Every three months, things have just kept getting progressively better, and now we're at this point where we're talking about full-on vertical AI agents that are going to replace entire teams and functions and enterprises. That progression is still mind-blowing to me. A lot of the foundation models are kind of coming head-to-head; there used to be only one player in town with 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 marketplace ecosystem for which consumers will have choice, and founders have a shot. That's the world I want to live in.
[Music] Welcome to another episode of The Light Cone. I'm Gary; this is Jared Harge and Diana, and collectively we've 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, especially young ones, are not fully appreciating just how big vertical AI agents are going to be. It's not a new idea; some people have been talking about vertical AI agents.
We funded a bunch of them, but I think the world has not caught on to just how big it's going to get. 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 think, 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 that they use as a consumer. As a consumer, you don't tend to use that many SaaS tools because they're mostly built for companies.
I think a lot of people have missed the basic point that if you just look at what Silicon Valley has been funding for the last 20 years, we've mostly been producing SaaS companies. Guys, that's literally been most of what has been coming out of Silicon Valley. It's over 40% of all venture capital dollars in that time period went to SaaS companies, and we produced over 300 SaaS unicorns in that 20-year time period, which is way more than every other category.
Software is pretty awesome. I was thinking back to the history of this because we always like to talk about 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 that was quite literally the catalyst for the SaaS boom. Ajax!
In 2004, browsers added this JavaScript function, XML HTTP Request, which was the missing piece that enabled you to build a rich internet application in a web browser. For the first time, you could make things in websites that looked like desktop applications, and then that created Google Maps and Gmail and set up this whole SaaS boom. Essentially, the key technology at lock was that software moved from being a thing that you got on a CD-ROM and installed on your desktop to being something that you used through a website and on your phone.
Yeah, Paul Graham actually shares in that lineage in that he was one of the first people to realize that he could take the HTTP request and then actually hook it up to a Unix prompt, and you didn't actually have to have a separate computer program that would change a website. So Viaweb was an online store, kind of like Shopify, but way back in the day. It was basically like the first SaaS app ever.
PG actually invented SaaS in like 1995; it's just that those first SaaS apps kind of sucked because they didn't have XML HTTP Request. Every time you would click a button, you would have to reload the whole page, and so it was just a shitty experience. It didn't really catch on until 2005 when XML HTTP Request spread.
Anyway, I see this LLM thing as actually very similar. It's like it's a new computing paradigm that makes it possible to do something fundamentally different. In 2005, when cloud and mobile finally took off, there was this sort of big open question of like, "Okay, well, this new technology exists; what should you do with it?
Where is the value going to accrue? Where are the good opportunities for startups? " I was going through the list of all the billion-dollar companies that were created, and I kind of had this realization that you could kind of bucket 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 that could be mass consumer products. So that's like docs, photos, email, calendar, chat—all these things that we used to do on our desktop that obviously could be moved to the browser and mobile. The interesting thing is that zero startups won in those categories; 100% of the value flowed to incumbents, right?
Like Google, Facebook, Amazon—they own all those businesses. Folks forget that Google Docs wasn't the only company that tried to bring Microsoft Office online. There were like 30 companies that tried to bring Microsoft Office online.
But they all lost. Google won. Then there was a second category, which was like mass consumer ideas that were not obvious—ideas that nobody predicted.
Um, that's like Uber, Instacart, DoorDash, Coinbase, and Airbnb. Those ones came out of left field. The dot dot dot between XML HTTP request and Airbnb is very not obvious, and so the incumbents didn't even try competing in those spaces until it was like too late.
Startups were able to win there. Then there's a third category, which is all the B2B SaaS companies—there are like 300 of them. By number of logos, way 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's no Microsoft of SaaS; there’s no company that does SaaS for every vertical and every product. For structural reasons, it seems to be the case that they’re all different companies, and that's why there are so many of them. I think Salesforce is probably the first true SaaS company.
I remember Marc Benioff coming to speak at YC, and he tells the story that very early on, people just didn't believe you could build sophisticated enterprise applications over the cloud or via SaaS. There was just a perception issue. It was like, “No, you buy your box software, and that's the real software that you run the way we always do it.
