The $30M Zero-Employee Challenge: Building with AI Agents, Not Humans

Speaker A: Welcome everybody. This is Ryan Staley and I am back with a very special guest today. I have Amos Bar Joseph, who is the founder and CEO of Swan AI. Something cool that, uh, Amos is doing is he's a trailblazer using AI agents. And what he kind of defined and stuck a fork in the ground, or whatever you want to say in the ground is him and two other founders are not going to hire anybody. Grow to 30 million in ARR with just 3. 3 people and a swarm of AI agents. So I thought he'd be awesome to have on the show. I thought you would enjoy him. Amos, welcome. Happy to have you, man.

Speaker B: Thank you so much. Ryan, I have a question. Can I actually take you with me to like every social gathering so you can introduce me to these people there? That would be amazing.

Speaker A: Uh, so you like the podcast voice? Yeah, my family makes fun of me. My kids are like, or my wife actually is probably the key one who likes to make fun of me. She's like, oh, are you going to use your podcast voice to do this? And I'm like, shut up. Right? So anyways, man, like you got, you got to bring the heat, right? Because everyone's like, yeah, you know, you got this. So anyways, I appreciate you appreciating me appreciating you, if that makes sense.

Speaker B: So that makes a ton of sense. I wish I had a voice like yours, Ryan.

Speaker A: Well, you're doing some cool shit, man. You got a voice that, in writing that I love and that's kind of how we came about. I saw what you're doing on LinkedIn. I'm like, hey, this would be a cool guy to have on the show. So walk me through, man. Give us a little bit of background about you, your unique point of view in terms of how you're tackling the problem that has been classically handled in a different way.

Speaker B: Yeah, yeah, so it's actually, it's not my first rodeo. Um, I've built two, uh, B2B startups before. Did one with enterprise sales, the other one was more like SMB sales, but both of them was, uh, you know, were purely about the unicorn playbook, the growth at all costs playbook where um, you raise big before you hit product market fit you uh, grow to a 30 people team before you get to uh, the first million ARR. And then you try to brute force your way and some lucky, you know, old fashioned way to get to $10 million ARR. And if you are lucky to get there, like your entire company is built on a very sick Foundation. And so when I actually started this company, Swann, together with my two co founders who were with me in these two previous shows, we actually had a, uh, interesting discussion. We said, why don't we actually, you know, rewrite the playbook? It's actually time to write a different playbook to how do you scale a startup in 2025? And it's not that that, only that the growth at all cost model is dead. Is that um, you know, when you look at AI agents, there's finally an opportunity for startups to scale with intelligence, not with headcount. So we, you know, told each other, why don't we have like a bold bet with twin friends. We said we'll scale to $30 million and ARR, just the three founders, we won't hire anyone, will be just us and we'll build an army of AI agents. Instead of investing time in hiring people and building org charts and doing a lot of meetings, we'll just work with these agents to scale. And you know, what started as just a close conversation as you know, over time evolved into a movement that we call the autonomous business. Today we're getting close to already 200 businesses using the platform. It's, these are like real businesses that a three person team supports them. So it's getting uh, it's getting excited every day.

Speaker A: That's awesome and I appreciate that. It's funny because like really I was looking at hiring someone for a role and as I was going through like kind of the jobs to be done or the pieces on what I wanted them to handle, I started thinking I'm like, I could automate and have most of this done and I don't need to hire a person for this role. So after I went through the exercise, what I originally thought I needed for the role evolved into something completely different. So uh, I could uh, relate to what you're talking about because like I said, I literally just did this a couple weeks ago and I'm like, shit, this is mostly automated. Or I could string together workflows or prompts and use case anyways, so it makes a lot of sense. So let's get into it. I'll go overhead. What are you going to say?

Speaker B: Yeah, already I have something to say. Right? We started off the bat with something interesting because you actually touched like the most important part of building an autonomous business, right, um, at the start, which is the fact that you know, when you revise that job description, you've actually discovered that there are like different things that you actually need. Right? So if you look at AI as core to that role, then you were able to reimagine that entire function. So what I wanted to start with is that the entire goal of know, building an autonomous business, to be the first movers into that space, is to discover how an autonomous business actually look like. And it's not about automating ourselves away, it's about reinventing these processes with human AI collaboration at the core of the operating system. So I loved your I.

