Summary
Notes
Transcript
One quick note on this. It doesn't matter to you guys because you're not getting graded. The best piece of advice I got going into GSB was about where to sit in the room. So if you want to make RJ, which is like a top-down thing for GSB students, if you sit right where you need to sit in, like here at my house, that is eye-level with the professor who's speaking, and they're staring at you, so you get called more frequently.
So I think it's nice to be here. By the way, that said, his wife, his now wife, who was a student of mine the following year, actually was sitting over here. And she was an honor to know this woman.
All right, she's all right. Yeah, she's done. She's done. She's done better. Fine. So, folks, it's my pleasure to introduce you to Jake Safer. Jake is aGraduates from JST 12 years ago and then went to a merchant's capital, leading venture in the valley and since then emerged AI and services. As we've talked about, we just sort of see services as, that's like the number one thing that's gonna die because of AI.
So we're having a discussion today, and Jake has talked a lot about what are called AI-made services, which honestly, Jake is, so first of all, thank you for coming. Stoked to be here. By the way, I forgot to say, to add this, I just found out from Jake yesterday. So as you know, yesterday was a propitious day. Messi scored his first hat-trick 20 years after scoring his first World Cup goal.
And it turns out he's an LP in earnest. Ooh.
He has no idea that he's an LP in earnest.
We are trying to make him more money. Yeah, yeah. He needs more money. Stips?
There's a much larger LP, and we're trying to work on the top of nonprofits. Oh, OK. It's good. So you talk a lot about AI-based services.
What does it mean to be a native versus like a gentic? I don't know that I understand. Like I read all your stuff on LinkedIn, like every time it pops up, I'm like, oh wow, Jake's writing. But kind of can you explain what you see as the difference in this court? And by the way, this is open discussion, interrupt, ask, et cetera. Thank you. Well, I'll start with defining the services, which is important because it's important to define it so people know what I'm saying.
So an AI data service is a company that delivers an existing service primarily with AI in a way that's better, faster, and/or cheaper than any other. Those services can be any service. It could be a legal service, it could be an accounting service, it could be an insurance service of some sort. But the idea here is doing the discernment that currently exists, and currently covers 80% of the American economy, so 80% of our economy is discerned.
Trying to rebuild those businesses from scratch, which is what we might do in the community. AI is delivering the majority of the service. There's human experts that sit on top of AI to ensure that it's accurate, correct, and guarantee the outcome. and the service is delivered better, faster, and more cheaply. You have to start from scratch. That is sort of the hottest topic right now in this space. So the majority of the capital that has gone into this idea of AI-ifying services has gone into roll-ups.
of our pride and general catalyst to raise funds dedicated to do this specific strategy. And then basically 80% of private equity is not saying they're doing this. The idea is like we're going to buy a bunch of legacy services businesses, we're going to stitch them together, and we're going to kind of jam AI onto those legacy services businesses and try to convert them into live businesses. I am less bullish on my strategy than I am on the concept The reason I came up with the term Amity Services or AIMS for short is to distinguish this concept from the kind of roll-up strategy that has been more well-funded than the strategy on the other end.
The reason I think that a native strategy is better is because when you're doing a roll-up, you're inheriting all of the legacy org structure, all of the legacy, frankly, people, certainly legacy technology, and legacy customers from the companies that you're rolling up. while you're trying to convert all of that into something that is AI first is a really tough task. I see this just for the companies that are native and start this way. It's hard enough to try to develop this technology and sell it to customers that are intentionally buying something in AI first. It's much harder to try to do this retroactively.
Like if you're an accountant who's been working in the same accounting firm in Chicago for 30 years, you getting bought and then being told by your corporate overlord, now you have to do what they ask you to do. or it's just happening. So I think this is a really good idea in theory. It makes sense. That's why people are putting a lot of cash towards it. I think in practice, and we're already starting to see this if you talk to some of the investors behind these vehicles, it's much harder to execute. I'm just going to say Gaines is easy to execute.
Gaines is extremely hard to execute. One thing I'd like to say about these businesses is because you are building both an AI platform and product, and you are delivering external service, accounting law, whatever it is, It's basically like building McKinsey and Stripe at the same time. trying to have them talk to each other and work really closely together. It is insanely hard. These businesses are much harder to build than a software business.