” It was quite contrary because the early web apps sucked. You had to be a visionary—like PG—and understand that the browser was going to keep getting better and that eventually, it would be good. That feels quite reminiscent of today, right?
It’s the same thing: “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’re just toys. ” But yeah, that’s like the early SaaS story—exactly the same. So when I think about the parallels with LLMs, I could easily imagine the same thing happening, which is that 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 that’s something like a general-purpose AI voice assistant that you can ask to do anything, and it’ll 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’s a little slow on that one. Why is Siri so stupid still? What year is it?
It makes no sense. I mean, as a counter to that, the very obvious thing is search, and maybe Google will still win on search, but Perplexity is definitely giving them a run for their money, right? Yeah, this is the classic innovator's dilemma at the end of the day.
I mean, you could argue, going back to what you said about Uber or Airbnb, that these were actually really risky things from a regulatory standpoint. So if you’re Google and you have basically a guaranteed giant pot of gold that comes to you every single month, why would you endanger that pot of gold to pursue these things that might be scary or might ruin the pot of gold? I think that’s probably the primary reason why the incumbents didn’t end up building those products and didn’t even clone them, even after they got 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 that he said that really stuck with me is that in the early years of Uber, he was very scared that he was going to personally go to prison for a long time. He was actually personally risking going to prison in order to build that company.
So yeah, no highly paid Googler 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 a lot of the use cases are very diverse?
It’s a great question, and I’d love to hear what you guys think. My take is that it’s just too hard to do that many things as a company. Each B2B SaaS company really requires the people who are running the product and the business to be extremely deep in one domain and care very deeply about a lot of really obscure issues.
You know, take Gusto, for example. Why didn’t Google build a Gusto competitor? Well, there’s no one at Google who really understands payroll and has the patience to deal with all the nuances of all these stupid payroll regulations.
It’s just not worth it for them; it’s easier for them to focus on a few really huge categories in the B2B SaaS world. It’s sort of about the unbundling and bundling of software argument that comes up a lot as well. Why did all these vertical B2B SaaS products evolve versus just Oracle or SAP or NetSuite owning everything?
And I think it. . .
Might be AI is another thing that's attributable to the shift to SaaS, and the internet is in the old ways of selling software again. Like, you had this box software that was really expensive to install, and you had a whole ecosystem around it. Any time you wanted something custom, like the integrators would just say, "Oh no, we can just build you a custom payroll feature," or something like that.
And then Salesforce comes along with a SaaS solution, and it just seems like it could never be as powerful or sophisticated as the expensive enterprise installation you just paid for. But they proved that it totally was the case, and I think that just opened the gates for all of these vertical SaaS solutions to emerge, doing exactly what you're saying. The other problem 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 bad.
They're trying to be a jack of all trades but master of none, so it ends up being a bit of a kitchen sink type of experience. This is where, if you go and build a B2B SaaS vertical company, you could do literally a 10x better experience and more delightfully, because there's this stark difference between consumer products and enterprise user experience. Yeah, well, there's only, what, three price points in software: $5 per seat, $500 per seat, or $55,000 per seat.
That maps directly to consumer, SMB, or enterprise sales. I think time immemorial has taught us that in the past—and this is less and less true with new software, thankfully—enterprise is terrible software because it's not the user buying it. You know, some high-up muckety-muck inside a Fortune 1000 is the person who's getting whined and dined for this mega seven-figure contract.
They’re going to choose something that maybe isn't that good actually for the end user, the person who has to actually use the software day-to-day. I'm sort of curious to see how this changes with LLMs, actually. I mean, to date, one of the more salient things that we've seen for both SMB and enterprise software companies—or all software companies, all startups, period—is that there's a sense that as revenue scales, the number of people you have to hire scales with it.
