About this talk
This talk explores the development of four AI-driven companies and shares insights on building businesses that leverage AI agents effectively. The speaker, founder of Full Stack Ventures, emphasizes a strategic approach to product development without relying on human labor for sales or support. He highlights the importance of starting with a clear intention and quickly iterating on product design by utilizing AI for user experience and optimization. Key focus areas include using AI for infrastructure design, the value of self-serve models, and the significance of an API-first approach to enable seamless integration and scalability. Throughout the session, the speaker provides practical advice based on real-world experiences in launching and managing these companies while navigating the challenges of growth in a crowded market.
Full transcript
Okay, good. Everyone can hear me? Headphones on, great. Great to see everyone. Thanks for joining. I will try to keep you entertained before lunch in this next half hour. Um the talk of today is it's called the AI first playbook. And what I want to share with you is the learnings from the past 12 months of building companies just using AI agents. Um so really we built
four different companies during this time and the learnings from the first to the fourth are completely different. And I want to talk you through how we're approaching things now and how we used to do it. And you can apply those in in any tool you're using. Doesn't mean matter if you're using cloud code, if you're using cursor or or codex or anything else. It's more the way
of thinking and and what you're building that that I'll get into and and then also how to think about growth when you get there. What I will not cover for you today is how my AI AI agent does the work of a $30,000 marketing team. What you need to comment to get all that magic that doesn't exist. These are These are, you know, LinkedIn headlines that usually
end up being a very simple end-to-end workflow that doesn't actually work. What I'll get into today is like the actual nitty-gritty of yeah, what you can build to to build your business before you hire anybody Um so quickly about me, my background. So I'm the founder of Full Stack Ventures, which is a company studio venture studio where we do exactly this. So we focus on building companies
AI first and we avoid anything that requires humans. So we don't do sales. We don't hire onboarding teams, customer support. All those things we don't do. It means also that certain things we cannot do and we're okay with that. We're trying to figure out where the limits are of running companies just with AI. And the only thing you can do that is the only way you can
do that is by starting from zero because before this I did run other companies. I've been a founder four times. I won't get into all the companies here, but I've also been at bigger companies like ZoomInfo, few thousand employees. I was CPO at DealFront for 3 years, which has 300 employees. And if you're at these bigger companies you just can't change certain things. You have too much
revenue, you have all these customers, you have teams that all need to be taken care of and looked after. And and that's true, but that also means that you can't just rip out your onboarding team because you're going to lose revenue, customers, and people, and it's just really hard. I found it really hard from that position to get exposure to AI, which is why I decided to
to leave about 12 11 12 months ago now and start Poolside Ventures. I also organized an event for founders in Barcelona. I won't get into that too much, but I also wrote a book about product-led sales. More relevant than it might sound at first because it's all about when do you use automation and when do you use people to drive the sales of your business, which is
something that I didn't start thinking about 12 months ago, but long before that. And it will feed back into our story later. All right, so enough about that. Let's talk about the companies that I'm going to use as an example today. So, we'll use each of these four companies to explain to you what works and what didn't. I'll give a quick intro on each of them. I
won't spend too much time because I want to get actually into the learnings. The first one is MeetKite. It's a Calendly alternative focused on AI agent use cases. The second one is called DataMerch. It is a company in contact data solution. So, if you are a company and you want to sell to other companies, you can use this solution to uh to find that data. Um there's
Scope, which is a workspace for non-technical people to start working with Open Claw. Uh and then there's a fourth one that um allows agents to get access to paid APIs like Data Merge or others that sell data or or paid services without having to go and sign up everywhere, create an account, get a token, put a budget, and so on. Um so those are like the four
products we've worked on. Three of them are live. Uh the fourth, it's marked secret because we haven't announced it yet, but it's running in the background. Uh and we did all of that in 12 months. These are not MVPs or betas, these are products that are actually production-ready uh and where we have customers on them that are using them and that uh in some cases, not all,
are also paying for them. you'll notice that I'll keep saying we, but uh I left the team size uh obfuscated at first. Team size is one, it's me. Uh it's me, myself, and my AI agents, I always say. And I try to keep it like this until something works because if I hire people before that, I feel like I will have failed. So trying to really push
the boundaries of of what's possible there. Good. So first, like a in terms of mindset, something I want to get across before we get into the examples, is that everyone now thinks like oh I can build something great, so I can start a company. And I think most of my calls nowadays, when I speak with people on uh through a video call, they're like oh can I