But it's 80% of your time. So if you do it right, your time's the time. So those that are here that have questions. services and businesses, or even not having existing legacy businesses that they're addingNo, I don't think, like, it's not a hopeless endeavor to try to make your organization more stable. I don't want to rain on anybody's parade. But, like, I will say, like, if someone's got an insurance brokerage that they've been operating in their family business for 50 years, and I don't know if any of you have that, I don't know, I didn't look at it when it was in business, if you are going to go up against an AI-native insurance brokerage who started from scratch doing this, it's probably going to be harder for you to build the data systems necessary You can do it primarily with AI and you get better as you write your policy.
The advantage you have as a company, and this is super important, is that in services, because you're not selling a product, you're selling a service, you're selling yourself. And so your reputation matters more than the software. And so the incumbency advantage in the services space, in some ways, is actually bigger than the software space. Because in the software space, you really care who made the thing.
As long as it works really well, you can try it. You can see if it works. In a services business, you're really just selling yourself and your reputation. And so the biggest challenge that the innovative services companies have at the beginning is this cold start problem. And the cold start problem is really about credibility. So what AIMS businesses have to do when they get off the ground is find a way to borrow credibility.
They can do that in a variety of ways. The best way to do it is that the founding team Include someone who is a baller from that service. Some crazy account that everyone knows. Some crazy insurance broker that I was. The second best way to do it is that an early hire has that profile and has credibility. Ideally, that early hire has a go-to-market background. So that they're customer-facing. Yeah, I can see him. Just get that back.
The third way to do it is some sort of advisory relationship with incumbents, where you get them to join you in more advisory capacity. And that's kind of hit or miss, depending on how serious the advisory arrangement is. And the fourth way to borrow credibility is to do a partnership with an incumbent. And this may be relevant to folks that are kind of running these businesses that are a lot older and have credibility.
There are ways that you can partner with some of these AIMS companies and basically sell your credibility in exchange for some of this AI juice. So I've seen a few of our companies partner-like a good example, one of our A's businesses is a company called Mechanical Orchard. which does mainframe migration services. So most of the Fortune 500 are still running their core applications on mainframes. This is not a well-known fact, but basically if your target, your inventory management system was built in the '90s and it's still running on-prem, it's your most important system, you don't want to mess with it. But everyone knows that you need to go to the cloud so you can use AI, et cetera. This has been a service that people have been attempting for years manually. It's really, really complicated because these services, these products have millions of lines of spaghetti code that's related on top of each other.
We're doing it with AI, so it's primarily AI that's taking this code and rewriting it in the cloud, and we have some expert human developers sit on top of it to make sure it works well.
That's all great.
However, the incumbency advantage here is huge because the massive SIs, be it Accenture, Deloitte, ThoughtWorks, Infosys, all these companies that have deep tentacles into the CIO and CTO organizations and portion of hundreds are going to say, don't use my startup, use me.
And so what we figure out over time is that we can actually partner with the incumbents. Because the incumbents don't have the ability to develop this technology in-house for a variety of reasons. but they have their credibility. And so we've done this kind of go-to-market GED where we'll go to a customer together and say, let's do this deal together. We'll provide the technology and some of the labor, and you provide any other kind of human labor necessary to get this job done. And we'll split it or we'll do some sort of.
They're already buying these services. So your job is just to convince them that your thing is better, faster, and/or cheaper than the incumbent. So most AIMS businesses, often they start with a lower cost value product and find a way to exceed costs over time. Most AIM businesses get a lot of demands. And so the founders are excited and they want to say yes to those customers and they build. The challenge there is you get to the state of what I call mirage product market feeling.
So garage product market fit is a situation where youHave an AI data services business that's growing really quickly. And they even have high customer retention and high customer satisfaction, which in software land, that means your product market fit. Growing quickly, great customer retention, great MPS, product market fit. In Aids, that's not true. It's not necessarily true. It's necessary but not sufficient.
Because you could just be growing with human labor. It could be the case that the service is not truly being delivered by AI. When I talked about this nonlinear relationship between cost and revenue, this is what I'm trying to get at. So in the green, you've got traditional services businesses, which grow cost and revenue linearly. In services, often with a start, they start with this cool bump where they get a lot of revenue quickly.