When you look at unicorns, even in today's YC portfolio, it's quite routine to see a company that reached a hundred or $200 million a year in revenue but they have like 500, a thousand, or even 2,000 employees already. I'm just going to be very curious—like, even the advice that I'm starting to give companies that are a month or two out of the batch feels a little bit different than the kind of advice I would give last year or two years ago. In the past, you might say, "Let me find the absolute smartest person in all of these other parts of the org, like customer success or sales or things like that.
" I want to find someone who I've worked with, who I know is great, and then I'm going to sit on their doorstep until they quit their job and come work for me. I want them to be someone who can build a team for me, hire a lot of people. That might still be true, but I'm starting to sense that the meta is shifting a little bit; you actually might want to hire more really good software engineers who understand large language models, who can actually automate the specific things that you need that are the bottlenecks to your growth.
So it might result in a very subtle but significant change in how startups grow their businesses post-product market fit. It means that I'm going to build LLM systems that bring down my costs, that cause me not to have to hire a thousand people. I think we're right at the beginning of that revolution right now.
I mean, we talked about this in a previous episode—we talked about there will be a future unicorn company that’s only run, if we take it to the limit, with only 10 employees. That’s completely plausible, and they’re writing the evals and the prompts. What you're saying is like a trend that was already underway pre-LLMs.
I remember when I was running Triplebyte, for example, we needed to build marketing or user acquisition, basically. Especially after we raised a Series B, the traditional way you were supposed to do that is to hire a marketing executive and build out a marketing team, and just basically spin up this machine to do sales and marketing. But I'd actually met a Y Combinator founder, Mike, whose company was basically building a smart frying pan.
Sounds bizarre, but he was an MIT engineer. Yeah, you remember this. He was an MIT engineer, and to sell the smart frying pan, he had to get really, really good at understanding paid advertising and Google Ads and just a whole bunch of stuff.
So he had taken this engineer's mindset approach to it. I remember just talking to him about it and realizing it would be so much better to have an MIT engineer working on our marketing efforts than any of the marketing candidates. I've spoken to him, and he was able to scale us up to, like, we were spending, like, a million dollars a month on just marketing and various campaigns.
Triplebyte had great marketing. I remember, like, the Caltrain station takeover that you did, all the out-of-home stuff that you did. It was really high-quality stuff; you could tell it was not being done by some VP of marketing person, and that was all Mike.
The comment I would often get when people would ask me around that time, like, "How big is Triplebyte? " and we were, like, 50 people—so much, yeah, yeah. People would be like, "I thought this was hundreds of people!
" I was like, "No, it’s all because if you put a really smart engineer on some of these tasks, they just find ways to make. . .
they find leverage. " Now, like, LLMs can go even way beyond the leverage you had, which is pure software. Okay, so here's my pitch for 300 vertical AI agent unicorns.
Literally, every company that is a SaaS unicorn, you could imagine there's a vertical AI unicorn equivalent in some new universe because, like, most of these SaaS unicorns beforehand were some box software company that was making the same thing that got disrupted by a SaaS company. You could easily imagine the same thing happening again, where now basically every SaaS company builds some software that some group of people use. The vertical AI equivalent is just going to be the software plus the people in one product.
One thing might be just Enterprises in general right now are a little unsure about what exactly they like, what agents they need. One approach I've seen from especially more experienced founders, like Brett Taylor, the CTO of Facebook, who started his company Sierra—I don't know all the details, but as far as I can tell, it's essentially more broadly about letting Enterprises deploy these AI agents and spin them up, like, custom for the Enterprise versus, like, "Oh, hey, we have this specific agent to do this. " It's something I've seen from one of my companies called VectorShift that funded about a year ago.
They're two really smart Harvard computer scientists, and what they found is that they're trying to build a platform to make it easy for Enterprises to build their own, like, use no code or SDKs to build their own internal LLM-powered agents. But Enterprises often don't know exactly what they want to use these things for. So bringing it back, I wonder if, like, in the box software world, you started off with just a few vendors who basically were 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, and you get lots of vertical SaaS players. Do we go through that same period with LLMs, where the early winners might just be these general-purpose, "Hey, we make it easy for you to do LLM stuff," 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 things worked first. Like, 2005 to 2010 was mostly consumer applications, like email, chat, and maps. People, as individuals, got used to using these tools themselves, and I think that made it easier to sell SaaS tools to companies because, you know, the same people are both employees and consumers.
