Transcription
[Music]
This AI business model will create millionaires or even billionaires in 2025. It could be bigger, or it could even completely replace SaaS, the best business model we've had so far, which is software as a service—a market of more than $300 billion.
In this video, I will explain exactly what this new model is, share all you need to know to get started, even if you're a complete beginner, and even provide you with a complete step-by-step roadmap you can follow to profit from this opportunity before everyone else.
At the end, we will even do an interview with someone who has already gone through all of these steps and currently generates $50,000 per month with this new amazing model.
Before we dive in, if you're new to this channel, welcome! My name is Arseny, and I've been building AI solutions for the last four years. We were one of the first AI agencies to start building AI agents for businesses worldwide using my own AI agent framework, with over 3K stars on GitHub.
Now, unlike all other agency owners on YouTube, I won't even be selling you a course or a paid community at the end, and it's not even in my plans for 2025. We'll be doing exactly what I teach you here in this video ourselves.
Lastly, for all my long-term subscribers, sorry for the recent slowdown in content. I'm currently in the process of relocating to Dubai, which is why I'm recording at this beautiful podcast studio.
Now, let's dive right in. First, let's talk about what this new opportunity is and how exactly it could replace SaaS. Some of the ideas in this section for this video were inspired by a Y Combinator podcast episode, which I'll link down below for you to check out later.
The key idea from that episode was that just recently, back in the early 2000s, all software was sold on-premise. There was no SaaS model because all software was sold on CDs. You would buy a CD, install it, and this is how you'd get your software.
Then, when the XML HTTP request, or Ajax, came out, it finally became possible to build interactive software directly in your browser, which is exactly what allowed for this new SaaS model.
Today, we're experiencing an even bigger shift because, with the rise of affordable and smart LLMs, software is becoming more interactive than ever before. You can now interact with your software as naturally as you would with any other human. There's no longer even a need for a browser anymore.
Check out my previous video on GPT-4 if you want to learn more about that. But in a nutshell, what LLMs truly unlock for the world in 2025 and beyond is a completely new way to interact with technology, which is exactly what enables this revolutionary new AI business model called "Agents as a Service."
AI agents are so much more powerful than SaaS because AI agents do exactly what SaaS does, but they do not require any manpower to run. For instance, HubSpot is just a CRM, but you still need a marketing person to use it.
On the other hand, if you buy a marketing agent, you don't even need anyone else because the agent will use your CRM for you. It will generate leads, schedule appointments, follow up with them, and update that CRM for you.
This is why I believe Agents as a Service could actually be even bigger than this $300 billion software as a service market. Agents do not just give you the tools to automate a specific process; they actually automate the entire process for you, so it's infinitely more scalable.
Another important point I want to mention here is that it's actually much easier to build agents than traditional software products, unlike many people think. You can trust me on this because before we started our Agents as a Service subscription, we actually built four large-scale SaaS products for ourselves and for our clients.
You see, at its core, all SaaS products are essentially just wrappers around databases. HubSpot, Salesforce, QuickBooks, Zendesk—it doesn't matter. All of them are just user interfaces built on top of differently structured databases.
So, Zendesk is just a pretty-fied database of your customer support tickets, while QuickBooks is a database of your transactions. The beauty of AI agents is that they can work with these databases directly.
There's way less code required because the agent can just hook up to your database without you having to build any super complex APIs or user interfaces. So, not only do agents provide significantly more value for businesses than SaaS, but they also require way less effort to build.
This is why I believe this is such a massive opportunity for anyone watching this. And to be clear, you are still extremely early in this huge $300 billion market. In 2024, enterprises only spent $1.22 billion on vertical agents.
This is the type of agents, by the way, that we will be discussing in this video. Don't worry if you don't understand the term; in the next section, I will explain everything you need to know about vertical versus horizontal agents.
The key thing to notice here is that the growth has actually been the highest among all other categories—12x from the previous year—and the use cases are still extremely basic. It's pretty much just co-pilots, rack, and support chatbots.
Okay, now before I can present you with this complete roadmap you can follow right now to fully benefit from this opportunity, we first need to understand the difference between the two types of AI agents that you can build under this new Agents as a Service model, which are the horizontal agents and vertical agents.