But they have to scale the head count really quickly to service it because the AI product isn't ready. It takes a while to develop this platform. You have to learn from the actual execution, take that data, and use it to improve the AI product that you are using to deliver the service. The best companies have this initial bump, get a bunch of data, and then they slow sales. This is really kind of tough. So the idea is you start saying no to customers. And the reason you do this is to invest in AI large.
They also built agents to sit on top of the GL to do most of the services, and they have really highly paid, really impressive people from legacy services, companies to sit on top of all that and deliver the service. They would get to a 70% gross margin providing services historically been a much lower part because they've got AI to be inclusive and generally speak the long way. The most notable thing about this business though, as it relates to this, his being Chris, the CEO of that company, is that every other quarter since we've invested, and we invested about a year and a half ago, he stopped selling.
And this is like-I've never seen this in 12 years of New England Capital. The CEO intentionally stops the entire quarter and he takes no more calls. He has no more subtleties. It's a very intentional decision to ensure that-All customers close and B, that the platform team has time to build out the AI product in such a way that it can actually deliver the service with AI and not with. And so this revenue growth is a super weird and scary stuff thing.
It's growing very quickly, but it looks very awkward relative to those companies. But I think it's a good way to avoid. So how do you know if you've got market market? The first thing to understand is, What are the product metrics you're looking at to know if the service is able to be delivered with AI? The best-aimed businesses do a very granular deep dive of what are all the tasks necessary to do this audit, to do this fund administration, to do this-insurance brokerage, whatever the service is.
And they track how much of it can be done with AI at a high enough quality water. And they actually assign individual metrics to individual engineers to own, so that they're moving them in a granular basis. All of that rolls up into this more macro metric that we call ARR per service FTE, which There's just a ratio of how much revenue you're getting divided by the number of people that are doing this.
And obviously, you should be able to win on that as the year-round gets better. And then of course the ultimate test of this business model is, are your gross margins improving? There's a little bit of funky accounting that happens right now when people aren't fully burdening gross margins the way they should with all the costs. But the broader point is, ultimately, if I'm right, if this business model works, you will see services businesses creating somewhere in the 60% to 70% gross margins over time.
I haven't figured out the clever graph on this yet. As you'll see, I'm really into frameworks. I did consulting at a college, and so I really like two-by-twos and everything. So there will be a bunch throughout this presentation. I haven't figured out the clever visual vision of this yet, but the best AI native services businesses asymptote at like 80%, maybe 85% doable by AI. and never get to a place where they can be 100% of AI.
It's the--Advice technology vendor in that world, in my opinion. It's simply We guarantee that the product does what? Be safe. And so at the limit, A technology vendor may look more like an insurance company. where their job is just to guarantee some outcome. And what I think is cool about the services angle is that there are certain services where there's a legal requirement for there to be a human-to-day TV on-call.
So in most insurance services, there's some sort of license rule where you have to have a human being oversee whatever a younger guy is,All that stuff, there are three species involved. In most financial services, or in many financial services, I should say, there has to be a regulatory requirement that a human has to stay in a loop in some capacity. Certainly in the law. Your bar not to lose against it.
I think those are in some ways the most interesting services to go after because you can get to a place where like, yeah, AI can do 85% of it, But there's still a human being who exercises judgment and ownership of the outcome. And it's very unlikely that Entropic hires a bunch of those licensed insurance brokers to provide insurance. Even if the model can do all of those things, in some ways the value of the service providers at the limit is a warranty that whatever service they're providing is good.
One thing I learned last week-is relative to software. AIMS businesses in many ways have a better ability to improve their AI. And the reason that's the case is because they own the platform. And as a result, they can develop a better what's called eval. Do you guys know what evals are?
And the system can, whatever output system does, you can judge it against this golden data set, see if it's right or wrong, and use that to improve the system.
With a software company, you sell them an agentic piece of software to help prepare a fancy stimulus. But you as a software vendor never really see the full, you don't get the full outcome of like what happened with that? Did it work or not? Because an AIMS business owns the outcome, with the audit completed or not, and do they approve the audit or not, you actually, by doing the service, develop the golden data set, develop EOS.
So my, it's a great point, and the reason why my suggested advice to founders of the building in this space is to go after services that are already housed. Because if you're going to a business and saying, hey, you're doing this service internally, I'll just do it for you, that's the exact reaction that you get. Now, what you'll probably find is you'll try to do it yourself, and then that's not a competency.
