Most Founders Are Using AI Agents Wrong (Just Get THIS Right)

Last month, $1,000 of AI infrastructure replaced $27,000 worth of work at my company. But here's what nobody is telling you. The way you use AI at $100K is completely different than how you use it at $1 million or $10 million or even $100 million. Apply the wrong playbook for your stage and you're either wasting money or leaving a massive amount on the table. I'm Eric Siu. I bought a failing marketing agency for $2, turned it into Single Grain, and we work with companies like Amazon, Uber, and Airbnb. I've been early on every major shift. SEO, content marketing, and now AI agents. Today, I'm breaking down the exact AI playbook for each revenue level: $100K, $500K, $1 million, $10 million, and maybe I'll even give you a preview of $100 million. What works at one stage will actually slow you down at another. Operators who get this right in the next 12 months are going to be running laps around everyone else. Let's get into it. So if you're at the $100K level, most people, they tend to just default to using ChatGPT and they're okay with it. At the bare minimum, with $100K, you have to sign up for a subscription of Claude and then add Claude Cowork to the work that you do so you can start to work in a more agentic fashion, meaning that you can actually have it organize your folders for you, you can actually have it send emails for you, you can actually have it help you with spreadsheets or designs and things like that. But because of the power of all these other tools available to you right now, at the bare minimum, you have to be using an agentic type of product like a Claude Cowork. And I think of Claude Cowork as a copilot or a coworker that works with you and actually helps you get things done. So at $500K, it gets a little more complex. So the way you want to think about this is you want to be adoptive with AI. And what that means is you are starting to build agentic workflows. So what that also looks like is in the past, you might have used products like Zapier to help you do different steps in workflow. So if someone sends you an email here, maybe you have an email that auto-replies over here. An agentic workflow is something that actually thinks. So you have different steps in your workflows, but you have an LLM at each every step, and they're actually thinking through, well, is this lead actually qualified? How should we respond to this lead over here? And I would just encourage you to think about creating these agentic workflows to help you get things done and then having them run. Now the cool thing is, in the last section I mentioned Cloud Cowork. Now you've got to graduate to Cloud Code. Don't get scared. Cloud Code doesn't mean you need to know how to code, because I don't know how to code. But the key thing is With Claude Code, you can actually start to run these loops. And what I mean by that is when you go into, let's say, your Claude, there's a section where you can open up code. If you literally just type in /loops— loops are basically cron jobs. Cron jobs are chronological jobs. So it's like setting a timer for things that you want to get done. Okay, so you might say, every Monday, I want you to look at all my top competitors on YouTube. I want you to take a look at what the outlier videos are. Okay, what's actually standing out, what isn't, and look at that skeleton format and then think about what I'm trying to stand for when it comes to my content and give me those skeleton ideas. Literally, that's what I did today. By the way, there's also a by the way command too. So when you're having Claude Code do work for you, if you type in /btw, you can chat with it while it's doing work for you. Now, the reason why I'm bringing Claude Code up right now is because at my company, Single Grain, we're mandating across the entire organization. If you don't know how to use Claude Code, or when we're interviewing you now, if you can't beat Claude, you're in trouble. Okay, so what we're doing is we're saying, hey, Claude Code, the special part about it is it allows you to build products. So you watching this right now, many of you, you probably don't know how to code. I don't know how to code either. I can build whatever product I want to build, and I'm literally going just back and forth with it all the time. And here's the other thing, when you're doing $500K, you can afford a $200 a month Claude Max subscription. And Anthropic themselves, the owners of Claude, they have said that for $200 a month, they're actually subsidizing $5,000 a month in tokens. Okay, so you're actually— you're getting a very good deal on that. You should get in on that. Right now, start using it. Because if you don't start using this as a business owner, as a founder, your entire team's going to be in trouble because you're in effect hurting your entire company by not getting them moving them into the future. That is not something that you can hand off. So whether you're at $100K, you know, $1 million, $10 million, $100 million, you cannot think about handing this off to somebody else and expecting that they're going to figure it out for you and that you can delegate this. All right, so at $1 million, you are now able to take it to the next level. You've got Claude Cowork, you are armed with Claude Code, and you can build stuff like this. What I have on my screen right now, this