show you what I built? Let me show you what I built. I built this thing. I made this thing with my kid, it's better than the the real thing, and so on. I I hear this all the time. And I, as a builder myself, I know why this is exciting, and now way more people having this experience of like you built something, someone else or yourself uses
it, and it works. It's amazing. It's a It's a really good feeling. Doesn't mean it's a business, though. And uh AI cannot help you with all the parts of the business in the same way as it can help you create the product. There's some of these models that are really good at getting a good user interface out. Uh lovable Claude, uh Codex got better. Um and so
I would say if if I take like four examples here of things you would need to do to create a company. These are not all the things, but these are just like four things that are all critical to get right. Uh the user experience, you can definitely do that with AI agents right now. Uh if you kind of know what a good experience looks like, you can
have it created. However, if we're already looking at infrastructure and we're still staying on the building part here, if we look at infrastructure, it becomes harder already. If I'm not telling the agent what kind of infrastructure I want, then it'll do weird things. It'll create a database on a .md file. Uh there will be security gaps left and right. There've been many talks about this at at
the event today and yesterday. Um and it it won't be scalable. It won't be efficient. And there's just lots of problems that you'll run into as soon as you start hitting uh high volumes or even just expose it to the real world. And so if you don't know what to ask it at that point or what to tell it, you're already going to have problems. So AI
can help, but requires already expertise. Luckily, the room here, a lot of us are technical. So we probably have that expertise. So let's say we can get that step done as well, right? Then we get to the more the business part. Uh first of all, the business model, like what do I sell? What do I charge for it? How can people buy it? Where can they buy
it? Those kinds of questions. And even one step further, like how am I going to find customers? Am I going to do ads? Am I going to do sales? Am I going to do uh is that is it some kind of product led growth motion? These kind of things, uh this is where it becomes much, much harder because building got easier, growing got a lot harder. 10
people can build the same thing at the same time. Customer only going to buy one. So, like the dynamics of growing and selling are completely different from building. And it's a lot more noisy in the market because there's a lot more products out there right now. And so, when you ask an LLM, which is meant to give you the most likely response, uh what you should do,
then what you'll get is a very average outcome, not a go-to-market strategy that will give you an edge. And so, I found it um challenging at best to use AI for that and found that really it's personal insight there that will help. So, that brings me immediately to like my first out of three main um steps that I see in in getting a company going, getting it
off the ground. Uh and the first time is really to be intentional. And it's not so much AI, it's often experience you've had or a point of view that you've had yourself. Um if I look at the three companies, I'll just cover three out of the four because the last one isn't live yet. Um but for each of these companies, I had a specific point of view
why I think that would work. I don't know if that's the right point of view, but I need to have this intentional point of view and validate that if I want to actually do something that stands out, if I want to actually be successful. Um for Meetbot, for the scheduling solution, that was more uh business model focused. It was like, "Okay, people don't want to pay $20
a month to get meetings scheduled, but they'll pay for meeting." Uh and that's a much bigger audience. That was the idea behind that one. So, we had a different business model that allowed us to distribute differently because we could go agent first and so on. For Data Merge, the intentional idea, why did we start to work on that, um is that the data that we have, like
there are company data APIs, but the the specific hierarchy data and legal data that we have wasn't easily available through API. Again, something I knew because I'd been in the space for over 10 years. Same with scheduling. I built like eight scheduling solutions over time. So, I had a point of view that I think was an opportunity. And that's what I used as my input to then
start defining what I wanted to build with AI. Um and for scope, this one was like I was getting a lot of requests where people were asking me, "Hey, can you help me get set up with Open Claw? I really want to get working with it, but you know, they get stuck on I don't even know how to get the the server up and running." Cuz these
are non-technical people. And I was every time doing the same steps. And I was like, "Okay, I think I can automate this. I think there's like a product here that can help people get like can give a lot more people access to this." Um and so, you start with this intention, and then with that, you start building your product around it. If you don't have this intention,
then you're just going where the wind blows, and it's very hard to get real results there because depending on what you're trying to achieve, it might make you do different things. And and the neutral middle is never a good place to be, uh especially not in the noisy environment that we are in right now. So, you have your intention. The next thing is not just build, but