You might do OK, but it might be harder for you to do it to a plus type degree. And so often, you might come back to a certain ender. But the broader point is that's an objection. That if you're going to sell auditing, fund admin, insurance brokerage, bank processing, automation, and whatever, all those things, those are all things that people are already outsourcing. So you're competing against the legacy vendors. Something else that's also really interesting.
So take the software they built, the AI they built, and then start using it themselves to deliver the service of doing. When I talk to CEOs, a few things they say about CEOs. The first thing he said was my competitive set changed completely. I went from competing against all the hotshot Stanford CS kids who were building AI tools for investment banks to a bunch of sleepy old people who were doing good marketing for investment banking. I just crushed that.
I should just start selling that service. and sell the app. So instead of selling software to these people, I'm just going to sell, like, did your customer really This is a very bold move. This is a large SaaS company that has announced that they are now going to be calling a gains business and stop selling software and start selling the outcome before, as I said on the front slide, stop selling the pole and start selling the fish.
I have no idea this is going to work. And frankly, I feel like-I don't want to wait around this because I hope it works because he's doing it. But you're going to see an increasing movement in this direction. Last week, they launched services on it. You're going to see more and more SaaS businesses move in this direction in the next five months.
SAS businesses are reconsidering their business model for one of two reasons. The first is, If their persona that they sell to is going away, then maybe themselves are in trouble. So for example, if you are selling software tools to product researchers, and the product research function can now largely be done using AI, and the product research teams are getting fired, then you as the SaaS vendor who's selling tools to them, you're in a tough spot.
So you've got to do something different. So considering becoming an AIMS company and selling the research itself versus software is a really smart thing to consider. So that's like the first times if your customers are going to pay. And the second, of course, this is the more buzzed about on the internet one, is if your product can be vibrated, if your thing can be done by going through co-work and building it.
I was gonna ask, what's your thinking on the ability to get sort of outsized VC returns with this type of model? 'Cause it feels like to me,Maybe harder to get like a thousand X time return, less sort of winner takes all. This is the next essential question for this whole thesis. And it also relates to number three here, which is what is the gross margin? If these businesses get to scale and they have bad gross margins, If there's enough scale, then the terminal value cash flows could be interesting enough that you could actually make some good money on it. But there's no way to know. I have no idea. I think you point to a second question, which is, is this going to be a winner-take-all the way that software often is, or is this going to look more polygopolistic the way that services are often done?
But like we compete within this 10 and we compete just for contract management software, which is like smaller than that. Services is so much lighter. And so even if You build an AI-indated contract law firm in Ains business that only has the full Fourth place in the market. If it's that large, you may end up still building enough terminal cash flow to haveBido. Don't bother.
I have a company that's doing that or it's experimenting with that right now because it's the path of least resistance. The obvious risk there is that you don't have any customer relationship and you can be disintermediated. So I think that companies will try it if they have to, but I think it's not the optimal.
The first is that Often the data actually hasn't been tracked or stored. And so like, intellectually, like, yeah, you've been around for a long time, you must have all that data. The reality is you were never collecting data to be used by agents. And so as a result, you haven't actually collected it. So there's a little, I think there's a little bit of overselling that it comes in sometimes too about the criteria.
Why don't you solve it really quickly? So this is my second point, which is the valuable data is often in SOP of how your work is done. SOP stands for procedure of how the work is done. So the way agents work is that they did data behavior. And the only way an agent can do a good job is if it understands all the granularity of the humans that are currently performing the service. And what is true is that almost no legacy providers are documenting in extreme granularity what each person does.
The level, at least in the US, the level of fear and dissatisfaction and sort of anti-posturing against AI is increasing quickly for good reasons, right? Like people are scared about their job. Also, you've got Dario and Sam who are saying that everyone's going to lose their job. So of course, like people are scared. Also, you've got data centers in people's backyards they don't like, our energy price is going up.
It's like PhDs. It'll be PhDs with pitchforks. It's going to be legit if this happens. So we as a collective group need to be... I want to say PhDs are the most...
But yeah, I don't mean to make light, I spent three days at a retreat with a bunch of people from top tech companies, a bunch of people from New York finance, a bunch of people from DC think tanks, trying to work on this issue because it is a huge issue. And if it becomes a thing where, let's say the Democratic Party in 2028 becomes the anti-AI party and puts a bunch of moratoriums on data centers, That's not for America.
