Vertical AI Agents Could Be 10X Bigger Than SaaS

Every three months, things have been getting progressively better. Now we are at a point where we are discussing full-on vertical AI agents that will replace entire teams, functions, and even enterprises. This progression is still mind-blowing to me. Many of the foundational models are now coming head-to-head. There used to be only one player in town: OpenAI.

But we've been seeing in the last batch that this has been changing. Thank God. It's like competition is, you know, the soil for a very fertile market ecosystem where consumers will have choices and founders have a shot. And that's the world I want to live in.

Welcome to another episode of The Light Cone. I'm Gary. This is Jared, Harge, and Diana. Collectively, we have funded hundreds of billions of dollars worth of startups right when they were just one or two people starting out. Today, Jared is a man on fire, and he's going to talk about vertical AI.

Yes, I am. I am fired up about this because I think people, especially startup founders—particularly young ones—are not fully appreciating just how significant vertical AI agents are going to be. It's not a new idea; some people are discussing vertical AI agents. We have funded several of them, but I believe the world has not yet recognized just how massive this will become.

And so I'm going to make the case for why I think there are going to be $300 billion-plus companies started just in this one category. Nice. I'm going to do it by analogy with SaaS. I believe that, in a similar fashion, people don't understand just how big SaaS is because most startup founders, especially young ones, tend to see the startup industry through the lens of the products they use as consumers. As a consumer, you don't tend to use that many SaaS tools because they are mostly built for businesses.

Companies. I think a lot of people have missed the basic point that if you look at what Silicon Valley has been funding for the last 20 years, we have mostly been producing SaaS companies. That's literally been the majority of what has emerged from Silicon Valley. Over 40% of all venture capital dollars during that time period went to SaaS companies, and we produced over 300 SaaS unicorns in that 20-year span, which is significantly more than every other category.

Pretty awesome software is indeed pretty awesome. I was reflecting on the history of this because we always like to discuss how the history of technology informs the future. The real catalyst for the SaaS boom was, do you guys remember XML HTTP Request? Oh my god! I’d argue that it was quite literally the catalyst for the SaaS boom—like Ajax. Yeah. In 2004, browsers added this JavaScript function, XML HTTP Request, which was the missing piece that enabled the creation of rich internet applications in a web browser.

For the first time, you could create things on websites that resembled desktop applications. This innovation led to the development of Google Maps and Gmail, sparking a significant SaaS boom. Essentially, the key technology at play was the transition of software from being something you installed from a CD-ROM on your desktop to a service you accessed through a website and on your phone. Paul Graham is also part of this lineage, as he was one of the first people to realize that he could connect an HTTP request to a Unix prompt.

And you didn't actually have to have a separate computer program to change a website. So Viaweb was an online store, kind of like Shopify, but way back in the day. Yeah, it was basically the first SaaS app ever. Paul Graham actually invented SaaS in around 1995. It's just that those first SaaS apps kind of sucked because they didn't have XML HTTP requests.

Every time you would like to click a button, you would have to reload the whole page. It's just a frustrating experience. It didn't really catch on until 2005 when XMLHttpRequest was introduced. Anyway, I see this LLM thing as being very similar. It's like a new computing paradigm that makes it possible to do something fundamentally different.

In 2005, when cloud and mobile technologies finally took off, there was a significant open question: what should we do with this new technology? Where would the value accrue, and what were the promising opportunities for startups? I began reviewing the list of all the billion-dollar companies that had been created.

And I kind of had this realization that you could categorize the different paths that people took into three buckets. There's the first bucket that people started with, which I would call obviously good ideas. These could be mass consumer products, such as documents, photos, emails, calendars, and chat—things that we used to do on our desktops but that could obviously be moved to the browser and mobile. The interesting thing is that zero startups succeeded in those categories; 100% of the value flowed to incumbents, right? Like Google, Facebook, and Amazon—they own all those businesses. People often forget that Google Docs wasn't the only company that tried to bring Microsoft Office online; there were about 30 companies that attempted to do so.