Yeah, I think the answer might just be that this is all just a continuation of software and there's no reason it has to reset back. LLMs don't have to reset back 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 things will just be a lot more powerful.
So if Enterprises have already built the muscle of believing that startups or vertical solutions can be better than legacy broad platforms, they are probably going to be willing to take a bet on a startup promising a very good vertical AI agent solution today. I feel like we're all seeing that in the batch now, with some of our companies getting faster traction in Enterprises for these vertical AI agents than we've ever seen before. I think we're just early in the game.
All software sort of starts quite vertical, and then as the industries actually get much 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 actually that early in the game, everyone's making these specific point solutions and then, at some point, you've got to go horizontal.
You're already doing this crazy spend on sales and marketing, and then the only way you can actually continue to grow, once you sort of get 100% or, you know, some large majority of the market, is you actually have to do not just a point solution, but things that sort of work together. The other point of why the bull case for vertical AI agents could be even bigger than SaaS is that SaaS, you still needed an operations team or a set of people to operate the software in order to get all the workflows to be. .
. Done. I don't know approval workflows, or how you have to input the data.
The argument here is that you will not only be replacing all that set of SaaS software—so that would be like one-to-one mapping—but it's also going to eat a lot of the payroll because, for companies, a big chunk of spend is still payroll. Software's tiny; exactly, they spend way more on employees than they do on software. So, it'll be these smaller companies that are way more efficient, that need way fewer humans to do random data entry or approvals or click the software.
I agree. I think it's very possible the vertical equivalence will be ten times as large as the SaaS company that they are disrupting. I mean, there's two cases: it could be that the vertical point solution is just big enough, and you don't need to do that “bro 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 news from the front on how it's actually going. Your former head of product, Aaron Cannon, is working on a YC company called Outset that I worked with, and basically, they're taking LLMs to the surveys and Qualtrics space. So, Qualtrics is almost certainly not really going to build the best of breed large language model with reasoning, and then the funny thing about surveys is, you know, who's it actually for?
It's for people who run products for marketing teams; it's for people who are trying to make sense of, like, what do our customers actually want, and what are surveys? Like, guess what? That's language!
So, I feel like these types of businesses actually have to thread this needle because enterprise and SMB software often is 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 whole job, and/or their whole team's job, is going to go away. Totally.
That's kind of the move that I've seen a lot of companies that sell need to do, because if you're going to go and sell to the team that's going to get replaced by AI, they're going to sabotage it, man. It just does not work. So, I think this is an interesting way that a lot of these are top-down, and you have to go through at some point to even get the CEO to sign off on it.
A company I'm working with, UMCH, that's essentially an AI agent—but for at least where they're starting is like QA testing—they're getting really great traction right now. And it's interesting because, you remember a decade ago, why can't we work with Rainforest QA? Like, Rainforest was a QA-as-a-service company, and they had this exact tension of where they couldn't actually replace your QA team, so they needed to build software that made the QA team more efficient.
But really, that obviously meant trying to replace as many of them as possible. They couldn't replace the whole team, and so they were always on this sort of tightrope between trying to sell the software to the head of engineering as like, “This will mean you'll need fewer QA people,” and “Great,” but then you also have to go sell that to the QA team, who don't want to be replaced. So, I think that was always a friction for that business, for how it could scale and grow.
But now, like MTIC with AI, can actually just replace the QA people. So their pitch is not, “Oh, this makes your QA people faster;” it's like, “This just means you don't need a QA team at all. ” So they can just focus their sell onto engineering, and engineering doesn't need to buy-in from QA at this point.
You can also go in—I mean, to start with—you can go and sell to companies that don't even have big QA teams at the moment. They just use something like MTIC, and then it will just keep scaling with them, scaling, and they'll just never build a QA team ever. Yes, that is a real-life case study of what Diana was saying about why these vertical AI agent companies are going to be ten times as big as the SaaS companies.
