Building an AI-Native Startup | GrowthX's Marcel Santilli

Marcel Santilli 0:00 And what I started to realize was most people are really shitty at delegation. That's just the reality. They're horrible at sharing context. They're horrible at communicating their thoughts. They don't write or they don't write very well. So they don't organize their thoughts very well. They don't know how to bring people along. And so if you're pretty bad at delegating something to another human, you're probably going to be even worse at delegating it to AI.

Conor Bronsdon 0:31 Welcome back to Chain of Thought, everyone. If you are watching on YouTube, you have already figured out we are not in our homes recording today. We are in fact live in San Francisco. I'm delighted to be at the offices of GrowthX with their founder and CEO, Marcel Santilli. You may also know Marcel as the former CMO of Deepgram and the former CMO of Scale AI, which, I mean, if you pay attention to news at all, was just purchased by Meta, well, 49% of it was, for about $15,000,000,000.

Conor Bronsdon 1:01 So he knows a few things about scaling and building AI businesses. Not to mention, GrowthX is having incredible success. In just the first six months, they've already raised $12,000,000 from Moderna Ventures, and they've actually already scaled to 7,000,000 in revenue run rate in just six months, an incredible amount of growth. They have a unique perspective on how they're building their business with customers like Reddit,

Conor Bronsdon 1:29 Ramp, Webflow, Abnormal Security, and Galileo. You're fresh off a $12,000,000 raise. You are having so much success here.

Marcel Santilli 1:39 Welcome to show, Marcela. It's great to see you. Yeah. Thanks. Awesome having you. This is our new headquarters,

Conor Bronsdon 1:44 I guess. You know? So it's always awesome to to hang out. Yeah. Yeah. It it's a ton of fun. We're really glad you had us down here, and always great to see you and catch up and hear about some of the incredible ideas that are percolating for you. Because you obviously have years of success, years of experience building businesses within the AI space, building businesses in general.

Conor Bronsdon 2:04 I'd love to kind of tell our audience, what does it take to build an AI native business? Yeah. I

Marcel Santilli 2:11 guess maybe I can go back a little bit and give a little background, right, and kind of how, what has informed my thinking. But pretty early on, I was at HashiCorp, we scaled that business open source. And so from I saw community building from the ground up and seeing the scalability of that kind of fast forward a little bit when I was Scale AI, obviously, I got this

Marcel Santilli 2:35 awesome behind the scenes view into all these AI labs before things kind of started to blow up Totally. AI and kind of seeing what data they were using and whatnot. And, and scale is just a very unique business for, for different, different reasons. Won't get into, into that. But, and, but then at, at Deepgram, I was at this very unique position as well because there's, at the time, at least there's not that many companies that were,

Marcel Santilli 3:00 you know, had their own data centers, they were training their own models, they had their own in house data labeling operations, their own research teams, and serving their own models, having models as a service via an API. So it's a very, very unique business as well. And we built a startup program as part of that. And so I was, my team was running that startup program

Marcel Santilli 3:23 and I just kind of got this view into things. And so from my years at scale and then Deepgram seeing all of that, one of the things I started to realize was there's this kind of big disconnect between, it felt like all the money, all the attention was going into these frontier models and training them more data, more compute and everything else. But from my experience doing conferences where we're interviewing a lot of these researchers and whatnot, you realize some of them actually never had a job. And what I mean by that is they never actually had to deliver on a revenue target or built a actual product. You know, these are researchers. These are amazing,

Marcel Santilli 4:00 super smart, smarter than you can imagine people, but they haven't actually apply and done the actual job. And so, you know, as, as a operator, you know, over the years, I think what, makes good operators is you having empathy for what does it take to do a good job, right? And what does it take to run a good company as well? And so a lot, it just felt like a lot of what was happening in AI was just building models, right? And so disconnected from actually applying that to, to real businesses.

Marcel Santilli 4:32 And here I am, I saw all this context. And then at Deepgram, it felt like it was closer to, you know, applying it to, to increase productivities of companies and help companies grow more efficiently. But then I think where things started to connect for me was when I just started building myself, you know, so I said, all right, let me just start stitching together. How do I go about doing great work?