There are about thirty companies that tried to bring Microsoft Office online, but they all lost to Google. Then there was a second category, which consisted of mass consumer ideas that were not obvious and that nobody predicted. Examples include Uber, Instacart, DoorDash, Coinbase, and Airbnb. Those innovations came out of left field.

The connection between XML, HTTP requests, and Airbnb is not very obvious. As a result, incumbents didn't even attempt to compete in those areas until it was too late, allowing startups to gain an advantage. There's a third category, which includes all the B2B SaaS companies—about 300 of them. Consequently, in terms of the number of logos, far more billion-dollar companies were created in that third category than in the first two.

I think one reason why that happened is that there is no Microsoft of SaaS. There is no company that somehow provides SaaS solutions for every vertical and every product. For structural reasons, it seems to be the case that they are all different companies, which is why there are so many. I think Salesforce is probably the first true SaaS company. I remember Mark Benioff coming to speak at Y Combinator, and he tells the story of how it all began very early on.

People just didn't believe you could build sophisticated enterprise applications over the cloud or via SaaS. There was a significant perception issue, right? It was as if people thought, "No, you don't. You buy your box software, and that's the real software that you run." This mindset was deeply ingrained: "You buy your box software, and that's the real software that you run."

Mark Mandelski: This is the way we always do it. Mark Mandelski: It was quite contrarian because the early web apps sucked. Via web, you had to be a visionary like Paul Graham and understand that the browser was going to keep getting better and that eventually it would be good. This feels quite reminiscent of today, right? Yeah.

Where it's like, yeah, the same thing. Like, oh no, you won't be able to build sophisticated enterprise applications that use these LLM or AI tools because they hallucinate or they're not perfect, or they kind of feel like toys. But yeah, that's exactly like the early SaaS story. And so when I think about the parallels with LLMs, I can easily imagine this same thing happening, where there are a bunch of categories that are mass consumer applications that are obviously huge opportunities, but probably the incumbents will win all of those. So something like a general-purpose AI voice assistant that you can ask to do anything and it will go do that thing.

That's an obvious thing that should exist, but all the big players are going to be competing to be that thing, right? Apple is a little slow on that one. Why is Siri still so limited? What year is it? It makes no sense.

I mean, it's interesting. The obvious factor is search, and maybe Google will still win in that area, but Perplexity is definitely giving them a run for their money. This is the classic innovator's dilemma at the end of the day. You could argue that, going back to what you said about Uber or Airbnb, these were actually really risky ventures from a regulatory standpoint. So, if you're Google and you have essentially a guaranteed, enormous pot of gold that comes to you every single month, why would you endanger that pot of gold by pursuing these ventures that might be risky or could jeopardize your revenue? I think that's probably the primary reason why the incumbents didn't end up building those products or even cloning them, even after they became big and it was obvious that they were going to work.

Google never launched an Uber clone. They never launched an Airbnb clone. I was listening to this talk by Travis, and one of the things he said that really stuck with me is that in the first use of Uber, he was very scared that he was going to go to prison for a long time. He was actually personally risking imprisonment in order to build that company.

And so, yeah, no highly paid Google executive was going to do that. What do you think about why the incumbents didn't go into B2B SaaS? Is part of the reason that many of the use cases have a very wide distribution? It's a great question. I would love to hear what you all think.

My take is that it's just too hard for a company to manage that many things at once. Each B2B SaaS company requires the people running the product to be extremely knowledgeable in one domain and to care deeply about a lot of obscure issues. For example, take Gusto. Why didn't Google build a competitor to Gusto? Well, there isn't anyone at Google who truly understands payroll and has the patience to deal with all the nuances of various payroll regulations. It just isn't worthwhile for them.