Yeah, I'm seeing this interesting trend now—in recruiting too. I had this exact same issue with TripleVet, where to build software that makes it easy to screen and hire software engineers, you need buy-in from both the engineering team that they're joining and the recruiting team. Effectively, the software we were building was trying to replace the recruiters, but we couldn't completely replace the recruiters.
But now with AI, the recruiters were always opposing it because it was a threat to them, so there was just always friction on how far you can get when the customer you're trying to sell to is worried about being replaced. But yeah, I think now it's still early days, but now with AI you can build things that do the whole stack of recruiting. We have a company we worked with last batch, Nico worked with them, a Prior, which is actually just doing the full technical screen—the full initial recruiter screen—and getting great traction.
So, I think as those things keep going, they won't have the same issues. You won't have the friction of, "Oh, I need to convince recruiters to use this. " You're probably just not going to build a recruiting team in the same way that you used to.
I mean, another example is that even for developer tool companies, they have to do a lot of developer support. I work with this company called Cap. AI that basically built one of the best chatbots, which responds to a lot of the technical details that are hard to answer.
I think a lot of the companies that started using them actually ended up having developer relations teams that are a lot smaller because it ingests a lot of developer documentation, even the YouTube videos that dev tools put up, and a lot of the chat history. It just keeps getting better and better, and it gives really good answers. Actually, it's one of the best I've seen.
Yeah, I also worked with a customer support AI agent company called PowerHelp—well, actually, we both did—last batch, and I learned a couple of interesting things from Parel. The first is that AI agents for customer support is a category that's famously crowded, where there's supposedly like, you know, a hundred of them. If you go and Google "AI customer support agent," you'll get like a hundred results on Google.
But what I learned through working with Parel is that actually, almost all of those companies are doing very simple, zero-shot LLM prompting that can't actually replace a real customer support team that does a lot of really complicated workflows. It just kind of makes for a nice demo. To actually replace a customer support team for an at-scale company that has like a hundred customer support reps who do lots of complicated things every day, you need really complicated software that does all the stuff Jake Heller was talking about.
There were only like three or four companies that were even attempting to do that, and cumulatively, they had less than 1% market penetration. So the market was just completely open. I could also see that being another case of hyper-specialization or hyper-verticalization.
There may eventually be a single general-purpose customer support agent software company, but we're in the first inning right now. Instead, you're going to have companies like Gig ML, which is doing it for Zepto, handling 30,000 tickets every single day and replacing a team of a thousand people. But it's very specific, and it's not a general-purpose demo—we're talking about 10,000 test cases in a very detailed eval set that is basically just for Zepto and things like Zepto.
If you are any of the other marketplace companies, you're probably going to use it because that's a very well-defined kind of marketplace, like an instant delivery marketplace. I think this is the kind of dynamic that led to there being like $300 billion in SaaS companies rather than one like $1 trillion meta SaaS thing that provides all the software for the world. It's just that the customers require really heavily tailored solutions, and it's hard to build one that works for everyone.
Exactly. I mean, we already gave three examples of customer support, but they are very different verticals. It's like developer tool companies need a very different kind of support than you need for marketplace training sets, which are very different, right?
Yeah, I guess whether you have agents or real human beings working for you, you end up with the same problem, which is that every company bumps up against Coase's theory of the firm, which states that any given firm will grow only to the point where it becomes inefficient to be larger than that. That's why they sort of network and create ecosystems. In a full-blown economy, every firm will specialize in what it is particularly good at, and the outer limits of what those firms can be are actually based on your ability as a manager.
So, yeah, that part a little bit breaks my brain because, you know, when we spend time with Parker Conrad at Rippling, one of his favorite points is that everyone's very obsessed with the fact that the rocks can talk and maybe they can draw, but the more interesting thing for him—running HR IT software—is that the coolest thing about the LLMs is that the rocks can read. From his perspective, he has, I think, 3,000 employees, and he still runs payroll for all 3,000 employees through Rippling. So I think he spends a lot of time thinking about how one person can extend their ability as a manager.