Marcel Santilli 4:55 And I like to call it kind of the messy middle. How do I go about researching my audience? And I go and I research and ask these questions. What is the mental model that I have when I'm asking these questions? What are the sources of place or places I go for information? And then how do you process that information? How do you plan your work? How do you go execute against that plan? How do you iterate on, on that execution

Marcel Santilli 5:20 and do a bunch of other things? And so what I started to realize was, as I was doing that, I was playing with the inputs. I was, you know, prompting it differently. I was, you know, adding more steps, changing how we were doing the plans and whatnot. And there was kind of this moment where I realized, well, you get to a point where actually what made me good at what I did could be pretty reproducible.

Marcel Santilli 5:44 Right. But it requires still a lot of work to apply these models in a way that actually can, can help you reproduce these plans. Right. And, and so for me, was the biggest disconnect was you see all these frontier models, they're gray, but it just felt like you're doing so much work to get them to be applied to more complex knowledge work. Right. And then here was seeing,

Marcel Santilli 6:09 like after doing a lot of this work, how it could be applied. And, and so that, you know, long story short, of led to GrowthX where at first I was just teaching people how to build these workflows and whatnot. I quickly realized there's a different way you can build companies, you know, and we can get, I'm sure we'll, we'll get into some of that. And, and so when I think about AI native,

Marcel Santilli 6:33 the way I kind of think about it is completely different. It's, to me, there's two principles and two skills, if you will, that are the most critical ones in, in, in how in AI native businesses. The first one is first principles thinking. And so how do you decompose things down to, to the very core foundation, the essence, and then rebuild it, but rebuild it in a different way. And then the second one is delegation.

Marcel Santilli 7:00 And what I started to realize was most people are really shitty at delegation. That's just the reality. They're horrible at sharing context. They're horrible at communicating their thoughts. They don't write, or they don't write very well. So, don't organize their thoughts very well. They don't know how to bring people along. And so, if you're pretty bad at delegating something to another human,

Marcel Santilli 7:22 you're probably going to be even worse at delegating it to AI. Right? And so if you don't have really good first principles thinking, and you don't know how to delegate, that is literally how to build an AI data business. You, if you're good at those two things, right? There is the infrastructure, there is the stitching everything together, but it's just completely rethinking every function in the company, you know, and we can go into examples how we're doing recruiting differently, how we're doing every function in the company

Marcel Santilli 7:49 differently,

Conor Bronsdon 7:50 But it comes down to those two, two things, I think. There are so many great threads to pull on from And what you just I love that. I'm like, okay, do we have more time? I, first of all, I love this insight of, look, it's great to build a great model. That's fantastic. It has a lot of value in the moment, but there's always a new frontier model coming. You're always chasing the dragon.

Conor Bronsdon 8:13 Whereas if you build a great business, you may be able to have more durability, whether or not you are always using your own frontier model. There are other models you can leverage and you can build a great workflow. You can actually understand your customers in-depth and build a great business around that. And that seems like that's the approach that you're taking here at GrowthX today.

Conor Bronsdon 8:33 The other interesting thing you bring up here, which I think maybe relates to this point, is this idea of like, look, we have brilliant people who are enabling the AI revolution. We have brilliant researchers. Some of them have come on the show. Some are coming on the show soon. But, they may not always be the best folks to build these AI businesses. They are crucial to doing so, but you also need great delegators. You need people who are incredible people managers

Conor Bronsdon 8:59 who can then apply those skills to, Hey, how do I manage all these different AI workflows, AI agents? How are you approaching recruiting and scaling GrowthX differently than you would in another business? Yeah. So, I think we're set up quite differently. So, started with services and the main reason we started with services is because

Marcel Santilli 9:19 services are a forcing function to figure out what does it take to do a great job And what does it take to deliver high quality work outputs at the end of the day? Right? So, someone is actually paying you for a work output, not for the tool that maybe you can do the work, but actual the output of that work, and hopefully that that is informed by good strategy,

Marcel Santilli 9:42 then that is what is actually valuable to businesses. Right? And so the process of doing that, what do you do? You hire a service provider, an agency, a consultant to help deliver that work output. That person normally, or company or agency, they're normally the experts, right? And so for us, the reason we started with services is because we want to understand what is that messy middle to deliver great work products.