It's easier for them to just focus on a few really huge categories. In the B2B SaaS world, it's often about the unbundling and bundling of software, which is a recurring debate. I also wonder why all these vertical B2B SaaS products evolved instead of just relying on companies like Oracle, SAP, or NetSuite. Yes, NetSuite, just owning everything.

And I think it might also be worth mentioning that there's another factor attributable to the shift to SaaS and the internet. In the old ways of selling software, you had proprietary software that was really expensive to install, accompanied by a whole ecosystem around it. Whenever you wanted something custom, the integrators would often say, "Oh no, we can just build you a custom payroll feature or something like that." Then Salesforce came along with a SaaS solution, which initially seemed like it could never be as powerful or sophisticated as the expensive enterprise installation you just paid for, but they proved that it absolutely was the case. I think that opened the gates for all of these vertical SaaS solutions to emerge, doing exactly what you’re describing. The problem, however, is that with a lot of this enterprise software, if you’re a user of Oracle and NetSuite, because they have to cover so much ground, the user experience is actually pretty poor.

They're trying to be a jack of all trades but a master of none, which results in a bit of a kitchen sink type of experience. This is where, if you build a B2B SaaS vertical company, you could create a significantly better and more delightful experience. There's a stark difference between consumer products and enterprise user experience. In software, there are only, what, three price points?

It's $5 per seat, $500 per seat, or $5,000 per seat. This directly corresponds to consumer, SMB, or enterprise sales. Time immemorial has taught us that historically—although this is becoming less true with new software—enterprise software can be inadequate because it's not the end user making the purchase decision. A high-ranking individual at a Fortune 1000 company is often the one being courted for a substantial seven-figure contract. Consequently, they may choose a solution that isn't necessarily optimal for the end user, the person who has to use the software on a daily basis.

And I'm sort of curious to see how this changes with LLMs, actually. I mean, to date, one of the more salient things we've seen for both SMB and enterprise software companies, or all software companies and startups, in general, is that there's a sense that as revenue scales, the number of people you have to hire scales with it. So when you look at unicorns, even in today's YC portfolio, it's quite routine to see a company that has reached $100 or $200 million a year in revenue, but they have like 500 or 1,000 employees already. I'm just going to be very curious; even the advice I’m starting to give companies that are a month or two out of the batch feels a little different than the kind of advice I would have given last year or two years ago.

In the past, you might have said, "Let me find the absolute smartest person in all of these other parts of the organization, like customer success or sales." I wanted to find someone I had worked with, someone I knew was great, and then I would sit on their doorstep until they quit their job and came to work for me. I wanted them to be someone who could build a team for me and hire a lot of people. That might still be true, but I'm starting to sense that the landscape is shifting a little bit. You might actually want to hire more highly skilled software engineers who understand large language models and can automate the specific system tasks that are bottlenecks to your growth.

And so it might result in a very subtle but significant change in the way startups grow their businesses, sort of post-product market fit. It means that I'm going to build LLM systems that reduce my costs and eliminate the need to hire a thousand people. I think we're right at the beginning of that revolution right now. I mean, we talked about this with a thousand people. I believe we are at the forefront of this transformation.

I mean, we talked about this in a previous episode. We discussed that there could be a future unicorn or a company that is run entirely, if we take it to the limit, with only 10 employees. That's completely plausible. They would be the ones writing the evaluations and the prompts. I think what you're saying reflects a trend that was already underway before large language models (LLMs).

I remember when I was running Triplebyte, for example. We needed to build marketing and user acquisition, especially after we raised our Series B. The traditional approach was to hire a marketing executive, build out a marketing team, and essentially create a machine for sales and marketing. However, I had met a Y Combinator founder, Mike. His company was building a smart frying pan, which sounds bizarre.