build fast. And this is something that really, especially in in the first part of this year, we got a lot better at that. Um I'll take the examples as well. Meet Bot took 12 months to build. Long, I know, but this was while I was still at DealFront, and I was kind of working with a developer on the side. This was a company where um where I
started with a developer and later reduced it back to just being me running it. And I want to highlight for each of these building experience experiences, what were the things that we took away from that, because the takeaways from Meet Bot became table stakes for Data Merge. And what we learned there was then like the default plus more on scope. Um the first thing, and this is
something that is heavily underestimated by by people, is that it's important to optimize for cost efficiency. So, you might be used if you're at a larger company, you might be used to like buy a piece of software for this and that, and you spin up some extra services for uh running it, but actually like the more cost you accrue, even if it's a few hundred a month,
that's going to kill your business at some point because you need time to figure things out. And so, the cheaper you can run your business, the the more chance you have of succeeding. And if you look at some of the most successful companies out there, the biggest ones, uh they were actually super efficient and focused on efficiency in the very beginning, including Amazon, Google, and whatnot. especially
if you're doing this on the side, or if you're not sure where to go with this yet, the more time like by keeping your cost low, you'll have a lot more time, and you can use building capabilities to do that. You can use open source much easier now by deploying it yourself with your AI agent, or using one of those services like I use Elastio a lot,
where I just click a button and now it has an open source running. So, I'm not paying for email marketing anymore uh where which starts at 250, 300 a month. No, I'm hosting a machine, it costs 15, and I'm done. And these considerations used to they it used to be the thinking that this would distract you from the mission and uh like they say, it doesn't make
your beer taste better, so just buy it somewhere. But at this point, because there's so much available in open source, and it just fits so well with building capabilities, um it just doesn't make sense to commit to that kind of cost anymore. And it gives you a lot more possibilities to uh to have time to build, but also to be a lot more profitable once things do
start working, which means you can spend more to win a customer, which means you can win against your competitors. Um so, cost efficiency, super super important. Uh two other parts there that that we learned is one self-serve is has always been important but in a world where you could sell a product and people may or may not start using it. People companies got away with not having
great products. Uh in a world where everyone can build a product, if your product's not usable, people won't use it. It's as simple as And there's going to be more and more products that just get used by technology. And so self-serve super super important. Not just self-serve for humans. I'll get back to this on one of the next examples, but also for agents. Can they start using
your solution? Um and really obsess over this part. Like how does someone find you? How do they from there get to uh to their first value and to a habit then? And um yeah, using the public API as a building block, I think Salesforce is a great example. Recently, we were already focused on this uh last year where it's like uh Salesforce says, "Okay, all our services
are now available as an API endpoint." That's the right way to think about it cuz it allows you to do two things. One is that uh you you and if you have a team, your team can now build anything against these APIs by just telling an API agent, "Here's my API spec. Now build me a new onboarding thing. Build me a new feature. Build me a new
model." Build like you can just have it build on top of your API. And uh you don't have to think through the spec in that much detail because the core is available to the agent. That does mean you need to include all the features up to ideally account creation. So not just your main thing that you would want to expose to some partners, but like it needs
to actually be possible to run your product through the API, uh to configure it, to use it, and to get the value out of it. And if you can do that, then you unlock a lot of uh other ways to build against uh product. And the second way then is for other companies to do that and for agents to start using it by themselves. So the API
is actually the core of everything. If you want an MCP, you need an API. If you want to integrate with agents, you need an API. So thinking API first, even if you're going to create a a user interface on top of that, thinking API first is really where it starts. And at this point, we built the API first and then we start figuring out if it still
needs a user interface or not. At the end of it though, or at the end of it when we launched and where we are now, the biggest challenge I have with with Meet Bot is that we started with a human in the loop and it's very hard. These processes are kind of made now for a person being there. So even though uh I took them out, the
processes are still kind of expecting in some cases a human. So it's very hard to undo a setup where you have people in place. Like even one person, I was like afterwards, I'm like that the the company is is like a bit handicapped if you will in terms of how it can operate because all these processes were made by a person who assumed someone was going to
