 clients with ease because we developed a lot of systems. So I was always involved in that tech development thinking about applications and throughout the whole career of being a legion agency, we’ve been always thinking what we can do in a way that can be scalable. So of course with AI and everything, my other two co-founders are computer guys. They’re the guys with hands dirty in ai, but we’ve been always really thinking how and what to do. And from talking to so many salespeople and sales leaders, it always felt okay, there is all that follow up game and there’s a lot of technical skills that salespeople don’t have. So more user-friendly tools to help do stuff at scale that we’ve been already doing as an agency could help.
Initially we started developing an app for just one communication automation, realize, no, it’s not going to fly for a lot of various reasons. We learned last year and now we are developing a system where, funny enough, our hunch is that salespeople will need to be really good at prompt engineering, not developers, not necessarily developing their own apps or systems, but prompt engineering is what will help you really leverage AI in the coming years. Funny enough, just one day in the morning when I was listening to one of other podcasts talking about the new deep seek model, the host said it might be more feasible than anytime before now to fine tune your own model. On the other side, a lot of experts are saying it’s not about fine tuning, it’s about better prompt engineering. Every model that is better is more or less biased, and with fine tuning, you’re just giving more bias in a way to the system where if you do prompt engineering, you can have the power of a much better model and still navigate it in a way that gives you the results you need.
Mark Lerner:
Not to get too much into the technical weeds. I did early on attempt to fine tune before I even really knew what that meant. So I think when I had access to GPT3’s API, it just came out absolutely gibberish afterwards. So I definitely did it wrong, but I think what would be interesting is that we can talk a little bit about the areas in which there’s opportunity for AI to really change things in the rev ops world. I think the use case of tech generation in a large language model, I think that there’s a lot people, it’s already here, everybody kind of knows that and there are different ways of doing it, and then it’s kind of the stuff on the back end that makes it unique is really kind of what sets it apart, but some of the areas that are ripe for disruption.
Before we got on, we were talking about one which is kind of data quality and there’s a lot of opportunity there. So one of the things that you had brought up was this idea of using large language models for a dedupe problem. So I think everybody out there their entire careers, if you’ve ever dealt with a database of contact CRM, the duplicate problem is a forever problem. It’s always been there, especially before AI. You could write as complex rules if you want, but there are always going to be edge cases where you’re just not going to get all of ’em. But with AI, it’s increasingly possible to potentially get almost a hundred percent of the problem kind of automated away. Yeah, I mean, how do you think that problem can be approached with ai?
Klemen Hrvat:
Yeah, definitely a good question. A good case where I see a huge value for ai, especially as I said now, if it’s clear, it’s first name, last name, company name is matching. Obviously now these are two concepts that should be merged, but it’s not often that obvious where you might have incomplete data to some extent, and then with some clever mapping and AI thinking you can do that. Another part of data quality and in a way DUP is when you’re trying to update your CRM, you most likely don’t have the information when a specific data was entered and from what source, which means when you bring a data, which is somewhat different to what you have in the CRM, which one is correct, is it LinkedIn job title? Correct, or the job title in your CRM, is it outdated? Should you rely and someone just didn’t update the LinkedIn profile, how to go about it, how to go about all the cases. If you want to upfront define all that, I mean good luck to cover all the edge cases, but with AI there is a good opportunity to just approach that more holistic, more broadly if you will, and let AI combine all the thinking of possible scenarios and case by case, decide what should be the data you trust, and then whenever you get another point, you again just reconsider everything and then through that you can clean up your CRM to be AI ready from the data perspective.
Mark Lerner:
Yeah. One of the things that is very apparent to me, at least in the current iteration of AI, is that databases are crucial. And I know that sounds silly, but I think they’re even more important now because especially vector databases for sure, but it’s like you were talking about you have a CRM database, it may not necessarily have timestamp of different changes for every single field or for which one is the more recent or whatever. I think this may have started with having people using things like a warehouse or data lake as kind of the place where a lot of these messy things were worked out. But when I talk to people working on AI solutions, almost always what they’re doing is taking your original source, bringing it into a kind of different database or data model and doing stuff there and then passing it back because the AI needs a very specific or the right kind of model to parse the data and then it can send it back so it’s not really messing with whatever you have set up. Am I kind of seeing that correctly? Do you agree?