is my operations center. This is our entire dashboard. We call this Single Brain. And within Single Brain, we can do a host of things. First and foremost, you can see on my screen over here, these are the different agents I have running. So I have Alfred, who's my chief of staff. I have Arrow, who is focused on sales pipeline. I have Cyborg, who is focused on recruiting. I have Flash focused on social media. And I have Oracle focused on SEO intelligence. And you can see These cron jobs over here, you can call these loops as well. All these cron jobs that are running, I can see how many are healthy because these things often break. I can also see what's happening with the calendar over here. I can see what's happening in progress. There's a Kanban board over here, like the roles that we're hiring. All these things are built. There's a memory log here too. And so my point of showing you this is to say that, guys, I built this myself. I don't know how to code. Everything that you think about within your business from a workflow standpoint, you should be thinking about at $1 million. AI is going to replace people. It's more so AI is going to replace processes because everything that has value within your business is a process thing. So for example, I work with my friend Josh right now where we're trying out his company for YouTube long form and short form and things like that. So I'm really judging his company based on his processes, right? And the more they know how to automate their processes and take themselves out of the equation, the more they're proving to me that they know how to leverage this stuff, the more I'm like, right, I'm like, amazing, right? You need to know how to do this first and foremost, you as a founder, Because if you don't know how to do this, you can't show other people how to do it. You can't communicate the vision. Okay, so I'm going to go a little deeper into this too. Recruiting is a big deal for us over here at Single Grain right now. We want to make sure that we're hiring very AI-forward people. We want to hire people who are humble, hungry, smart. And to fulfill this vision, we got to handle recruiting. And so here's the thing. Literally, this is our recruiting bot over here within Single Grain. And this is where we almost have it working right now. I'm paying for a handful of other tools, but literally you can see the pipeline of what we have over here, like who we're reaching out to, and you can see what's happening over here. You can see the approval queue. Do we approve these people over here? What's the justification for these people? This person's a VP. You know, what level are they at? The LinkedIn profiles and all these things. And then there's a CRM. I mean, I can set this to, you know, I can see if we're behind, on track. I can see how all these sequences are doing over here. All these things I would be paying for, and we are paying for another recruiting tool right now. I think we're probably paying $1,000 or something like that, but this eliminates the need for that. Not only that, this single brain combines everything. It has an entire memory of all the things we're doing, all of our goals, if we're on track, we're off track. And again, with this, with this dashboard, now we're able to manage these agents end to end. So that's how you got to think about this stuff, right? When you know how to use Claude Code, you know how to use Claude Code, you start building this stuff. And then more importantly, you get to the next level where you think about automating processes. And when you build out these dashboards, it's to manage these agents, not for humans to manage these dashboards. I see a lot of people at my company right now when we're going through these hackathons, they're like, oh yeah, they're just building dashboards and they look cool and they're building new reporting tools, but that doesn't matter because they're building stuff for humans. You want to think about at $1 million, you got to think about hiring humans to do human work. Don't hire humans to do robot work. That's what a lot of humans are making mistake with right now. And by the way, this is, guys, this is still very fresh right now. We're talking about like a couple months or so. And so if you're ahead of this right now, if you're ahead by a year or so, I actually think you're ahead by 3 or 4 years. So now is the time to get in on this right now. That's a million bucks. Let's go over my agent dashboard here. So you look at the team, this is me up here and Alfred, my chief of staff. The way Alfred works, by the way, I probably would be paying a chief of staff anywhere from, depending on experience, we're talking about $150 grand to $250 grand a year. Okay. So my chief of staff knows everything about our goals at the company. It has the most connectors. So it's connected to our CRM, it's connected to Granola Notes, it's connected to Gong, it's connected to all these different data sources, it's connected to our analytics, it's just whatever I ask it, it's connected to the YouTube API too, so the information I queried this morning for my YouTube team, I actually pulled it directly from there. I asked it questions in natural language and it popped it out inside of Slack and then I just forwarded it over to the team and that's now allowing us to rotate a lot faster. Not only that, it's connected to, for our