be there to take care of it. And that's what I did completely different then with Data Merge as a second company where I was like, "Okay, let's see." And this was November last year. So the models weren't as good as they became in December and I was working with Cursor at the time trying to see like can I build something with only a coding agent? Can I
just do it without a developer entirely? And so that by itself was an experiment because today we will say, "Yeah, probably." But in November, it was actually not that obvious if that would be possible. Um and um yeah, so so work through that and got it live and it it but it ended up being a similar kind of setup as you'd expect when when you're developing with
a regular team, right? So we had a front end instance, we had a back end instance, we had a worker for async jobs. That doesn't really change. That that doesn't change. You're still building in a similar way because that's just what you need for scalability and for maintenance and those kind of things. What we had focused a lot more on with data merge was AI indexing. So,
like can we be found by AI? And some of these things people are going to know already, but I'll I'll highlight them anyway. So, the most basic one is like does your page require JavaScript to render? If yes, then many AI agents in in their processes are not able to read it. You're done. Like they're just not reading it at all. So, we started to I now
use Astro for all my projects, which allows me in the front end to use to to have it server-side rendered, but at the same time have react components load in I haven't explored other solutions, but this is working well for me. But but like the main thing to think about is not have JavaScript rendering block your content. If you put in an FAQ because you want AI
to read it, don't put it behind some kind of JavaScript thing because it might not be able to see it. Like just put it in there so that it can read it. and of course people have heard about the LLMs.txt. Yes, it should exist, but what should it actually look like? And if you look at the companies that are doing this very well, Recense people might have
heard about that they got very famous about being found about AI that they they they caught that wave very very well. And so I started to look at how did they do their their LLMs.txt. And it's not like one big blurb of text. Actually, the first file that you find that is on the on the naked domain is actually an index file to further files that have
further information about the the specific topics you might be looking at, right? So, it's actually a text-based database for for the model to find what it actually needs. applying that uh to make sure that the AI agent understands what does the product do, who is it for? Uh like that by itself many [clears throat] many websites fail just at that. And it's so easy to test. You
don't need any solution. You just put your URL into an AI agent and you ask it what does this company do? And it'll tell you or not, right? And then you know exactly what to work on. And if you go one step further uh I call this AI indexing, but then the second step there is uh AI interpretation. So does it actually know how it works? And
this is especially with software. Does it know how to use your software? And there the skill.md file, which you can then reference from your LLM.txt is very important. Again, I'm going to speed up a little bit looking at the time, but again you want this to be multi-layered and have examples how to use your API, coming back to the API as a building block, so that when
I ask my AI agent how should I call how should I get data from Data Merge? I always get the same answer. Here's how to do the call. I basically just run the curl and it works right away. That's That should be the level of quality you get from it. If not, it's not working, right? And you're not going to be found. You're not going to be
There's a detail around MCP optimization but um I'm going to skip that one for now to look at the looking at the time. The biggest challenge we still had at the end of this because we use a lot of AI uh in the workflow of providing the data to the customer is to get consistent responses. When you're providing data using the LLMs we all know they never
they don't always ask answer the same thing. And even when we kept the prompt exactly the same and the model pinned and everything, then the API would change somewhere else. So there's just a lot changing. Um so that's still a challenge that that we're working through on that front. And then on scope because we've just launched recently uh the the biggest thing to highlight here is that
we were able to launch within 4 weeks. So if you look at the timelines at the top, right? 12 months, 4 months, 4 weeks, and the quality of the products actually went up, not down. And that's because I got a lot more comfortable with how to ask like what do I let the AI do and what do I want control over. So, my workflow is like I
first go for a walk and I get the full brain dump transcribed into a text file, and then I take that, I put it into the AI agent, make sure it asks me some questions to come up with a first product description of what I'm trying to build. That I review in detail. That's usually just a few pages, so I can read through that. I make sure
that I'm fully aligned, and then I make I have it make an implementation doc out of that, which is like 100 pages. I don't look at it. I just say go, and I'll look at the product afterwards and see if anything needs to be adjusted, but usually at that point I've given it some guidance on the infrastructure, but from there it it can it makes a lot
of the right decisions with just that. All right. So, I want to get to the growing part. So, iterating on growth is very important. And the thing is that you usually get growth wrong from the beginning. Like almost nobody launches and magically starts getting customers in, and I just want to show you how much of the things I tried failed, and this is not my first rodeo,