But he was an MIT engineer. Yeah, do you remember this? He's an MIT engineer, and to sell the smart frying pan, he had to get really, really good at understanding paid advertising, Google Ads, and a whole bunch of other things. He took this engineer's mindset approach to it. I remember talking to him about it and realizing that it would be so much better to have an MIT engineer working on our marketing efforts than any of the marketing candidates I had spoken to.

And he was able to scale us up to spending a million dollars a month at one point on marketing and various campaigns. Triplebike had great marketing. I remember the Caltrain station takeover that you did, along with all the out-of-home advertising. It was really high-quality work. You could tell that it wasn't being done by just some VP of marketing person.

And that was all, Mike. I would often get comments from people asking me around that time, "How big is Triplebyte?" We would respond, "We're about 50 people." It felt so much better, yeah.

I thought there were hundreds of people involved. I realized that if you put a really smart engineer on some of these tasks, they just find ways to create leverage. Now, with large language models (LLMs), the potential for leverage goes far beyond what traditional software can offer. So, here's my pitch for 300 vertical AI agent unicorns. Literally every company that is a SaaS unicorn can have a corresponding vertical AI unicorn in a parallel universe. Most of these SaaS unicorns were once software companies providing boxed solutions that got disrupted by SaaS innovations.

And you can easily imagine the same thing happening again, where essentially every SaaS company builds software that some group of people uses. The vertical AI equivalent will combine the software and the people into one product. One thing to note is that enterprises are currently a bit uncertain about what agents they actually need. One approach I've observed from especially more experienced founders, like Bret Taylor, the CTO of Facebook, is that he started his company, Sierra. I don’t know all the details, but as far as I can tell, it’s broadly about enabling enterprises to deploy AI agents and customize them specifically for their needs, rather than offering a specific agent designed for a particular task.

It's something I've seen with a company called VectorShift that was funded about a year ago. They're two really smart Harvard computer scientists. What they found is that they're trying to build a platform to make it easy for enterprises to create their own internal LLM-powered agents, using no-code tools or SDKs. However, enterprises often don't know exactly what they want to use these tools for. This makes me think about the boxed software world, where I often encountered a few vendors who were essentially trying to convince people to use software at all.

And it was just like, it does everything. Then it gets more sophisticated and higher resolution. You see a lot of vertical SaaS players. We are going through that same period with LLMs, where the early winners might just be these general-purpose tools that make it easy for you to work with LLMs.

And then the vertical agents will come in over time? Or do you think there are reasons it's different now, and the vertical agents will take off on day one? Yeah, that's interesting. Because if you think about the history of SaaS, the consumer applications worked first. From 2005 to 2010, it was mostly consumer applications like email, chat, and maps.

People as individuals became accustomed to using these tools themselves. I think this made it easier to sell SaaS tools to companies because the same people are both employees and consumers. Yeah, I believe the answer might just be that this is all a continuation of software, and there’s no reason it has to reset back. LLMs don't have to revert to a few general-purpose enterprise LLM platforms doing everything, because enterprises have already been trained on the value of point solutions and vertical solutions.

The user experience is not going to be that different; these solutions will simply be much more powerful. If enterprises have already developed the mindset that startups or vertical solutions can be better than legacy broad platforms, they are likely to take a chance on a startup promising an effective vertical AI agent solution today. We are witnessing this trend now, as some of our companies are gaining traction in enterprises for these vertical AI agents faster than we have ever seen before. I believe we are just at the beginning of this journey.

All software typically starts quite vertically, and as the industries become more developed, I mean, I just answered my earlier question. It's like, you know, why does a company end up having a thousand employees? It's because early on, everyone is creating these specific point solutions. Then, at some point, you've got to go horizontal. You're already making a significant investment in sales and marketing.

And then, the only way you can actually continue to grow once you capture 100% or a significant majority of the market is by not just offering a point solution, but by providing integrated solutions that work together. Another point to consider is that the bold case for vertical AI agents could be even bigger than SaaS. With SaaS, you still need an operations team or a set of people to manage the software in order to complete all the workflows, such as approval processes or data input. The argument here is that not only will vertical AI agents replace the entire suite of SaaS software—resulting in a one-to-one mapping—but they will also significantly reduce payroll costs. When you examine a lot of corporate spending, a substantial portion is still allocated to payroll, while software costs are relatively small.