SaaS products, for Carrot or for ClickFlow, Within there, we can ask questions around our Stripe profile or our Mixpanel profile saying, hey, where are we seeing the most churn right now? Where's the most drop-off? What do we need to fix exactly? What is the biggest bottleneck? What do you recommend? So it's not only doing data polls, it's actually strategizing with my team in public. And it's also asking tough questions to people in public too. Because by the way, people don't like being held accountable, right? You think about the last time you had to hold someone accountable, they didn't like that, right? But when you have a robot holding people accountable, That's a little different because a robot doesn't really have emotions. So that's Alfred. So Alfred has the most context, but Arrow, Cyborg, Flash, Oracle, these are separate agents and they have their own context. I'll give you an example here. Oracle is focused on SEO, and because of that, it has access to the Ahrefs API, which is like an SEO tool. It has access to Google Analytics, Google Search Console. These are all SEO tools, right? It doesn't have access to all these other things. It has access to the YouTube API, but it doesn't have access to the CRM. Mm-hmm. Gong because it's just focused on driving more SEO traffic or AEO traffic. Same thing with Flash. Flash is only focused on a couple areas, right? So you want these agents to be focused on their own things. And that's what I would say. Like, by the way, at $1 million, you have multiple agents running for you and these are autonomous agents. And I am using something here that I'll probably talk about for the next level. So when you get to the next level, I'm going to talk about these autonomous agents that you can be running and how I think this is the future of everything. Quick break. If you're doing AI SEO or content marketing in general, you know what the problem is. The problem is you're spending hours writing yourself, or you're paying agencies lots and lots of money, and you're getting AI slop that sounds just like everyone else. That's why we built ClickFlow. ClickFlow creates production-grade content that actually sounds like you. Plus, it helps you with other areas such as building internal links, creating FAQs, and more. One customer even reported saving 90+ hours a month just by using ClickFlow. You can get a 14-day free trial just by going to clickflow.com. No more AI slop, just production-grade content that drives traffic and conversions. Back to the show. So at $10 million, here's what I would do differently. At this point, you want to think about having your AI agents collaborating with your team, invading your Microsoft Teams or your Slack channels. Okay, so in this case, on my screen over here, I'm working with Alfred here, and it shows me a leaderboard here on how autonomous our agents are right now. So this is one thing. So if you have your agents working on tasks for you, such as email marketing, like cold email, or maybe it's helping you with recruiting, what happens is you want to think about this as a calibration meter, almost like a level bar. Okay? So if you played any video games, you have to be constantly feeding context to your agents. Otherwise, that context is going to decay over time. Maybe your product or your service changes. Maybe someone in your company leaves, or maybe your messaging has changed, or maybe you have a different offer. So you're constantly looking to feed it. More context. And I can see here how my agents are doing with certain things. And so for cold email, we have a fully autonomous agent where our team has been very consistent with approving cold email prospects, right? Because once you approve a few of these and it looks good, then you can basically let the agent kind of do its own thing more or less, right? So there's a trust meter that you need to build with the agent. You can't just let it YOLO initially, right? I think the challenge is when you're at $1 million or $500K or even less than that, people aren't thinking about that. You're just thinking about using AI for yourself. But when you're using it within your team and these AI agents have invaded your Slack channels, you got to think about how you can actually have it collaborate. You have to think about permissioning as well. You have to think about security as well, because now it's working with your team as collaborating and you don't want it to— for example, you don't want to just give it all of your client data and give it read and write access, meaning that it can read the data and it can also change information, because if they delete your your client's ad accounts, or they delete the entire, you know, email accounts, who's on the hook for that? You're on the hook for that, right? So we don't do that on the Single Grain side. And I will talk about how to get to $100 million right after this, what you should be doing there. But, you know, this is an example of us adjusting our email list. So here on the very right side, I basically, the way I work with it now, I didn't write all this, by the way, I dictated all this. I use Whisper Flow. And I use it on my phone quite a bit too. So I just start talking. Sometimes I'll yap for like 5 minutes or so and I give it context. So I said, here, look, I want you to take a look at our newsletter, the Leveling Up newsletter. Do you think you can