Speaker A: So, so let's walk through that, man. So like, and I, I know how I kind of approach it and how I approach my clients because I'm a big believer too that you could really grow an organization substantially just by multiplying intelligence. And so what do you think is the belief system and the mindset structure you need to have to start to go down this path and to start to orchestrate a company that is designed from the ground up to be AI first. We're not talking about all the crap of CEOs from big companies saying we're now an AI first company. Right. But uh, what do you really think it means from your perspective as it relates to that?

Speaker B: Yeah, yeah. So it actually, it all starts with a constraint, with, ah, a simple constraint. We can throw bodies at a scaling challenge. That's the number one constraint that you start with. And if you tell yourself that no matter what, if you want to scale a specific process, you cannot add more headcount to actually scale it, then what do you do? Right? And if you add another constraint, okay, we're going to use intelligence to do it and you connect these two then, okay, you start looking at problems differently and you start looking at solutions in a very different way. And what you start to follow basically is kind of like an uncharted territory where you try to reimagine that process in an iterative process. Okay. So it's about iteration of human AI collaboration step by step. And when you look at, for example the best use case in Swarm, that we actually seen it firsthand is the biggest scaling challenge that we had early on was support. So we're serving businesses. It's not like a consumer, like a prosumer solution. When you're serving businesses, you need to give that extra support touch. They want their answers, they want the service, they want to understand how to use the platform. They paid big bucks for it and now they want to bet. And so we start, you know, getting all these support tickets. And usually what a startup does, they hire uh, a support rep or maybe even support manager to build the Support organization right from the start. Um, and we said, okay, we can do that. What can we do? Let's add AI to the process. How can we do it? So we just built a simple, um, AI bot into Slack because that's where Swann lives by the way. It's a Slack app that you chat with and work with in Slack. And so we just added another layer where Swan is now able to answer simple support questions within that same interface within Slack. It has a knowledge base that we gave it and now it is able to answer these questions. But what happened initially is like, you know, 80%, 90% almost of the questions, we didn't have answers to them in the knowledge base. So Swap. Swan didn't know how to answer them. Right. So we added another layer. It was an iterative process. We added another layer where Swan is now able to escalate a question to the founder if it doesn't know the answer. And so you could use Swan to reply to that user so the process is more streamlined and you get that AI capabilities when you reply to the user. But then we say, okay, why not add another process there? Why don't. Swann will document that Q and A into the database, into the knowledge base every time you answer a question, so that you only need to answer a question once. And so what happened is just a simple exercise in adding a simple layer of AI. Automation evolved into a self learning support system that knows when to escalate to the founders. So there's that human in the loop but knows when to document it and learn from these interactions. And, and now it covers more than 75% of our support requests and thousands of tickets, uh, each month basically.

Speaker A: Okay, so I like how you just kind of kept building on top of that. So what did you build that on? Like, how did you build or construct that? Like, are you using Slack's AI? Cause I know Slack has an AI model. Are you using a no code model originally when you started or how did you design that support agent?

Speaker B: Yeah, so the brains behind the agents are, uh, ido, the CPO and Nev, the cto. I'm the spokesman, but I know how to speak about it. Basically we use N8N for the, you know, the logic and all of the orchestration around it. We use Slack as a conversational interface because we believe that every agent should have a conversational interface so that you can interact with it, you can give it feedback. Every, every agent has a memory.

Speaker A: Mhm.

Speaker B: And our design, so you could always tell it things and it could add it to the memory. So and so next time it will encounter a similar use case, it will be able to, you know, take that feedback into consideration. So we're building on top of N and Slack as kind of like the interface and that's how we connect things. Usually we also have a lot of different tools that our agents has access to just by the nature of our product. Because Swan, it's an agent for go to market companies, go to market organizations. And so we already have a lot of, you know, know, commercial agreements with data providers. We uh, already have built in interesting, you know, CRM integrations, things like that. So we just added these out of the box integrations to these N8N workflows and what you get are very powerful agents that have, you know, capabilities of enriching and getting, you know, B2B data and searching online using perplexity, getting into your CRM data, getting into your emails, things like that.

Speaker A: Okay. And then from a data store perspective with the uh, self learning agent, just to put a bow on this. Right. And then we'll move on. Because I'm curious, are you storing that data in Pinecone? Are you storing it in the cloud or kind of how are you basically capturing and storing that data?