I fully believe that. For our competitive landscape, it's just a bad thing. It's also bad if we do nothing to deal with the crazy side of the Trump inspector coming. So we have to be much more mindful and engaged and nuanced.
Because you have to convince someone to do something. Behavior changes are hard to fix. Raising someoneAnd say like, hey, I'm gonna buy a service for something I've never used a service order for, and you should buy my service. You have two sales. Whereas if you go to them with, you know,The, uh... And some people use an outsourced service for e-generation sales. If you went to them and said, I'm a better, faster, cheaper version of Legion, it's much easier for them to switch on to you and try it, than for you to say, I'm going to do this kind of amorphous service that you've never done before.
It's a really good point. I'll start with this two by two obviously, because two by two is too good. And then I'll give an example, and then I'll come back and work on. Part of the way, there's this guy,, who was a pricing consultant that I spent a bunch of time with. And we sort of worked on this two by two, where basically, this is beyond and in services. By the way, this is software companies and whatever. How do you know if you're well suited for outcomes-based pricing? This is all academic, and we'll talk about some of those things. So the first is, is your thing autonomous?
for two reasons. The first is there's variability on that. And it's really hard to forecast. And this is the forecast study. And the second is, that's just not how they budget. They budget per minute. That's how they always done it. That's how-when you're selling, one of the downsides of selling against the legacy service is the buyer's just applying a certain way. You've got a new model. It can be scary for them to try to change it.
Let's see this. For our first contract, Let's do 80% of the contract value in the old historical permanent way. It's not great for us, but it's the way you know to buy it.
But 20% of the contract was due on a top of business.
The buyer said yes. I was like, okay, a small enough amount on a relative basis that we'll try it that way. We said, when the contract comes up for renewal, let's stare at this data.
Let's do two things. First, let's see if we can get you comfortable with the fact that we're delivering these outcomes and that you're comfortable committing to priceing. And second, to Jonathan's point, let's figure out how to fine-tune how much to charge for the outcome based on So the offset, you know, pain that you otherwise would have felt. And buyers said yes. It has not been a year, so I don't yet know how that new contract negotiation is gonna go, but that was a way for us to like teach the buyer and like stair step our way or tiptoe into what I'm gonna call this model.
Like in law, for example, laws historically charge per hour. There's a bunch of law firms now that are saying, we'll do a contract for you for $1,000.
Doesn't matter how long it takes. We will take the risk. Like if this contract requires us to do a bunch of time,, we'll keep that cost. But if our AI is really good, we're going to capture the benefit. So this counter positioning thing of charging on a fixed outcome basis versus per hour, in some cases the buyer's like, I would rather not. In law, I think a lot of buyers want that. People just so hate the global hour.
That's historically a really long, that can take years to migrate millions of lines of code into the cloud. Just in the time between February and April, the average time it takes them to migrate a lot of Koval, which is the language that my members are in, improved by a quarter of a magnitude. Because this company, as the foundation model of coding proves, this company gets downstream value. And so we go to our customers and say, oh yeah, I can send you a quarter a year.
Think of that as a heart. prioritizing revenue growth over AI productization. So This is what this two by two attempts to describe. So in any business, you've got important stuff and you've got urgent stuff. The challenge with building a software business and a services business at the same time is that you've got to time version stuff because you have clients that need to answer it quickly. And you've got a lot of important stuff, which is like building the platform and product I would say the majority of businesses that I see in Ames right now are in the bottom right, which are called the trap.
where they're just responding to the urgent stuff. Like you close a bunch of business, customer says I need this new feature or I need you to do this service, whatever, and you say, okay, fine, I'll hire a bunch more people to do it. This customer needs it, they need it tomorrow. I'm just gonna throw the body as a problem. I'm not actually gonna take the time to invest in profitizing the answer to that request.
Turns out, like, maybe somebody bought into this, And so the opportunity to build an independent version of this, I think, is real. But just like any instructor, most of them are going to fail. And hopefully we'll get you in success. I will say that in the last technology transition, so the last major technology transition, and really the birth of my firm, is when On-Prem moved to cloud. So in the early 2000s, on-prem in the cloud, there was a whole on-prem software industry.
And many of the leaders in that era didn't make the jump. So like Siebel Systems, which was the dominant provider of CRM software, was beat by Salesforce as the cloud player taking advantage of the Google Earth business model.
I think there will be sales forces in this account.