make it significantly better? Also considering the fact that you have the content ingester, not in gesture, it'll, you know, take it in wrong. And you know what I'm talking about already and you kind of know what platforms already, what performs already. You have the Metricool data. So anyway, it'll misspell things, but I'm just asking questions here. I was like, hey, let's figure out how to make the newsletter better. I want to make it the best newsletter and we want to promote the right things. And you know, we can plan this out together and you can ask me clarifying questions. This part is important because the agent, oftentimes you're going to be saying a lot of things and it doesn't— if you don't ask it to ask you clarifying questions, it's just going to assume that that is all the context that you're going to give it. But if you say, hey, let's slow down, let's plan this out for a second, and why don't you ask me clarifying questions before you go, your output is going to be way better. And not a lot of people think about that, right? And by the way, when you're at $10 million, there's a bunch of other things that you should be doing within Claude Code, right? There's a bunch of plugins you could be adding. There's a bunch of skills that you can be adding to, to kind of level up what you're doing. And you can have skills at the very first level of Claude too, but when you have skills, combining and you're having these agents call on these skills whenever and you're running these loops. Or if you're running like an autonomous agent, like what I'm using over here with one of the CLAAs, you can just have it running at 10 AM every day and then you can just audit these, right? But it's very important for you to continue to upgrade the scaffolding that you have, almost like you're upgrading your business, right? Because at $10 million, you've got a management layer, you're trying to build something very serious. It is important that you take all these things into account— governance, security, all these things that you wouldn't think about at $1 million or $500K or so. So anyway, I digress. Let me come back to this for a second. Alvin starts talking to me. Okay, so look, it does a little hat over here when it reacts. All right, I read through the newsletter archive and the recent issues. Honest assessment, this is good. Okay, so here's what's not working. Voice inconsistency. The PS section is a junk drawer. No data from your system. No recurring structure. Recent issues are OpenCLaw heavy. Okay, this is all good feedback. I'm going to bring you to the bottom of kind of where we're getting with this. I'm not going to bore you with all the details in the middle, but If you look at the very bottom over here, okay, I started working with it. I'm like, hey man, you gave me 2 ideas over here in terms of what we can do to make our newsletter better. So idea A is we go with the podcast-first model, okay? So the newsletter that we have will extract from the sharpest atoms from my 5 to 7 weekly episodes from Leveling Up in Marketing School, which is the podcast or the YouTubes that I have. An atom is basically a key talking point, okay? Because we have a content ingester, okay? Again, that's like a little cron job that we have running, a content ingester that will pull from my content every single week and I'll pull the key points. And so again, you want to make a list at $10 million, you want to make a list of all the things that you're doing that are recurring, you want to look at the time tracking of your entire team. If you're like an agency like mine, we have a time tracking tool called Timely. I just pulled from the API yesterday, hey, where are we wasting the most time right now? What should we be automating exactly? That's what you should be thinking about, how you can unbottleneck multiple areas of your team. And that's how you start to scale to 10 million, to 20 million, to 30 million, to 40 million, eventually to 100 million or so. Okay, so back on this, the original idea, okay, which actually, I don't know if this was my idea, but it claims it was my idea, was like, oh, we should have a news and content hybrid for our newsletter. The trouble with that is that there's already a lot of newsletters like that, right? I was like, okay, well, that idea is not that good. Like, even the 80 out of 100 is not that good. Like, how do we make idea A a 90 out of 100? Then it's like, okay, well, you need to level up the voice. You need to make sure that you have data in whatever subject line that you have, because that's clearly what works. And, you know, we need to have call to actions in your podcast or your YouTube channels and promote it more, right? I'm like, okay, that's pretty good. When it gave me an example over here, I'm going to pull this example up for you guys in a second. This is what I was able to do after I went back and forth with it. I was like, hey, I want you to send me some examples of how this new newsletter might look. And bear in mind, like, this is just me dictating back and forth. Okay, so sample A, Monday. Subject is 95% of AI citations come from content under 10 months old. Preview, your 2023 blog posts are invisible to ChatGPT. Okay, so I put out 5 episodes this week. These 3 things are worth your time. Okay, your old content doesn't exist to AI. A16Z got tired of NYT, New York Times controlling