Speaker B: Yeah, so Pinecone is uh, um, the main database that we use. But we actually learned that a lot of information could be stored in the prompt level as well. So we try to differentiate where do we can we actually put that directly into the prompt. So for example, you know, maybe your ICP is something, you know, it's great for the agent to always have as if it's like a specific type, you don't need to save it somewhere else and then kind of like fetch it every time. So there's like things that you can actually save at the prompt layer, but then you know, um, for like more you know, bigger context that you need for specific use case you want to save it in Python.

Speaker A: Okay, yeah, that makes sense. I imagine that improves latency as well. So like you're not. If it's calling multiple sources, it probably slows it down I would imagine. Is that kind of why you did that?

Speaker B: It also makes it easy to not always try to foresee in which use cases would you actually need to call that database and you know, to specific enrich it with additional context. So it's an action, it's something that happens, you know, it's not natural for the agent to do it yet you need to go to Pinecone to get that data. So you don't need to think of all the use cases in advance. It makes it easier for you. And that's actually a super important point. What we try to do when like our uh, core principle when we are building agents is that we don't try to optimize for use case coverage at ah, you know, the early phases of building an agent. It's not about an agent that can solve all these different use cases. It's an agent that has the core capabilities structured in and can maybe could answer 70% of the use cases at the beginning. But then we have these feedback loops and so on that memory and that ability to actually tell the agent what, how to improve itself within Slack. That gets you from that 70, 80% to that 99%.

Speaker A: Okay, so you have your starting point, you obviously have your data store. You embed it like the prompting if you will to incorporate like the corporate support foundations and then like that you're using Slack as the feedback loop to continually update it and then it learns from that and then continues to optimize. Okay.

Speaker B: Yeah, exactly.

Speaker A: Very sharp man. Love that. So let's shift gears a little bit and I think what we'll do is, you know, I know I'm going to have you back for episode two, for episode one. Let's, let's stay on the like how you're leveraging this internal agent wise. And then I want to hear how in the next episode we're going to talk more about the go to market use cases with Swann. I think it'd be cool to go deep and maybe even show it if you'd be up for it, man. But let's talk about you because I know internally, um, being that you're the spokesman or the prime minister I think you said. Right. You mentioned also that you have 20 plus agents. I thought that was a number supporting you. What does that look like and how do you have that set up to support you? Because if you're you know, basically one person working on sales and you have 200 clients over, I think it's like a, you know, three to four month time frame. That's fantastic man. But like what does that look like and how are you making that work?

Speaker B: Yeah, definitely. So the first thing we realized when we started building the AI agents is that you know, we didn't look for these off the shelf AI agents to solve all of our problems with. We didn't look for AI SDRs or like for AI support agents. It wasn't about that. What we realized is that we have our own unique characteristics of our business. Me, myself, as the solo GTM practitioner at the company. It's a very unique situation. I cover a lot of areas that it's not normal for one person to do. And also I have my weakness and strength. And so what we started with is building an army of agents around my core advantages and core strength that I haven't go to market and around my blind spots. And what, you know, in hindsight, when we reverse engineer it, you can look at it kind of like as a funnel. And I'll explain how, you know, these 20 agents. And we're going to talk about not all of them right now, but how they actually work in an orchestration mode in a funnel, and how they're all, you know, part of a bigger system that is working together.

Speaker A: Right.