The biggest difference is that vertical agents are very niched down. They are made to perform only a specific single role or function. On the other hand, horizontal agents are not restricted by any specific use case.
When you are building a horizontal solution, it means that you or anyone else can adapt it for any niche role or function. For example, Harvey AI is a vertical agent made specifically for law firms, while my agent framework, Agency Form, although it can be used to make vertical agents, is a horizontal platform.
Now, what this means from a technical perspective is that vertical agents can be pre-trained for you, while horizontal agents require training from scratch. You see, a key insight that we discovered in our agency is that there are no two agents that are exactly the same.
However, although most companies follow completely different sets of steps in their own processes, a general process for each role is still there. For example, a good developer always typically first looks at a task, then finds the relevant files, then writes a test, then writes the code, and then tests it until it works as expected, regardless of which company he works for.
This is exactly what vertical agents are trained on. This pre-training on a general process for a specific role is what allows those agents to get up and running much faster.
However, just as this can be a blessing, it can also be a curse. You see, the fact that those vertical agents have already been trained on a general process actually makes them harder to customize for a specific process in your or your client's business.
For example, if your developers also need to write documentation for each feature, but that documentation has to be placed in a completely different repo following a specific format, this is where vertical agents can struggle.
On the other hand, horizontal agents are not restricted by this initial pre-training, so they can easily be customized for any process.
So, to fully grasp the differences, let's try to plot this on a graph. On the x-axis, we have the performance, while on the y-axis, we have the time or effort required.
This is what it looks like for vertical agents: you get a very steep return on investment in a very short amount of time because you don't have to train it from scratch. But then, as you can see, typically the performance plateaus as you've exhausted all its capabilities and you can't customize it any further.
And this is what it looks like for horizontal agents: with horizontal agents, you need to invest a considerable amount of time and effort before you get any value at all. But then what typically happens further down the line is that you are able to achieve a higher return on investment because you are not constrained by this initial pre-training, and you can customize it much further.
The problem is that most business owners have no idea how to do this, nor any desire to do it themselves. If they can get to 80% performance with 20% effort, they will just go with the vertical solution.
This is why, with vertical agents, you can get to extremely high revenue numbers in a very short amount of time, just like with B2B SaaS compared to B2C SaaS. It's much easier to grow at those initial stages.
Vertical agents are much easier to pitch and build yourself because you are only targeting a very specific type of business customer. This is why I don't even recommend you to consider building a horizontal platform for maybe 99% of people watching this video.
And this is why we ourselves, at our agency, will be expanding our offerings with vertical agents in 2025.
Okay, now let's take a look at a few real-world examples of vertical agents. First, we have a platform called 11x, which is building an AI SDR and sales rep agents. They are automating your go-to market and replacing platforms like Salesforce and Apollo, but again, not just giving you the tools, rather automating the entire process from generating leads to scheduling appointments and closing.
They have just raised a $50 million Series B round at around a $350 million valuation. In total, they raised over $74 million. Honestly, I think $74 million is a total overkill. It's not like they're training their own large language models; they're still just hooking up to OpenAI's API, just like everyone else does.
So don't bother with those crazy investment rounds. In my roadmap, you don't actually need any funding to get started at all.
The next example is Carmen. This is a more niched-down agent made specifically for construction project managers to help them automate administrative tasks.
I personally prefer more niched-down use cases like this, where it's not just an agent for a specific role, but it's an agent for a specific role in a specific industry or for a specific niche. I think these are much easier to pitch and can provide more value for your target customer.
After all, Normi is another example of such an agent for regulatory compliance teams. It helps them evaluate whether proposed content or actions are compliant with relevant regulations.
And lastly, we have the famous Devon, which is a $500-a-month development agent. Yeah, I don't even know what to say about this. Let me know, guys, if you want me to do a full review and compare it with some of the horizontal agents that we made in our agency.
By the way, it surprisingly works in Slack. Just as I explained earlier, you don't really need a UI to run an agent.
All right, now let's talk about what you'll actually need to build a vertical agent yourself. The first key component, which you will most definitely need to focus on, is data.
In any AI project, the quality of your input data will ultimately determine the quality of your output. Recently, people started to forget that AI agents are also first AI models.
So, to build an effective vertical agent, the first thing you need to focus on is collecting valuable internal data. This data will later either be used by you for training, evaluation, or for fine-tuning.