Klemen Hrvat:
Yeah, yeah, I would definitely agree. Yeah, you’ll have a single social through data and you won’t have that, and then you’re bringing other data sources to reach to do the magic with ai, whatever you do for different purposes. And then you keep the data clean at the source and you want to have the structured approach and definition of how you say, okay, and to what data you trust, because if you can’t trust the source of the data, whatever you do with AI will be crappy, happy. Just a case that we’re now thinking about now within Celestial is we are about to have a system to help you super personalize every message. Let’s say for post webinar where we want for post webinar highlights to be matched to the persona. A C-level executive is maybe interested in the first 10 minutes of a webinar when A SDR might be interested in the last 20 minutes of a webinar.
So why would you blast highlights from the whole webinar to every persona? But if you want to match the highlights to the persona, you need to have the persona. So it won’t help you to have a really good system to match the persona with the highlights. If your source data is outdated, if the job title of a person is SDR, but that person has already had sales for the past three years, it’ll be just personalizing in the wrong direction. And then you might conclude, okay, AI is not working for me, but it’s not necessarily AI that isn’t working. It might be your source data that you should take a look at.
Mark Lerner:
Yeah, I mean this is a tangent or kind of beyond deduplication is the data enrichment. So up until now there were two or three, let’s say big players that had huge data sets that we all worked off of. There was ZoomInfo and Apollo, or there was ZoomInfo and Apollo and those are the big ones. And then a few, there’s a bunch of other ones depending on where you are in the world and especially ZoomInfo for time on the market, they were the way to go. I think that with AI agents and large language models, we may not necessarily be bound to have to pay considerable contracts to these people to do that data enrichment. I know that we’re already doing some stuff on a small scale to solve some of those issues. What’s their current job role? Has it changed in the last three months? Did they change jobs? Right. I do think that that’s a place that’s already seeing AI innovation for the Rev ops use case, but do you think, we talked about some newer models that are coming out that are less expensive that could potentially be run locally, reducing costs. Do you think that as those come online, those kinds of big analysis projects will become more commonplace within companies?
Klemen Hrvat:
I would say yes. If now a project would cost you $5,000 on oh one with a similar capability of deep seek R one model, you would pay a hundred bucks or I’m making numbers up, but the range, but
Mark Lerner:
We’re talking about orders of magnitude.
Klemen Hrvat:
Yeah, the range is there. It significantly dropped. So yes, we’ll want to leverage that because you can do it now at scale before even marketing emails, why would you still want to send a templated marketing newsletter if you can personalize it at no cost now with models that if with oh one model that is really good for consistent output and it obey what you say in a prompt, if the costs are 96% lower, why would you only stick to the sales part for personalization? If you have all the marketing emails that can be written fully individualized for every recipient, now you can afford it, you wouldn’t be paying, and it’s not worth paying thousands of dollars for a marketing newsletter. But if we’re talking here about a hundred bucks for the whole database to send a really personalized message, I mean, you would probably start considering.
Mark Lerner:
Yeah, yeah. I mean for sure. And I think the thing that holds people back or holds companies back is there’s an inverse relationship between the size of the company and its embrace of some of the more cutting edge things, especially when it comes to letting the AI write stuff that isn’t done by a human and is going to be seen by the world. I think there’s been some, early on, there were some cases with companies that let AI loose, and there was some guy who interacted with an airline chatbot and was able to convince it to give him a hundred percent discount on flights or something, and they had to go buy it. And then there was a lawyer who had AI write his case and it was citing all these other cases that didn’t exist. Those were very big in the news like, oh, better watch out, better not let AI.
So I think that there are, especially a bigger company that has more to lose, is less willing to kind of let AI handle that though I think as guardrails come into place and it becomes just what’s expected that’s going to change. But the area that I think is really going to be a game changer for companies, for rev ops, for sales, for revenue in general is the predictive insights, predictive analytics, predictive everything kind of potential here. So it used to be that you’d have to run really complicated machine learning models that did all this kind of scoring and stuff, and you still can, but I think that with these smarter models that are less expensive to run, that could even be run locally and that could handle a huge amount of context to feed it your current ideal customer that you have and say, find matches like this and it may find behavior patterns that you or I wouldn’t see, but it does because AI and then being able to kind of leverage that for predictive stuff, I think there’s going to be a huge benefit there. I kind of want to see what you think about that.