their narrative. And then this one over here, your AI agent is stuck at 6 out of 10 because only one person is using it. Okay, and then it recommends an experiment. Here's a playbook every week. I'm like, dude, and then also, by the way, it has the URLs plus the UTM tags. I asked it to include the UTM tags in there as well. So this is pretty good. This is not bad. Like, I'll tell you what, this is better than our current process right now. This looks better. Like, I'm proud. I would be proud to share something like this. So those of you right now watching this, like, yes, this is a way for you to upgrade your newsletter, but zoom out. All the stupid processes that you have going on right now where you have humans doing robot work, you doing robot work, you should get angry. At $10 million, you should get angry that you're still doing that stuff. You should be the one driving AI fluency throughout your entire organization. Okay, everyone, we're going to do a hackathon. Okay, you know, 2 to 3 days, maybe 7 days every quarter or so. We're going to shut down the company. No client calls, no sales calls, none of that. We're going to focus on learning every week. Weekly office hours mandatory. Okay, there's a show and tell. Your performance reviews are going to be tied to your fluency with AI. Show us what you built. You as a leader, you telling that your managers too, every single week, maybe every other week or so, Asking the people that they're working with, what have you automated this week? What have you automated this week? What have you automated this week? Over and over and over. Because what happens is, if you don't apply the pressure, you will feel the pressure. And I guarantee you, like, sure, you can bring people, you can put the water right there, right? You can bring many horses there, but you can't make everyone drink. And it's on you as a leader to drive the energy, because if you don't, everything's going to fall apart, guaranteed. So the types of people that you should be hiring in this new AI world, especially at $10 million, You need to hire people who have a good relationship with change. And what I mean by that is it doesn't matter if you're 15 years old, it doesn't matter if you're 65 years old. Anybody that's very adaptable and is very open to trying new things. Like I've seen 50-year-olds that are really on top of the AI stuff. I also see people that are maybe 18 years old or so that don't want to touch AI. They're very against it. Oh my God, the electricity bill is too high. The water bill is too high, right? The water thing's a non-issue. Like it's just people are just looking for things to kind of, you know, point fingers at. So you want people who are adaptable, who look forward to the future, and who are very comfortable with being uncomfortable. And to be more specific here, at Single Grain, we are hiring people that have high agency, that have a strong bias to action, who love to learn. And by the way, the easiest way for me to measure that now is I just look to have an exceptional conversation with an exceptional person. Within that 30 minutes, I just want to talk to them about AI because AI covers all 4 of our core values. You got to know your 4 core values, right? Like, or whatever they are for you. Our core values are simple. We want people who are relentless learners. Okay, that means you cannot help yourself. You must learn. Okay, number two, you have a strong bias to action. Number three, you make bold long-term bets. And number four, you are unreasonably resilient, meaning that you can't even take one punch in the face. You take like twenty punches in the face and you still get back up. Okay, when you look at those four values, I'm like, oh wow, like all of that's within AI. If you have bias to action, you're strong with AI right now. If you're a relentless learner, you're all in on AI right now. If you're making bold long-term bets, AI right now. If you're unreasonably resilient, like when you're learning in cloud code, oh my god, it is. Painful. So many people give up, right? To give you something that's more actionable that you all can kind of steal from here, I'm going to show you our Beat Claude Challenge. Okay, so if you go to singlegrain.com/apply— again, singlegrain.com/apply— it will redirect you to our Beat Claude Challenge. And what this really is, is this is the hiring challenge for AI-native professionals. So we're looking for strategists who can think better than AI. Okay, and we have all these challenges in here. So AI can do 80% of knowledge work now. We need people that can do the 20% that AI can't. So creative leaps, pattern recognition, judgment under uncertainty. And so if you can't beat Claude on a strategy challenge, you're competing with a $20 a month tool. We're looking for people who make AI look basic. And so I'm not going to go too far into this, but you can see these are all the different roles that we're hiring for. And that's not to say we're going to hire like a paid media person immediately, but we always want to keep the bench building, right? And we found this to be a very good filter for us because the filter is basically Hey, you can use AI all you want, and beat Claude means like you're going to be competing with AI. And if you tie with AI, that means ties don't get to move forward. Like if you actually beat Claude, then you actually get to move