Note that the data you collect has to be internal because if it's not, and if it's valuable, most likely ChatGPT has already been trained on it. The real magic happens when you train agents on this most precious internal data that all companies keep secret.
However, if you don't have access to such data, it's not a big problem because in my roadmap, I will actually explain how you can collect it as you go.
The next key component for building vertical agents is industry-specific expertise. To build an effective vertical agent, you need to know your customer and their SOPs, or standard operating procedures, really well.
From those SOPs, you will then extract the general process that the agent must follow. If you don't have industry-specific expertise, I definitely recommend partnering with someone who does.
And lastly, you need some resources. It does not necessarily have to be funding. You can raise money if you want to, or you can apply to a startup accelerator. It seems like this is not a bad time to do so.
However, don't build your product only to raise funding. This is the biggest mistake you can make. Focus on actually delivering value with whatever resources you have. You can easily get started completely by yourself by investing only your own time and effort.
Okay, now that we've gathered these resources, you can choose one of the three methods to start building your own vertical AI agent. The first one is using a framework, the second one is leveraging a platform, and the third one is developing a fully custom-coded solution.
So, let's now explore the pros and cons of each approach, and then, as I promised, I will provide you with a roadmap.
First, as I said, you can build it using a framework. This means essentially using an agentic framework like Agency Form, Crew AI, Link Chain, Autogen, and others, which handle some of those lower-level details for you.
The advantage of this approach is that it saves a significant amount of time and effort during development. The downside is that you will still need some development experience.
The next way to build a vertical agent is through a horizontal platform. Yes, you can actually build a vertical agent on top of another or even multiple horizontal platforms like Google Cloud, Vertic AI Agent Builder, and others.
This is the approach that the person we will be interviewing at the end has selected, by the way, and it's pretty impressive how far you can take it. The benefits of this approach are scalability, meaning that you can serve as many end users as you want without managing the servers and without having that significant technical experience.
However, one major drawback of this approach is the costs because most of these platforms can get extremely expensive at scale.
The final way to build a vertical agent is through your own fully custom-coded solution. This is where you're not using a framework or another horizontal platform and just building everything from scratch yourself.
With this approach, you obviously have complete control over everything, and you're not constrained by any other systems, but you obviously need significantly more technical experience and development effort.
I generally recommend starting with a framework or with a horizontal platform and then potentially transitioning to a completely custom-coded solution later when you are ready to scale.
By the way, I have plenty of tutorials on how to build agents with my framework and even free cloud deployment templates on my channel that I will also link for you down below.
The final thing before you can start building your agent that you need to consider, and that can easily make or break your success, is how to price your agent.
So here are the four key AAAS pricing models that I've discovered so far. The first one is licensing. This is where clients pay either for a one-time setup fee or a monthly subscription fee to use your agent.
This model is easier to pitch since clients know exactly what to expect up front. However, it doesn't account for the actual value generated or usage of the agent, which means you risk either significantly underpricing or overpricing your solution.
I recommend this model when you are just starting out so you can get some cash flow, case studies, and valuable feedback.
The second pricing model is usage-based. This is when you charge customers based on their token or message consumption. For example, you can charge double or even triple OpenAI's token rate or set a fixed rate like 20 cents per message.
This is already a much more scalable model, but you need to be sure that your agent will be constantly used, preferably even by multiple employees at the same time. That's why I believe this model is best for enterprises.
The next model is outcome-based pricing. This is where it gets interesting. With some vertical agents, you can actually charge per result—for example, per appointment booked, per lead generated, or even per website built.
That's how many SMMA and other types of agencies are able to generate such massive revenues because the clients can easily evaluate if your pricing makes sense compared to the results that they get. You can charge them significantly higher rates. I recommend this model for SMBs or small medium-sized businesses.
Finally, there is one more model that we will be personally experimenting with in 2025, which is the hybrid approach. With the hybrid model, you can combine multiple pricing strategies together.
For example, you can charge a base fee of $3,000 per agent per month that includes a thousand messages and then charge an extra $50 for each lead generated or, I don't know, for maybe 500 other messages.
This approach lets you get the benefits of different pricing models; however, it does make your agent harder to pitch because it's harder to estimate the long-term costs.
Now that we've covered the fundamentals, let me finally present you with the complete roadmap to fully leverage this opportunity in 2025.
