How AI is Reshaping Labor Markets: A $Trillion-Dollar Opportunity Explained

Now you have software agents that are effectively doing what, for 65 years, has been human work. Is that going to increase software revenue 2×?

It could potentially increase it 10×. It’s not even on the same kind of playing field. All the data is here in the cloud, all the compute is in the cloud, and now you just kind of mix them together. Many of these incumbents aren’t going to evolve.

Moats still matter, and a lot of the moats in software today are the same as they’ve always been. It’s kind of both defense and offense for these companies to figure out what the hell to do.

Alex, you wrote an article recently, “Input, Coffee, Output, Code.” But this idea of turning capital into labor—hasn’t this always been true? What’s new here?

Well, it certainly has been true for a long time. If you watch some old movie about the Romans, you’d have all these Roman slave laborers or Roman soldiers rowing in unison on a boat. Then, of course, you got the steamship, and you didn’t need these 50 people rowing anymore.

Clearly, there has been this long historical arc of technology in some cases augmenting labor, but it was always kind of the brawn and not the brains. I have a bunch of people sewing clothes, and now I have the loom. But everything that was kind of what we’d now call white-collar work—that hasn’t happened before.

What I talked about was the 3 or 4 different eras of software. The first was just storing information. For the longest time, if I wanted to keep track of who was on my airplane, from the Wright Brothers onward, I would have a filing cabinet. It would be, “Here’s Pan Am Flight 192, and here’s who’s on it,” and I’d write down their names. Then they might call or send a telegram back in the old days saying, “I don’t want to be on that plane anymore.” I’d erase it, and then I’d refile it.

One of the first examples of digitization of the filing cabinet was something called SABRE. This was developed by American Airlines, I think in 1959 or 1960, in concert with IBM. This was one of the first examples of taking this filing cabinet that kept track of who was on a Pan Am or American Airlines flight and putting it in a database. Instead of having filing cabinets with lots of erasers and White-Out, you replaced the filing cabinet with a computer.

This, in turn, begot travel agents and travel agencies. What was a travel agency, and what was a travel agent? They had a thin client—a little computer, a terminal. I remember booking airline tickets with my mom in the 1980s. You’d go into the travel agent, and they had a green-screen computer that connected to a mainframe in Texas, which is where SABRE was.

The first realm of software, from 1960 onward, when computers became a thing, was: Take a filing cabinet—it could be the HR filing cabinet, the medical filing cabinet, or the financial filing cabinet—and put that in software. What was that? It was a database with a front end to actually enter things.

Quicken famously did this in the 1980s for financial statements. There was a company called PeopleSoft that famously did this; it was the first HR filing cabinet put in software. SABRE obviously did this for airline tickets. I would even argue that email did it for mail. You have files of mail, and now you just have a filing system, if you will, on your computer. But the actions that were done on the software were really the same.

Imagine an HR department that has 50 people working in HR. They might have 1 person in charge of the filing cabinets, and the HR person says, “Get me David’s file, because I want to talk to him about something,” which is very scary. Don’t worry. “Get me David’s file.”

The filing-cabinet person—the gofer—was what that person would be called. “Gofer” means “go for something.” There’s the gopher that Bill Murray tries to kill in Caddyshack, but there’s also the gofer: the person who goes for something. If you were a gofer at Creative Artists Agency, you went and got files.

That person went away, and the filing cabinet went away. It was more efficient from a space perspective, but the 50 HR people are still 50 HR people today. Round 1 of software was really: Take the filing cabinet and put it not as physical files, but as a database with a front end.

Round 2 started arguably in 1998 or 1999. This is what Salesforce did. The idea of a customer relationship management product had been around for a long time. The Rolodex is an actual physical thing where you’d put every business card, organize them alphabetically, and find the person you wanted to contact that way.

Salesforce put that not just in software, but in the cloud. What QuickBooks had done for a long time, or what something called Great Plains Software had done for a long time, NetSuite did in the cloud. Zendesk put email support in the cloud.

It was still software, but instead of having a giant mainframe in your office somewhere, you now had it in the cloud. It was much easier. You didn’t have to have a dedicated IT team worrying about your server exploding in flames if your office building burned down. It was more secure.

Everything that was software 1.0 then became software 2.0, which is in the cloud. That played out from, call it, 1998 to maybe 2010.

Then that grew a little bit with the insertion of financial services. The way I like to think about this is: How many restaurants need and will pay tens of thousands of dollars a year for software? Pan Am needed this in 1960, but does a restaurant with 1 location want to spend $100,000 on a server and pay for software? No in 1960, no in 2000.

But when the idea of bundling and payment processing became a thing, along with other financial services, the market became big enough for restaurant software to exist. This is where Toast came from. Toast is a $15 billion company that does this. ServiceTitan is another example.

What’s the software market in 1965 for an HVAC contractor? Zero. What’s the cloud-software market for an HVAC contractor? Zero. But once you bundle on these other things, it becomes big enough.

The point that I’m getting to is that the same 50 HR people who worked in 1960 are the same 50 HR people in 2024. The same email-support team in 2024 was the phone-support team back in 1985, which was the letter-writing, typewriter-support team in 1965.