Klemen Hrvat:
Yeah, I think you’re thinking in the right direction. I would even say I see probably sooner than we all think in the future where CRM or a system connecting to your CRM is just going through everything you have now. You have the scoring based on how actively they’re maybe responding, and then you’re trying to augment that to, okay, someone above 60 is a really good person to focus on, but if you’re thinking about A CRM, you not only have the emails you have, what PDFs they downloaded, what pages they viewed. If you combine that with their activities, maybe on LinkedIn, combine it with their press releases, you can just throw all that data into AI and say, okay, those are the companies I closed in the past five months. Go and find me. The others in my CRMI are not sure what the criteria is, but I know I successfully closed those 50.
Go figure it out. Go there. You have all the data. If anything is unclear, just ask me if you need other sources, if anything is ambiguous or too nuanced in a way that it might be hard to understand. Yes, we’ll add additional sources, but I definitely see a timing feature where you have a season that just goes through all that data and just analyzes it. And one idea might be you have a CSM that says, Hey, there’s a person who just messaged five days ago, and according to the behavior of your other customers, this is really high potential. But for now, we don’t have that information because it’s hard to gather, and as you said, it’s just too much data for your eye to gather or for any holistic approach where you implement some bias for triggering is not sufficient enough. But with a more broader look at the data, there are for sure patterns we don’t even mean.
Mark Lerner:
Yeah, sorry about that. I had a bit of a camera issue there, so my image here might look a little different for the folks at home, but I’ll take what you said and I’ll raise you not only that data, but also if it’s an existing customer, all the product data that you have and behavior within the product on an account basis as well as an individual basis, and then using that for upsell, churn signals, account expansion, things like that. It could be for anything. When we talk about within the sales cycle itself, giving sales reps an edge of that, whatever signals about how to approach the sale, things like that. I mean, those are all areas where this kind of insight that humans might miss because we’re not as good at finding patterns as ai, but that would give an edge. So yeah, I think those are going to be big.
Klemen Hrvat:
Yeah, yeah, definitely. It’s not only for marketing and sales, but even for customer success teams and of course for upsell to the sales team.
Mark Lerner:
So I think the one area that we didn’t cover, which is kind of near and dear to my heart that I haven’t seen in the folks that I’ve talked to, I haven’t quite seen this change happen as much as I feel like it’s happening within my purview, but is just the incredible ability to bypass either needing to buy an external tool or require additional tech resources internally to build. So if you have, it used to be that when budgets were as big as you wanted them, everybody got their own little special SaaS for their very small sliver of a use case. Then we had interest rates skyrocket, budgets got slashed and everything. Nobody had a budget for anything. And so there’s a lot of consolidation and people lost those little tools. But with AI, we can build those tools. I mean, the friction that would stop you from being able to do that is almost non-existent.
If we talked about prompt engineering, if you can prompt correctly and you have something like a cursor or a windsurf or there’s a whole bunch of ’em out there, you can kind of build internal tools in a way that was never possible before. I’m doing it and I’m not a technical person. So especially from a rev ops perspective, I think there’s a huge opportunity where folks in rev ops can really step up and take on a more traditionally technical role and be kind of the builder. That’s kind of the role rev ops is always meant to be is the builder. So
Klemen Hrvat:
Yeah, I would definitely agree. It’ll go in the direction. We’ll all be developing small tools for that. Even with, as you said, the expansion of all the SaaS, you always search for something and search for a long time, and the tool didn’t seem to be exactly for what you’re looking for, but then you try it and it might not necessarily work, and then you already spend probably the same amount of time that now you would spend just building that tool specifically for your own case. So yeah, I was just talking to our CTO who said no, he now started using Bold for web app.