forward to the interview round. So I'll show you one example of this for like a marketing strategy role. And this is beautiful because it saves us so much time. So you click on this, it's like, oh, this is go-to-market strategy for an AI customer support tool. So this is someone who's looking for like a marketing leader type role. Okay, so the situation is you're advising a Series A startup that has built an AI-powered customer support tool. Here's what you know. Here's the company snapshot, revenue, how much money raised and all that. Here's the product, here's the traction. And it's like, here, create a 90-day go-to-market strategy to get them from $0 to $50K in monthly recurring revenue. Here's what to submit: channel prioritization, specific campaigns, pricing, and 90-day milestones. Here's the constraints over here. Here's how we score. So we keep the scoring rubric very open, and then maximum 4 pages. Spend no more than 1 to 2 hours on this. Now, some people might complain, oh yeah, well, you know, you should be paying to do this work and you should be paying us to do this work in the beginning. It's like, no, if you're gung-ho, like, it doesn't matter. It does such a good job of filtering people out. And the people that, like, I've sent this to a few people on the team too, and they can't beat AI and they're offended by it, right? But this is just what the reality is. And if people don't want to confront reality, that's too bad. But for you as a leader of a $10 million company, you have to set the constraints, the boundaries, on what it is that you're looking for and what it is that you are willing to accept. And for us, we are not willing to accept anybody who can't beat Claude. So here's where this is all going. If you're trying to transition up to $100 million, even a billion or so, here's how you want to think about this, or here's how we're thinking about it, at least at Single Grain. But it's actually Single Brain. Okay, so we have Single Grain, but we also have Single Brain. Now, many of you have probably heard of the company Palantir. Obviously, we're talking— we've talked a lot about Anthropic, which are the creators of Claude. So I'm going to give you 2 kind of examples. Palantir is a software company and everyone's like, how do they make so much money? What do they do exactly? Well, I'll tell you one model they have that works. It's known as the forward-deployed engineer. So what they do is they have Foundry, which is their software that works. They customize their software for every single client. The forward-deployed engineer goes in and they embed themselves into the company. Okay, so you have an embedded engineer and they build everything custom according to that company. It could be a government organization or something else. Anthropic also was doing the same thing with the US government before they kind of, you know, ended their relationship. Now, the reason US government and Anthropic, they had a spat for a while, was because the government was actually trying to negotiate with them because Anthropic had 4 deployed engineers within all these government organizations and they were entangled. Like everything was entangled into their— the builds were entangled into government systems. So they actually wanted to work a deal out. The 4 deployed humans, that's how I look at this, is this is the future of it. Okay, so you're going to have, you know, the thing I talked about earlier, like I was showing you an example of my agents, you're going to have different CLAAs that you're using. So in this case, I'm using OpenCLAA on my own stuff. So you have, you graduate from like just standard Claude or ChatGPT to Claude Cowork, maybe you're using some Claude Skills in there, then you're moving over into Claude Code, and then you're graduating into autonomous agents like OpenCLAA, right? What I think is going to happen, you're still going to need humans very much. And I would say that for us, we are very focused on this 4-deployed model where we have very strong marketing strategists on one side and we have very strong AI engineers and AI automation engineers on the other side. And that's what we're doing. That's how our team's going to be organized. And the way I see this over here is like, look, you know, Palantir, Anthropic, Playbooks deploy engineers and build from the inside, starts from zero each time. So you want to make sure that you're building all these copilots, these agents that you've built for yourself and you're customizing for your customers. And that's what we're doing, right? That's how we're piloting this for our customers right now on the single grain side. Like a lot of people are reaching out, hey, can you help us with enterprise revenue agents? Can you help us with recruiting agents? We're doing all these things internally. We're eating our own dog food, if that's the expression, right? And then we want to eventually build these custom revenue engines that have embedded intelligence, shared context, and that way you have a moat and a high LTV. So when you have— when it's not just software, it's strong software with smart human beings, that's how you're going to capture the future of software in general. I think that's how the game is going to be played. Now, if you want to see me go more in depth on the latest AI tools and what's changing in the markets right now, you'll love this video.