Exactly. They spend significantly more on employees than on software. Therefore, it's likely that these smaller companies will be much more efficient, requiring far fewer people for tasks like data entry, approvals, or using the software. I agree; it is very possible that the vertical equivalents could be ten times larger than the SaaS companies they are disrupting.

I mean, there are two cases. It could be that the vertical point solution is just big enough, and you don't need to do that breath thing, right? That could be a nice scenario. Should we give some examples? I feel like we've all been working with so many vertical AI agent companies.

We've got some news from the front about how things are actually going. Your former head of product, Aaron Cannon, is working at a Y Combinator company called Outset, which I worked with. Basically, they're applying large language models to the surveys and Qualtrics space. So, Qualtrics will almost certainly not be able to build the best-of-breed large language model. The funny thing about surveys is, you know, who are they actually for? They're for people who run products and for marketing teams.

It's for people who are trying to make sense of what our customers actually want. And what are surveys? Well, guess what? That's language. I feel like these types of businesses actually have to thread the needle because enterprise and SMB software is often sold based on a particular person who is the key decision-maker.

You have to go high enough in the organization so that the people you're selling to are not afraid that their jobs, or their whole team's jobs, are going to disappear. Totally. That's the approach I’ve seen many companies take when selling, because if you're selling to a team that might be replaced by AI, they will sabotage it. It simply doesn't work. I think it's interesting how much of this process is top-down, and at some point, you even need to get the C-suite to sign off on it.

A company I’m working with, Momentic, is essentially an AI agent that is starting out in QA testing. They are gaining really great traction right now, and it's interesting because you might remember a decade ago when we worked with Rainforest QA. Rainforest was a QA-as-a-service company that faced a similar tension: they couldn't actually replace your QA service team. They needed to build software that made the QA team more efficient, but this often implied trying to replace as many of them as possible, which they couldn't do entirely. As a result, they were always walking a tightrope. On one hand, they aimed to sell the software to the head of engineering, emphasizing that it would mean needing fewer QA personnel. On the other hand, they also had to sell the solution to the QA team members, who were understandably resistant to the idea of being replaced.

And so I think that was always a friction point for that business in terms of how it could scale and grow. But now, Momentic with AI can actually replace the QA people. So their pitch is not, "Oh, this makes your QA team faster." It's more like, "This just means you don't need a QA team at all." This allows them to direct their sales focus onto engineering, and engineering doesn't need buy-in from QA at this point.

And you can also go in. I mean, to start with, you can sell to companies that don't even have big QA teams at the moment. They just use something like Momentic, and it will just keep scaling with them. Scaling, and they'll never build a QA team ever. David Gardner, yeah. David Gardner is a real-life case study of what Diana was saying about why these vertical AI agent companies can be ten times as big as the SaaS companies.

I'm seeing this interesting dynamic in recruiting as well. I had a similar issue with Triplebyte, where to build software that makes it easy to screen and hire software engineers, you need buy-in from both the engineering team that the candidates are joining and the recruiting team. Essentially, the software we were developing aimed to assist in the hiring process, but we couldn't completely replace the recruiters. As a result, the recruiters were often opposed to the initiative because they perceived it as a threat to their roles.

Yes, there's always friction regarding how far you can progress when the customer you're trying to sell to is worried about being replaced. However, I think it's still early days. With AI, you can now build solutions that handle the entire recruitment process. We worked with a company in our last batch called Nico, which is effectively managing the technical screening and the initial recruiter screening while gaining great traction. As these solutions continue to develop, I believe the friction will decrease; you won't have to convince recruiters to use these new tools in the same way.