Marcel Santilli 10:12 And so, just to give a very concrete example. So let's say you go to a Michelin star restaurant and you have this Michelin star chef delivers this amazing dish right in front of you. And you just look at it and it was like, this is amazing. And you take all the videos necessary, all the pictures and everything, and even come up with the perfect ingredient list and it's all right there. And then you bring someone and you say, here's all the pictures of the final product. And here's all the ingredients.

Marcel Santilli 10:39 Help me replicate this. You would not be able to do that. Right? So why is that? Like, it's not about the quality of the ingredients. It's not even about how much detail you have in the final output. The final output is a snapshot of something. Right? And, and so in order to replicate great, you need the messy middle, you need the try on errors, and there's not only one way to achieve greatness.

Marcel Santilli 11:05 And so knowledge work is that. And I think what, what we started to realize was just like how important that messy middle is to, and there's no easy, non messy way to capture that. You're not going to build this UI that some designer and product team is, you know, off in the corner designing and building for six months, then they launch and then it goes out and all of a sudden, like a new frontier model or something, some paradigm changes. And it's like, you're six months behind, right? Already. And so instead,

Marcel Santilli 11:38 when you think about like, forget the UI, forget the form factor by which you're delivering it, just sign up to deliver the thing. And then figure out what does it take to deliver in a more AI native way, but not in a way that's like, I want to replace humans or I only want to use AI or I don't want to use AI. It's more about like, doesn't matter to us. It doesn't, it really doesn't matter. The way we think about it is like, where are the places that you absolutely need expert interventions?

Marcel Santilli 12:04 Where are the places where you need to shape how to do the work? And, and then what are the places where you do need the humans in order to get higher and better quality inputs from your customers, you know, in order to deliver better work products. And over time that, that mix is going to change. And I think, you know, we have a lot in the works, but I think there's, the form, the current form factor, at least for knowledge work

Marcel Santilli 12:29 is, is not there. And I think there, there's, there needs to be a paradigm shift. If you think about what's happening with coding agents, you know, Cursor, Windsurf, AugmentCode, and, and a lot of these companies like they're, you know, in the, in, in just like the day to day of our engineers and how critical these tools are. And they're truly like professional tools that are making engineers and builders more productive. And then you compare what's happening to knowledge work of copying and pasting context over to ChatDPT or Claude and, you know, and you just kind of like, it feels like you're working for a chat window. Like, it's like, I work for this chat window now. Like, and I'm constantly trying to like wrangle and wrestle it to do the thing I needed, and I have to constantly tell her the same thing. It's like, my wife like, Did you not hear what I just told you yesterday? It's like, it's like, I feel like that's me, you know? So I have more empathy for everyone who says that to me, you know? I'm joking, but Oh, man. And and I think there there needs to be kind of some kind of paradigm shift there.

Marcel Santilli 13:32 Well, first of all, I'll say,

Conor Bronsdon 13:34 listeners, you heard it here, but when I, in the future, steal this great comparison to a Michelin star restaurant in order to describe how knowledge work and the messy middle works in the future, just pretend you didn't hear Marcel say it first because I'm gonna I'm gonna borrow that for sure. I'll I'll start there. But secondly, I, I just really love that you are taking us back to, as you said earlier, first principles thinking.

Conor Bronsdon 14:01 And in this case, yes, you're using services and deep engagements to understand and then build products that can help solve these problems. But more than that, you're taking on first principles business thinking of saying, we're going to be customer obsessed. We're going to understand our customer. We're going to get close to the bare metal. We're to understand what it takes to ship for them. And we've been doing that a bit with Galileo where we,

Conor Bronsdon 14:25 you know, you may know us as an observability and evaluations platform. And you're probably hearing us use the phrase reliability a lot more. A lot of us are starting to use this phrase reliability because we've realized that observability evaluations, while crucial, are means to what we're actually solving for customers, which is that reliability challenge.