but still a lot of stuff is failing. And the whole point of growth is that you actually want to be failing so that you can start learning what works. So, having a high failure rate on your growth experiments is a good thing cuz it means you're trying a lot of things, and you're you're hopefully getting closer to something that does work. I'll give you some quick numbers
for for each of the companies. So, Meatball we launched in November. The cost of running it for us is only 20 bucks a month. So, I don't mind if this takes longer. We've been a bit distracted by some of the other companies, but it it's it's also not generating any revenue at this point, but luckily it doesn't cost anything. We have users, but the freemium plan is
generous, and therefore revenue is taking a while. The first idea was that we would grow this through business development, um that we uh would get partners that would integrate our functionality. Uh but the partners we were aiming for, they had the need, but it turned out they usually just in they built it themselves. And there just wasn't enough market. Something I should have checked beforehand, probably. Um
but anyway, that that was then uh invalidated. Then AI optimization and search engine optimization, I'm going to cover it together with Data Merge because it's a topic there as well. Did a lot of work on that, and you can use AI for that quite well. If you have it create specific uh content, don't let it create general things for for general keywords, but specific to your product,
stuff that only your product can do, it can create pretty good content and get you ranked on there. However, if you don't have users and if no one's talking about you online, it won't matter because everyone's launching products now, and it just it won't even get picked up. We're we're barely getting any visitors from that, even though like if I I've had some consultants look at it,
and we we're doing everything right. But what we need at this point is people to give it reviews, to use it, comment about it, talk about it. And so, what that uh taught me is that for that first wave to get going, for that first momentum to build, I kind of ended up uh still needing to talk with humans first. And it's good to talk with your
customers, don't get me wrong. And you can do that yourself. Um because with Data Merge, what ended up working because we are generating revenue there, is that we got inbound interest from companies that wanted to buy data sets instead of just access to the product. And that allowed us to uh at least pay the bills and and have a a reasonable business out of that. Um so,
that's a good thing, but actually the product we built is not doing what it's supposed to do just yet. And so, what we tried differently with Scope, and which is now working very well, is that instead of just putting the page online and saying like, "Okay, we do SEO. People should find us. They should start using us. We started to do workshops and in these workshops we
help people to, for example, get started with Open Claw and with that they get introduced to our product and we capture that very early feedback and what's working and what's not working and are they staying actively are they staying remaining active users afterwards, yes or no. And I think that is in the beginning before you start doing SEO or AI optimization or anything, it comes back to
this very basic idea that I mentioned earlier, which is that usage is the core. If you cannot get people excited about your product when you introduce it to them personally, then it won't matter if you're putting it on SEO or or on AI because you need people to talk about this thing to get ranked properly and to get recommended. Um so, focus on usage is really the
key here and you do that in a few ways. Um I'm not going to cover each of these in detail because I've had uh them come uh I have mentioned some of them already, but this is all the things that matter to create a good user experience, a good product experience. Make sure that you integrate into people's day-to-day. Make sure again onboarding is good. Make sure that
agents understand what your product does and look at what people are doing when they use your product. Look at the recordings and ideally have them in a workshop and talk with them, see what's going on because if you crack this part, then actually the revenue will follow and this is really like the piece that's missing many cases. We're still going to have to talk with humans at
least at this point to get it right. The one thing I would tell you not to do is to launch an MVP. This concept of MVP comes from before AI and MVPs are uh not enough anymore to get people's attention and also there's no excuse, right? It takes you like maybe another extra day to make it a great experience instead of a crappy or minimum viable experience.
So, just make it a great experience. Make sure it integrates with the right tools. Make sure it's easy to use. There There's no excuse at this point to launch something that is like bare-bones. But at the same time, focus on your specifically intended use case, coming back to your intention. So, like this is the thing where we're going to win. Okay, let's see if let's make that
part great and get that out instead of building another 100 features around it and figuring out if if maybe some of it sticks. If you work like this, you'll be much more focused. You won't always get it right, but at least you'll be learning and you'll be methodical about getting to a place where you have steady customers coming in uh and where your business starts growing. And
you can do all of that before you need to hire anybody uh through the things we just went through. I think I have one more slide, but I need to finish, so thank you. Let me just >> [applause] >> Okay. Thank you.