You're probably not building a recruiting team in the same way that you used to. One example is with developer tool companies, as they need to provide significant support for developers. I work with a company called Kappa.AI, which has created one of the best chatbots for addressing various technical challenges and details that are difficult to answer. I think many companies that started using their services ended up with much smaller Developer Relations (DevRel) teams because the chatbot can process a lot of developer documentation, YouTube videos produced by dev tools, and even chat history.

So it just keeps getting better and better and provides really good answers, actually. It's one of the best I've seen. Yeah, I also worked with a customer support AI agent called PowerHelp. Well, actually, we both did in the last batch, and I learned a couple of interesting things from PowerHelp.

The first category is AI agents for customer support, which is famously crowded, with supposedly around a hundred options available. If you Google “AI customer support agent,” you'll find a plethora of results. However, what I learned through my experience with PowerHelp is that this perception is somewhat misleading. Almost all of those companies are using very simple zero-shot LLM prompting techniques that cannot truly replace a real customer support team handling complex workflows. These solutions may create impressive demos, but they are not capable of replacing an entire customer support team in a large-scale company that relies on numerous representatives to manage complicated tasks daily.

You need really complicated software that does all the things that Jake Heller was talking about. There were only three or four companies even attempting to do that, and cumulatively, they had less than one percent market penetration. The market was just completely open. I could also see this being another case of hyper-specialization or hyper-verticalization. There may eventually be a single general-purpose customer support agent software company, but we are currently in the first inning of this development. That will be more of an eighth or ninth-inning scenario.

So, you know, instead, you're going to have companies like Giga MS that are doing it for Zepto, handling 30,000 tickets every single day and replacing a team of 1,000 people. However, it's very specific; it's not a general-purpose demo-ware kind of solution. It includes 10,000 test cases and a very detailed evaluation set that is essentially tailored just for Zepto and similar companies. If you're one of the other marketplace companies, you will use it because it's designed for a well-defined type of marketplace—specifically, an instant delivery marketplace. I think this is the kind of dynamic that has led to the existence of $300 billion SaaS companies rather than one $10 trillion meta SaaS offering that provides all the software for the world.

It's just that the customers require heavily tailored solutions, and it's hard to build one that works for everyone. Exactly. I mean, we already provided three examples of customer support, but they're from very different verticals. It's like with dev tool companies; they have a very different kind of support.

You need a very different training set for marketplaces, right? Yeah, I guess whether you have agents or real human beings working for you, you end up with the same problem. Every company confronts Coase's theory of the firm, which states that any given firm will grow only to a point where it becomes inefficient to be larger than that. This is why there are networks and ecosystems, and, you know, a full-blown economy. Each firm will specialize in what it is particularly good at, and the outer limits of what those firms can achieve are actually based on your ability as a manager.

So yeah, that part breaks my brain a little bit because when we spend time with Parker Conrad at Rippling, one of his favorite points is actually, well, everyone's very obsessed with the fact that the rocks can talk and, you know, maybe they can draw. But the more interesting thing for him, being someone who runs HR IT software and spends a lot of time thinking about HR, is that the coolest thing about the LMS is that the rocks can read. From his perspective, you know, he has around 3,000 employees and still runs payroll for all of them through Rippling. So I think he spends a lot of time thinking about how one person can extend their ability as a manager.

And I think we're going to spend a lot of time considering how one person can extend their abilities as a manager. I believe we will see much more development in this area. This could be a counterargument suggesting that if we are at a moment where tools for managers and CEOs are becoming significantly more powerful, it might increase the scale of the firms they can effectively manage.

And that's certainly what Rippling is trying to do. They are attempting to build a suite of HR tools that, if successful, will enable them to compete with numerous billion-dollar SaaS companies and one enormous corporation. It's a very interesting point, Gary. What made me reflect on this is that having all these AI SaaS tools will empower leaders and organizations to effectively expand their capacity to process information. There is a limit to how many meaningful relationships we can maintain as humans; it’s related to the Dunbar number, which suggests around 150 meaningful relationships for each person. However, with AI, as these tools become more capable, I believe we will be able to extend that Dunbar limit.