Speaker B: Uh, they are not like completely isolated. They are working in tandem, basically. So my biggest passion, and that's where the AI agents are built around, is content, is storytelling, is writing. That's what I love doing the most. That's my unfair advantage as a human being. And so I Write posts on LinkedIn documenting our journey on becoming an autonomous business. And what we started building is these agents around that. So first of all, we built Shakespeare. That helps me shorten the amount of time I spend on actually writing these posts and performing research and bringing data, um, to make my post data backed. Um, it has a knowledge base of all the other posts that I've actually written, all our manifesto and content pillars and things like that, and is able to synthesize it. So together we collaborate in taking the time to write a post from like, you know, three hours, two days, sometimes to like 20, 30 minutes to generate consistently over 1.5 million impressions every month on LinkedIn. And then if you go like a step below that in the funnel, you look at the people that are engaging with the posts, not just the impressions, right? So I get over 15,000, uh, reactions to my posts, uh, each month. And so we have the observer, which actually monitors the engagement, uh, on these posts, uh, the leads that are engaging with my post to surface, ICP leads that are repeatedly engaging with these posts and showing kind of strong sentiment of alignment with the content that I'm talking about. And then surface these leads on Slack so that I can reach out to them, um, in a very, uh, and get like, extremely high reply rates based on that. Then if you go another step down in the funnel, you have connection requests, people that saw my content, but then they actually sent a connection request, like a follower or a connection request to myself and I get around 300 connection requests each day. And so we have the connector which goes over these connection requests, um, uh, high intent ICP leads basically and starts a conversation with them and LinkedIn so I can follow up if there's a reply. I don't need to be the first mover, I'm the second mover to these conversations. And um, I have like a very fixed process to take it from the reply to a meeting booked if they are indeed qualified. And so after that I have folks coming, you know, to the website actually. Then we have the SWAN agent which takes care of all, you know, the website visitors that didn't convert and actually identify who they are. Surface high intent leads and, and enables me to reach out to them on LinkedIn. Then we have 20,000 monthly visitors. So that's a lot of volume to actually work with. If you go below that we have the inbound forms, so we get around 2,000, you know, intakes inbound forms in um, swan in the website. And so we have an agent that goes over them, routes out these leads into high value leads with a demo with me we have free trial, which are low value leads and then we have non relevant leads that go to a waiting list. Um, and the process continues from there. Ryan. After that we have a prep agent that helps me prep for the demos. Then I have like a follow up agent that documents all the conversations into CRM, prepares tasks for me. So I don't need to think about a meeting, I just go into the CRM. I know exactly what I need to do. So this is kind of like an example of how they all work in tandem in the way that there's a funnel that is built around me, how I work, what I need help with and what I'm good at. Okay.

Speaker A: Huh. That's fantastic, man. I love the meticulous detail. So if I'm reading the tea leaves right or just kind of how you're, you're walking through this, are you using the pre website? All the pre website, basically agents, are those all uh, N8N and then the SWAN agents are everything from the website moving forward or like outside the research. Right. But like where's the delineation there of what's kind of homegrown and then like what's part of the solution that you guys offer?

Speaker B: Yeah, so the majority is actually homegrown Swan. Right now the main, main, main use case is actually turning your website visitors that didn't convert, that didn't fill out a form into qualified pipeline, you know, taking it all the process from identifying who they are, qualifying them, segmenting them, finding the contacts, personalizing messages, and reaching out on LinkedIn and email. So it takes that process and end to end. That's what Swan actually does. All the rest of the agents are homegrown, but I would say. And guys, if you're listening, you're the first, uh, to hear it actually, uh, because we're launching our beta soon, um, in kind of like the next two to three weeks. And it is an autonomous GTM engineer that will actually help you create agentic workflows for your go to market motions. From prompt to pipeline, basically you can turn any go to market idea that you'd like to have into an agentic motion. So all these agents that we're built that are homegrown, what you want to provide SMBs with is that engineering resource that could just build it for them. They were able to just chat with it in Slack and it will be able to take all these capabilities that is doing on the website visitors, but across different use cases.

Speaker A: Well, that's exciting, man. Super exciting. And what we'll do is, uh, we're up on time, unfortunately, so we'll get into that a little bit deeper in the next episode, get a little bit deeper into Swan and what you're doing right there. So I think this is a great point to stop. So where can people find you? Where can they find about more about Swan AI and then we'll take things from there.

Speaker B: Yeah, so I have a newsletter called the Big Shift where it's kind of like, uh, you get a front row seat, uh, behind the scenes of building an autonomous business where I share, you know, the wins and losses, the frameworks and the playbooks that we identify when we try to scale to $30 million. ARR, which is three founders. And also I've built a GPT in ChatGPT called the Autonomous Business OS that has already thousands of conversations. So if you want to learn how to build an autonomous business, you can get just go to that GPT and instead of just talking about it, just build it.

Speaker A: Love it, man. Well, thanks for being on the show, Amos. It was fantastic. You did an awesome job, delivered as expected. So I, uh, appreciate you being on, man, and uh, look forward to seeing you all on the next episode.

Speaker B: Thank you, Ryan, for having me.