Conor Bronsdon 14:45 And I see you doing exactly that for knowledge work and exactly that for growth teams saying, look, we're going to come in and figure out like what your problem is. We're going to actually solve it for you. And then we're going to break that down and say, how can we solve it better? And that is such a first principles approach as you put it. And it clearly opens up a

Conor Bronsdon 15:08 bevy of pathways for you in the future. What do you see as kind of the next stage of GrowthX's development as you continue to dive deeper into AI enabled workflows, AI products and much more?

Marcel Santilli 15:20 So, I'll, I can share a little bit of kind of how we're thinking. And I think a lot of the, a lot of these things will connect better over, over the next few months as we share more. But, so think of it as what are we trying to do? We, I worked at HashiCorp and maybe, maybe some of the audience are familiar with like Terraform, right? Like, so when you think about infrastructure as code, and Terraform, like almost like a config plan, right?

Marcel Santilli 15:46 That, that you're applying, And we're almost thinking of it similarly. And so, when we say workflows, I don't even know if workflows is the right word to use anymore, but for us, it's like, how do you codify what your best people do? Right? And in our case, the, our best people are the people delivering and ultimately accountable and being the drivers to delivering a work product, right? That work product in our case can be an article,

Marcel Santilli 16:11 it can be a landing page, it can be copy, it can be research on something, a topic, right, for, for a customer. And so when you, when you think about that, we're like, we're essentially building the engine that bill helps build those plans. So, just to kind of, you know, give an example here. What does it take to do a really good piece of content? Right. For us, as we started to break it down, it's like, first I need to understand all the context needed. So we call those artifacts.

Marcel Santilli 16:42 Right. So, so then we're starting to say, Hey, Galileo, what does Galileo do? What audiences do you serve? What do your competitors do? Like, how are you positioned in the market? What do people say about you? Who are for the audience we want to serve? Let's go understand them better. What are their top concerns? You know, what's their day to day like and all those things. And then you're generating a lot of these artifacts and those artifacts are always kind of being regenerated.

Marcel Santilli 17:09 Right. And so then as we go through that, we under, we saw that there's workflows that we could build to make generating these artifacts better and better. Right. Now you have the context. Now you got to go figure out where do I fetch the right inputs. So that can be retrieval based on a knowledge base we already built for Galileo, or that can be, Hey, let me go and,

Marcel Santilli 17:28 not just do a Google search, but maybe I'm going use Perplexity's API and grab, not just take Perplexity's answer, but I'm going to take the 100 citations it gives me, and I'm going to go citation by citation and see if it's relevant based on the context I have and based on the job at hand. Right. So as you go through that and you formulate a plan and you execute against that plan, what we started to realize is like,

Marcel Santilli 17:51 this is way better done in code. Right? Yeah. And coding agents just got a lot, a lot better. And the reason coding agents got better is because they have all the messy middle completely out in the open. So you look at a repo, you there's open source projects, every pull request, every common documentation is actually there. Like I cannot name a single, if you take a, one of the best open source projects out there and you look at how clean those documentations are for some of them. And, and then you compare to internal knowledge bases inside of companies for how things get done. Yes. I dare you to go to a sales department and find a document that actually describe how the sales process works, that even remotely

Marcel Santilli 18:34 as good as a doc for an open source project. Right? And, and good luck finding the pull request for that, or how you describe, how you go about doing good work. What are all the learnings? What are all the back and forth? The messy middle, right? Yeah. And so for us, like, what we started doing is like, we started with the service, then we build a coding agent that builds workflows in code, and then a runtime layer that can run that. And so there are steps where we might be running hundreds of things in parallel, right? Last month we did over half a million runs of workflows. And each one of those runs can be, you know, ten hours of human work, but that's not what we're focused on. What we're focused on is if you're an expert in a topic, I don't want you to spend the next hundred hours of your time

Marcel Santilli 19:15 reading a 100 articles on a, on a subject, knowing that 89 of those might not be relevant. Right. I want you to ingest that information. Then I wanted to figure out like, where is the right place for human interventions? Right? Or expert intervention, however you want to, want to think about it. And so for us, that's one of the things we're building right now is a expert hub, if you will, or it's an orchestrator layer that takes these workflows that are running in code, right? And figures out where should the interventions be?