Yeah, I think Flo Crivello had this interesting post on Twitter that went viral. I believe someone had created a voice chat as a weekend project, where the CEO would call all 1,500 of their employees. It was a very short call, and it sounded like the CEO was speaking personally. It sort of reminds me of that scene in "Her," where it zooms out, and you're witnessing the experience of one person using the OS, but that OS is actually communicating with thousands or tens of thousands of people at the same time. Large language models can engage in conversations, so to what extent can this power enhance the capabilities of one person—or just a few people—to understand what’s going on? I heard about that.

It definitely got me thinking because, as I understood it, the project is something like it will call up all your employees, and then your employees can just ramble about what they've been doing. It will then extract the meaning from it and provide the CEO with a bullet point summary of the most important information. There were a bunch of SaaS companies that attempted to create these sorts of weekly pulses from employees using traditional SaaS software. However, that version is literally a hundred times better than the pre-LLM version of this idea. But I wonder if that particular tool goes beyond just reading and summarizing.

This is the argument: if writing is thinking, then there is actually a huge amount of work involved in the effort of figuring out who is an effective communicator and what the most important things to focus on as a company are. I wonder if, at some point, the large language models (LLMs) will go beyond just summarizing and reading to actually thinking. At that point, who is truly running the organization? An interesting thought, I guess. The other thing that is particularly intriguing about how Parker Conrad is approaching this is what I recently learned from an interview with Matt McGinnis, the COO. There are more than a hundred founders who currently work at Rippling, each running an entire SaaS vertical within the company. It’s impressive how he has built the team. Harsh probably knows a lot about it since you’ve conducted several interviews with him. Yes, I mean, the focus on recruiting founders is definitely evident. Parker essentially positions Rippling as a case against the verticalization of software, right?

He's actively trying to horizontalize and take over all of HR and IT software. Yes, the whole thesis is basically that there's an underlying platform with a lot of value, and he wants to recruit founders and teams to build on top of it. It's somewhat similar to an Amazon-esque model, focusing on shared infrastructure. I think every product they've released, including things like time tracking and so on, reflects this vision.

I mean, basically they launch a product, and it hits multi-millions of dollars in ARR on the first day of launching. And that's exactly what we were talking about earlier. Once you have a vertical and a toehold, what you're saying is, "Well, I have to spend this money on sales and marketing anyway." I essentially achieve a higher LTV while keeping my CAC constant. If you look at all the top software companies today, that's essentially what Oracle is doing.

That's what Microsoft is. That's what Salesforce is. Rippling, knock on wood, is going to be next. But it's an interesting alternative to going from zero to one totally on your own. Do you guys want to talk about some of the voice companies that we have?

I think that's an interesting subcategory of this topic that is really blowing up now. I work with a company called Salient, which essentially uses AI voice calling to automate a lot of debt collection in the auto lending space. They call people and say, "Hey, you owe $1,000 on your car." This kind of job is actually one of those jobs that tends to be a bit of a 'butter-passing' job.

It kind of sucks because a lot of these low-wage workers are employed in call centers, and it's a terrible, boring job. There is a very high employee turnover and a large headcount required to manage these operations, mainly because there are so many accounts with these banks that need assistance. This is a perfect task that AI could automate, and what Salient has done is achieve a very high level of accuracy in this area.

It has been going live with many big banks, which is super exciting. This is a company that emerged last year, demonstrating that they were able to succeed because they sold through a top-down approach. The space feels like it's moving very quickly, and we have incredible companies that are voice-infrastructure companies, like Vappy. People can get started right away, especially in retail. These companies have achieved rapid scalability because it's one of the more exciting and mind-blowing things that you can set up literally in the course of just a few hours.