Mark Lerner:
So Bolt is the other one that I couldn’t remember. Yeah, bolt is
Klemen Hrvat:
Another one. Yeah, he built the app, as he said. Now would take him before that tool, probably a month or two months within a few days. And it’s not without any mockups, without anything, it just works. Okay. It’s not connected yet, but to the backend. But the web app works. He brought with a few prompts, drag and drop and all the fancy UI components. Of course, it’s not a hundred percent there, but for a developer even, it gets you to 96% and you just tweak a few lines of code and then you have a functioning app up and running, and it’s not really just a text editor. It’s a complex app. And with that, you can really tailor the apps for as a rev ops expert, for marketing, for sales, for customer success, for financials, for reporting. I mean for now it still feels like we’re just exporting data from A CRM and then you go in Excel or through Python script, you do some reporting and data gathering and tweaking and enrichment, but now you just have an app connecting to your CRM and you have what you need.
Mark Lerner:
Yeah, I mean, I’ll give you an example of my first taste of this was back in 2022 or whenever it was when Chad GPT first became available, I think it was around Thanksgiving, the Thanksgiving holiday here in the us. And so I went on a ski trip and I spent most of the time in a room trying to figure out how to use this thing to make me better at my job. And what I came back with was this Airtable with an API and essentially was able to build this thing that you give it kind of a broad topic and it’ll build you an entire glossary for SEO, the strategy of keywords with glossary. And it just filled in all the blanks basically. And then connected to Webflow and sent it. And I presented this to the company, my company at the time, and I could hear the devs on the call go like, oh, shit, what this had enabled me to do today. I don’t need Airtable. I can build the front end myself with Bolt or VO dev and connect it to all those things and actually have a SaaS. And this again, is, I have almost no real technical background. So anyone that is kind of that mind the opportunities to build internal tools and really establish yourself as a critical piece of the company, as a builder, it is just endless right now. So that’s what gets me excited. That’s what gets me going in the morning.
Klemen Hrvat:
Yeah, I mean, sometimes I hear HubSpot admins or whatever, CM and RevUp experts might be the ones behind, they just get pulled into some marketing projects and some sales projects and some CS projects, and then the work they should be doing is just left behind. But the support should be now much faster for everyone because you can just build something much quicker. You can analyze the data, and every team as a result will be much more efficient because you sit down, you quickly tweak it. Just earlier today, we’ve been playing around with some APIs for some data enrichment, and if you can do that on your own with an agent, we’ve been searching for a tool to have really precise revenue information. You have an agent that goes to Google, synthesizes all the data right there for you, and it does a pretty good job, more or less for free.
Mark Lerner:
Yeah. Yeah. I mean, there’s so much, right? The opportunities are endless. And I think the only thing holding anyone back would be either fear or a lack of imagination. So I think the thing that I would say as we wrap up here is that I personally say you slice off like 5% of your time to identify bottlenecks and problems within the org and try to think about if technology or resources were not an issue, how would you solve that problem and go from there. And I’m guessing that off the bat you’ll be able to find a few things where you become the hero. So that’s my suggestion to the folks at home.
Klemen Hrvat:
Yeah, I would only add that. Yeah, even if it seems unreasonable to address that with AI today, the next model might just solve it. So think what’s coming in three months, in six months, understand the problems inside out, understand the processes so that once the technology, the models are there, you can just go and automate or solve those bottlenecks because who knew a year ago what models will be capable Today, we probably all anticipated it’ll happen now in five years, but it happened in a year. So who knows what’s going to bring the next five, six months.
Mark Lerner:
I’m very excited. So we will leave it at that. But before we go, if you want tell the folks at home maybe where they can read a little bit more about some of the stuff you’re doing, maybe learn a little bit more about Celestial,
Klemen Hrvat:
LinkedIn is probably the place to go. We have some basic information on the website, but I’m super active on LinkedIn, so you can find me there. I’m sharing a lot of tips and tricks around personalization using AI to help you with the sales process, write more and better messages around sequences for HubSpot. So feel free to reach out there. I’m always happy to chat.
Mark Lerner:
All right. And we’ll put a link in the show notes to your LinkedIn. So thank you so much for joining us today and I’m really looking forward to our next conversation in a few months when oh seven comes out and we can maybe talk about that. So thanks again.
Klemen Hrvat:
Definitely. Thanks for having me. Mark.