The first step is finding your niche. Ask yourself which industries you know best. Do you have any unique market insights, or do you know anyone else who could provide those insights? If so, this is the industry that you need to be in because, as I said before, understanding your target customer and their problems is key for vertical AI solutions.
The next step is to identify a suitable problem for an AI agent to solve. Keep in mind that not every problem should be solved by AI agents. You need to find the recurring problem that appears repeatedly across the entire industry and that companies have struggled to automate before.
It needs to be a dynamic process that traditional automation tools like Make or Zapier can't solve.
The next step is to sell. This is the approach I always recommend. You shouldn't build your vertical agent before selling it first. You need to find the client, and only then you build the solution specifically for that client.
This helps you to reduce any upfront risk and eliminate any upfront investment because your clients will essentially finance your vertical agent solution. This is why licensing works so well at the start.
After finding a client, the next step is to build an MVP, a minimum viable product tailored for that specific customer. This is the step, by the way, where you can collect the necessary data if you haven't had access to it before.
Simply ask your clients for their SOPs and any internal knowledge required to train the agent. However, obviously, make sure to be clear about how you're planning to use it later and make sure to anonymize it if needed.
Don't worry about your solution being not reusable across multiple clients at this stage. The key right now is gathering feedback and understanding how your agent differs between those initial clients.
After building solutions for a few clients, you'll start noticing patterns and similarities between them. You'll see which features remain consistent and which features need customization.
At this point, you can begin productizing your agent. So, at this stage, you need to identify which components remain constant and make sure that everything else can easily be modified from a single config file or a template, depending on which approach you selected.
On a high level, all vertical agents are just pre-trained models combined with templates that you can easily customize for different clients.
So, for example, if some of your clients use QuickBooks while others use Xero, ensure that you can easily adjust your agent's tools and switch between those two platforms.
The same applies to prompts. Make sure that you only change certain sections in your prompt and not rewrite every prompt every single time.
After that, the next step is to evaluate. This means setting up the evals. Evals are a complex topic, and I'll probably be doing a completely separate video on that later, but tracking your agent's performance at this stage is crucial because this is what's going to allow you, or maybe the agent itself in the future, to improve its performance.
With proper evals, you can completely destroy all of your competition because the more agents you deliver, the better they will become.
And the final step is scaling. So, increase your marketing spend, hire more people if needed, and try to deliver as many of those agents as possible. If you priced your agent well, I hope to see you in Dubai here soon.
Now, let's talk to someone who has actually gone through all of these steps and successfully scaled their vertical AI voice solution to over $50,000 per month.
After this interview, I will also share my personal thoughts on the future of vertical versus horizontal agents because, for some of my long-term viewers, you might be wondering why I am building a horizontal platform after all.
Okay, now let's get to the interview.
All right, so welcome, Chase. Chase here built one of the most impressive vertical AI voice solutions that I've personally seen so far.
So, Chase, can you give us a brief overview of what you've built?
Yeah, what we basically did was we integrated the old good old technology that everyone's been using for a long time with CRMs and automations, and we married that into real authentic conversational AI.
So, voice AI, conversational AI—we married all those things together. We really focus on everything after a lead comes in until that lead leaves the company.
We can handle everything almost solely. When they fill out a Facebook form, our AI agent can call them immediately, text them immediately. The voices are absolutely amazing. They can take all the information, update the CRM, book appointments, reschedule, and update the names and emails and addresses in the CRM live on the call.
Then we have another AI that will read the transcription, dictate what happens, update the CRM, and then they get automatically moved into the right bucket.
So if it was like a not interested, then it would automatically tag them as not interested, move them into the not interested pipeline, and it would remove them and put them on the do not call list—all without any human intervention.
So just as I said, vertical agents don’t just give you the tools; they don’t just give you the CRM; they actually use your CRM for you. This is what makes them so freaking scalable.
So when you were getting started, how did you initially find this idea?
That's a great question. So that's a solar panel. Everybody thinks that's a curtain; that's a solar panel. I've helped about 600 people go solar, and I was already really deep into running my own team.
So we had four people on the phones, and we used an awesome CRM. Everybody and their mother was trying to copy us and chase us for almost five years with what I thought were basic automations.
But every time I checked out a marketing agency, they're