And then some of the questions that remain unanswered, and we hope to figure out, are: How do you hold on to customers, especially as you encounter things like the new OpenAI voice APIs? Do you go direct? It's probably a lot more work to try to use the underlying APIs right off the bat. However, these platforms clearly set a low barrier to entry. The question then becomes: Can you continue raising the ceiling to retain customers indefinitely? Harts, you were making an interesting point earlier about how the apps that people have built on top of the language models have changed from early 2023 until now. Voice, which we were just discussing, is a great example of this. I think even if you look back just six months, it felt like the voices were not realistic enough yet, and the latency was too high.

It felt like we were probably a ways off from having AI voice apps that could meaningfully replace humans calling people. And here we are. Yeah, we're zooming out, thinking back to the first YC batch where LLM-powered apps first emerged. It was probably winter 2023, so almost two years ago now. The apps were essentially just tools that generated some text, and not even perfect text at that.

The rocks could talk. That’s about it. Yeah, it was more like copy editing, marketing edits, and email edits. It felt more like incremental changes.

Yeah. I had a company, and the one that sticks in my head is a company called Speedy Brand. What they did was make it very easy for businesses to generate a blog and produce content marketing. It's a very obvious idea, and it wasn't perfect, but it was pretty cool at the time. We've talked about that a lot on the show, but that's when the ChatGPT wrapper came around. So, hey, this is what an LLM app looks like: it's just a ChatGPT wrapper that produces very basic text. It was bound to get crushed by OpenAI in the next release. And it did. I don't know if that specific one did, but that first wave of LLM apps mostly got overwhelmed by the following wave of GPT advancements.

But I feel like we've experienced this sort of "boiling frog" effect where, from our perspective, it seems like every three months, things have been getting progressively better. Now, we're at a point where we're discussing full-on vertical AI agents that are set to replace entire teams, functions, and even enterprises. That progression is still mind-blowing to me. We're only two years in, which is still relatively early, and the rate of progress is unlike anything we've seen before. What's interesting to observe, as we discussed in the last episode, is that many of the foundational models are beginning to come head to head.

There used to be only one player in town with OpenAI, but we've been seeing changes in the last batch. Claude is a huge contender. Thank God! Competition is, you know, the soil for a very fertile marketplace ecosystem where consumers have choices. Founders have a shot, and that's the world I want to live in. So, people are watching and thinking about starting a startup or may have already started. They’re hearing all of this, but how do you know what the right vertical is for you? You're likely to find some boring, repetitive admin work somewhere, and that seems to be the common thread across all of this. If you can identify a boring, repetitive admin task, there is likely going to be a billion-dollar AI agent startup if you keep digging deep enough into it.

But it sounds like you should pursue something that you have direct experience or a relationship with. There is definitely a common thread I've noticed in the companies that I see promise in. Another one that just comes to mind is Sweet Spot. I think I mentioned this before; they are essentially building an AI agent to bid on government contracts. The way they came up with that idea, which was about a year ago, was because they had a friend whose full-time job was to sit on a government website, refreshing the page and looking for new proposals to bid on.

And they were pivoting. They thought, "Ah, that seems like something an LLM could do." There’s a company from a recent batch that pivoted into a new idea and is gaining great traction. They are essentially building an AI agent to handle processes like medical billing for dental clinics. The way they discovered this idea was that one of the founders' mothers is a dentist. He decided to work with her for a day and observe what she did.

And she's like, "Oh, all of that processing claims seems really boring. An LLM should totally be able to do that." He just started writing software for his mother's dental clinic. So, I guess, in robotics, the classic maxim is that the robots that are going to be profitable and that are going to work will be doing dirty and dangerous jobs. In this case, for vertical SaaS, look for boring, butter-passing jobs. Well, with that, we're out of time for today.

We'll catch you on the light cone next time. Thank you.