I think we're going to see a lot more there; that would be a reverse argument. If we're at this moment where tools for managers and CEOs are getting much more powerful, it could increase the scale of the firm that you can run, right? And that's certainly what Rippling is trying to do; he's attempting to build this suite.
Of HR tools where, if he wins, he's going to eat a whole bunch of billion-dollar SaaS companies and like one giant company. It's a very interesting point, Gary. I think what made me think about this is that, with having all these AI SaaS tools, it's going to give the ability to all these leaders and all these orgs to basically open the aperture of the context window of how much information they can parse.
Because there's a limit to how many meaningful relationships humans can have—there's like the whole thing with the Dunbar number: it's about 300 people, that you have 150 that you can have a meaningful relationship with. But with AI, because all these rocks now can read, I think we will be able to extend that Dunbar limit that we have. Yeah, I think Flo Crel had this interesting post on Twitter that went viral around, I think someone had made a voice chat as a weekend project as a CEO, but it would call all 1,500 of their employees.
And, you know, it was a very short call—it kind of sounded like it was from the CEO, just asking personally. I mean, it sort of reminds me of that scene in *Her*, where it zooms out and, actually, you're following the experience of one person using the Her OS, but actually that Her OS is speaking to 15, you know, thousands or tens of thousands of people all at one time. How many others?
8,316? Yeah, I mean, large language models can talk and can have conversations. And then, to what extent can, you know, this power actually extend the capability of one or a few people to understand what's going on?
I heard about that Yuk; it definitely got me thinking because, as I understood, the product is something like—it just it will call up all your employees and then your employees can just ramble about what they've been doing, and it will just extract the meaning out of it and give the CEO, like, a bullet-point summary of here are the most important stuff. And there were a bunch of SaaS companies that attempted to do these sort of like weekly pulses from employees using traditional SaaS software, but like that version is literally 100 times better than the pre-ELM version of this idea. But I wonder, with like that particular tool, just like it's not going beyond just like reading and summarizing.
This is the argument of like, if writing is thinking, then there's actually just a huge amount of work that's involved in the effort of figuring out like who's an effective communicator and like, what are the most important things to be like—what are the key things to be focused on as a company? I just wonder if at some point do the LLMs go beyond just summarizing and reading and do actual thinking, at which point, like, who's actually running the organization? An interesting thought.
I guess the other thing that's kind of interesting about how Parker Conrad's thinking about it is I found out about this recently off an interview with Matt McGinness, his COO, that there are more than a hundred founders who work at Rippling now as sort of specific people who run like an entire SaaS vertical inside Rippling. It's super cool the way he's built the team—Har probably knows a lot about it because you've done a bunch of interviews with him. Yeah, I mean, it's definitely very focused on recruiting founders, and I mean Parker, like, Rippling is essentially the case against verticalization trying to horizontalization—lots of value.
And he wants to recruit founders and teams that build on top of the platform—it's almost a little bit more sort of like Amazon-esque, whereas like shared infrastructure. Yeah, I think every product that they've released—I mean things like time tracking and whatnot—I mean basically they launch a thing and it hits like multi-millions of dollars in ARR on day one of launching. And that's exactly what we were talking about earlier: once you have a vertical, once you have a toehold, what you’re saying is, "Well, I have to spend this money on sales and marketing anyway; can I, you know, basically get higher LTV and hold my CAC constant?
" And that’s sort of what you—if you look at all the top software companies today, it's like that’s what Oracle is, that's what Microsoft is, that's what Salesforce is. Rippling, knock on wood, going to be the 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 like an interesting subcategory of this stuff that's really blowing up now. I have a company that I work with called Salient that basically does AI voice calling to automate a lot of that collection in the auto loan space, which—traditionally, so they like call up people and they're like, "Hey, you owe $1,000 on your car.
" Yeah, which actually, this kind of job is one of those butter-passing jobs—it kind of sucks because a lot of these low-wage workers work in all these call centers and it's like a terrible boring job. So very high churn and a giant headcount to run these, because there are just so many accounts with these banks that have to do that, and this is a perfect task. That AI could automate, and what Salient has done is it has been able to actually get very, very accurate.