Marcel Santilli 19:46 What kind of intervention should it be? Is it an open ended thing? Edit this? Is it a comment on this? Approve or reject? Is it a multiple multiple choice? Is it a ranking thing? You know, is it more of like a reinforcement learning task or, or, or what is it? Right? Yeah. And, but then the really interesting thing is as you start to learn from that, you can start using LLM as a judge in the loop to figure out if those interventions are being good. But not only that, you can actually teach the people doing the interventions to do a better job. So let's say I give you an outline of an article that is for abnormal security, right? One of our customers in the security space and it's very technical, right? And the person doing the intervention, maybe they know something about security,

Marcel Santilli 20:29 excuse me, but they might not be the ultimate expert. But you can say, Hey, by the way, here's some reminders. Here's some guidelines, here's some things you should watch for as you review this. Right? So now AI is actually, are actually helping the humans do better interventions, right? And add more value and connect the dots for them or remind them how to do it. And the opposite is also true, right? The, the right human that you calibrated and that, you know, they're an expert in something can look at something and go

Marcel Santilli 20:59 A versus B, right? And you do enough of those and now you have data, enough data to start to fine tune these, what is now a workflow into essentially task specific domain specific company specific models, or doing a mixture of expert type of approach. Right. And so, but, but the opposite doesn't really work. So today, if you go into some of these models that have a lot of chain of thought, you know, and they do a lot of thinking and you start to see them describing the thinking they're going through, you just want to hit stop, right? In Cloud Code, you can do that. You can use escape,

Conor Bronsdon 21:32 right? And, and so like that's, those are some of the layers we're thinking about. And then the final one is kind of the, the learn engine that kind of layers on top of the whole thing. I love it. Marcel, I'm so excited to see what comes out of GrowthX NEXT. You have such an incredible team you've built here, and I've had a sneak peek at some of the products you're doing as well. And it's really exciting. So definitely make sure you're checking out GrowthX,

Conor Bronsdon 21:55 for all your growth needs. We're big fans of them, and you can also follow Marcel on LinkedIn where he puts out incredible content about things like the rise of voice as a modality in AI, and so much more excellent advice for folks who are building and scaling AI businesses. I also just wanna call out before we end here, like Yeah. I love that you brought up this coding example because,

Conor Bronsdon 22:15 as you point out, there's great documentation. There's a great dataset out there with all these different repos. And we had the co founders of Poolside on the podcast a few weeks back, and they talked about this. They're like, We think we can solve coding first because, you know, the English language is is messy. It's it's fun. It's useful, but it is not as clear. Whereas, hey, look, we have a ton of incredible code that is clear. We have good examples. We have bad examples. We have clear best practices.

Conor Bronsdon 22:39 There's such an opportunity though to go beyond that with knowledge work. You know, coding is just the start of where we are gonna enable people around the world to build faster, build more, and build better. And I'm incredibly excited to see what GrowthX does on that. Thank you so much for coming on the podcast today. It's it's great to have you on Chain Thoma. Yeah. Thanks for thanks for having me. Yeah. It's our pleasure. We're gonna have to swing back around, do another recording here once you move to your new offices after you, you know, you hit that 100,000,000 ARR mark. We'll we'll come back and do another recording. So That sounds good. And for folks who are listening, make sure that you are watching this episode on YouTube because you're not gonna get the full experience without it. I I have to say, Marcel, I you still have to have your Charlie cameo, though, which you you almost forgot to do. So

Conor Bronsdon 23:17 This is Charlie. We

Marcel Santilli 23:19 Do do do we just should we just leave it as a mystery? We'll just leave it as mystery. Yeah. I Cool. Gonna mask it here. Maybe we'll do a behind the scenes telling telling people Yeah. I like it. I like story it. Behind Charlie. Yeah. Well, thank you all. Thank you everyone for listening.

Conor Bronsdon 23:31 We'll see you again soon, and make sure that you have subscribed to the podcast on your platform of choice. And can you leave a review, leave a comment. We really appreciate it. It helps folks find the show. Marcel, thank you again for coming on. Thanks.