It has been going live with a lot of big banks, which is super exciting. This was a company from last year demonstrating that part of it; they were able to get in because they sold through top down. I guess the space feels like it's moving very quickly, and that we have incredible companies that are voice infra companies like Vapy.
People can sort of get started right away in retail. I mean, these companies have reached pretty fast scale just because it’s one of the more exciting, mind-blowing things that you can get up and running within, I mean literally, the course of hours. Then some of the questions that remain unanswered, and we hope they figure it out, are: how do you hold on to them, especially as you run into things like the new OpenAI voice APIs?
You know, do you go direct? It’s probably way more work to try to use the underlying APIs off the bat, but these platforms are clearly low bar. Then the question is: can you keep raising the ceiling so that you can hold on to customers forever?
Har, you were making an interesting point earlier about how the apps that people have built on top of LLMs have changed from early 2023 when it started until now. Voice, which we were just talking about, is a great example of this. I think even if you went six months back, it felt like the voices were not realistic enough yet; the latency was too high.
It felt like we were probably a ways off having AI voice apps that could meaningfully replace humans calling people up, and here we are. Yeah, I was just zooming out, thinking back to the first YC batch where LLM-powered apps first came in—probably winter 2023, you know, almost two years ago now. The apps were essentially just things that spat out some text, and not even like perfect text; kind of like a marketing edit or email edits.
It was just a kind of more incremental progress. I had a company—I mean, the one that sticks in my head is a company called Speedy Brand. All they did was make it very easy for a small business to just generate a blog and spit out content marketing.
It’s like a very obvious idea, and it wasn't perfect, but it was pretty cool at the time. That’s something we’ve talked about a bunch on the show—like this is what an LLM app looks like; it’s just a ChatGPT wrapper. It does very basic things; it spits out some text.
It’s going to get crushed by OpenAI in the next release—like, and it did. Yeah, well, I don’t know if that one did, but that first wave of LLM apps mostly got crushed by the next wave of GPT. I feel like we’ve had this sort of boiling of the frog effect where, from our perspective, it's sort of like every three months things have just kept getting progressively better.
Now we’re at this point where we’re talking about full-on vertical AI agents that are going to replace entire teams and functions in enterprises. Just that progression is still mind-blowing to me; like, we’re two years in, which is still relatively early, and the rate of progress is just unlike anything we’ve seen before. I think what’s interesting to see is—we discussed this in the last episode—many of the foundation models are kind of coming head-to-head.
There used to be only one player in town with OpenAI, but we’ve been seeing in the last batch this has been changing. Claude is a huge contender! Thank God!
It’s like competition is the soil for a very fertile marketplace ecosystem, for which consumers will have choice, and founders have a shot. That’s the world I want to live in. So, people watching and thinking about starting a startup—or maybe have already started—and they’re hearing all of this.
How do you know what the right vertical is for you? You’ve got to find some boring, repetitive admin work somewhere, and that seems to be the common thread across all this stuff. If you can find 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 go after something that you directly have some sort of experience or relationship with. That is a common thread. There’s definitely a common thread I’ve seen in the companies that I’m seeing promise with.
Another one just popped into my head: Sweet Spot. I think I mentioned this before. They’re basically building an AI agent to bid on government contracts.
The way they found that idea a year ago was they just had a friend whose full-time job was to sit there on a government website, refreshing the page, looking for new proposals to bid on. They were pivoting, thinking, “Ah, that seems like something an LLM could do. ” Another company from a recent batch, which pivoted into a new idea that’s getting great traction, is basically building an AI agent to do medical billing processing.
For dental clinics, the way they found the idea was, um, one of the founder's mothers is a dentist. So, he just decided to go to work with her for a day and just sit there, seeing what she did. She's like, "Oh, all of that, like, 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, I mean, in robotics, the classic maxim is, uh, the robots that are going to be profitable and that are going to work are going to be, um, dirty and dangerous jobs; and 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.