What’s exciting about AI is that it’s taking this filing cabinet and now allowing actions on the filing cabinet. That’s what I think is really revolutionary, because you can actually ask the software filing-cabinet application—Workday, for example—“Hey, I want to add a dependent,” and now Workday will do all the work involved with that. Workday can charge a premium for that.

Why is it “Input, Coffee, Output, Code”? The whole idea is that you now have software engineers who can build products on top of these digital, or digitized, cloud-based filing cabinets that now do the job that the end user of that software product did in 1960, 1970, 1990, 2000, 2010, 2023, and 2024.

From 2024 and 2025 onward, you now have software agents that are effectively doing what, for 65 years, has been human work.

That sounds really important, so I want to underscore it. You went through the eras: first the non-software era, then software, then the cloud, then this financial-services-enabled cloud era, and now we’re in this new era. Can you talk about how, as we chart through these different eras, the scale in this new era is fundamentally different?

It’s completely different because it’s really comparing wages to software. To take an example from a completely different field that also has no software market, there are about 4.7 million registered nurses in the US. The average wage for a nurse is a little over $120,000 a year, so it’s a high-paying profession.

That means the annual nurse—not software market, but wage market—is over $600 billion a year. That’s a lot. The worldwide software market is under $600 billion. And that nurse figure is just for the US. Of course, there are nurses in the UK, France, Angola—every country on Earth. The labor market is enormous.

But what is the dedicated software market for nurses? Probably zero, because nobody took the time to develop software, and you could try developing software, but there’s no budget. The economics didn’t make sense.

If every hospital in the US currently has a nursing shortage and wants to hire a nurse—or maybe you’re in Minneapolis, where there’s a giant Somali expat community, and you need nurses who speak that language—how do you find them? It takes 3 to 4 years to get trained as a nurse.

Now you have a software product that can deliver not everything a nurse does, of course. It can’t be a phlebotomist, and it can’t perform CPR. But it can call you the night before your colonoscopy and say, “Don’t eat food,” and it can say that in 45 languages. It can actually have a conversation with you.

That’s a labor example. Going back to financial-services land, NetSuite is used by, I think, 70% of companies that go public. It’s some very, very large number. I’m sure I’ve heard a podcast ad with that exact statistic.

Maybe it’s higher. If I’m wrong by 10 points, it doesn’t belittle the point. The key thing is that you always have people who are paying you late. You look at your accounts receivable and see, “Wow, a bunch of customers owe me millions of dollars.” That’s what you’ll see when you look at your financial statements.

Again, this is where humans take the operation on that information. I have teams in collections that will call you and remind you to pay, or I’m going to cut off the product. That’s an operation that can now be done within NetSuite.

They haven’t done this yet, but I’m sure they will. Versus charging for the filing cabinet, they can say, “We know that you wanted to hire 5 collections people. We know that you pay those collections people $80,000 a year, with benefits and everything else. We know that it takes a year to train them. Now our software product can do that—not for $80,000 a year, but for $2,000 a year.”

That’s incredible. It really is augmenting labor. The question is, how will the customer think about this? Will they say, “Once we start paying NetSuite 10 times more than we paid NetSuite last year, our software budget has ballooned. We have to cut our software spend”?

Or will they say, “We’re saving so much money. As opposed to these 5 job openings, where I’m waiting to pay $400,000 a year for 5 collections people, I can pay $10,000 a year to NetSuite”? I’m paying more for software but less for labor.

That part is very new. A lot of the hypergrowth we’re seeing in this category of company is because they’re really moving into the labor market, not just the software market. The nurse example is a prime example of that.

It’ll be very interesting to see the appetite. If you think back to the App Store or the early days, people were so reticent to pay 99¢ for an app for arbitrary reasons, just because it was what they were used to paying.

I want to talk about pricing, and whether companies and consumers are adapting. But, super quickly, as we round out these eras of cloud, it might be interesting to touch on how the previous eras have set us up for this one. You mentioned things like PeopleSoft, Quicken, Zendesk, and all these companies that are capturing data. How are we uniquely set up now because of the previous eras?

I would argue that if we just went straight to AI in 1960, this wouldn’t have worked. You still need human input to collect the customer information and put it there.

We have everything built out. Right now, this isn’t an API, but if I want to answer a customer question, the question is already on the computer, already on the internet, or already in a database. All of these things have set this up to be the ultimate platform.

This is why we love investing in what I would call systems of record. A system of record is something that has every single piece of minutiae that runs a business.

It could be the strangest business you could imagine. If I’m running a laundromat, there is actually dedicated laundromat-management software. All of these systems of record have popped up for all sorts of different businesses and all sorts of consumer use cases.

The fact that the mainstream form of communication for many adults and children is texting or email—not voice, not calling your parents whom you haven’t talked to in a long time—it’s all in the cloud. Now you can perform these operations.

I would almost say there’s been a 60-year period of digitization of physical things, putting them in the cloud. Why is the cloud part important? If this were the mainframe era of the 1970s, how would you get the AI to access information that was in some Google server somewhere, or spread across thousands of servers? How would it access all the information that was in a server basement in Indiana? That’s really hard to do.

All the data is here in the cloud, all the compute is in the cloud, and now you just kind of mix them together. The fact that systems of record for so many different types of businesses and so many different kinds of consumer use cases are now widespread—and hundreds of billions of dollars of company market cap have been created from these systems of record, either horizontal or vertical—sets us up for this moment.

A vertical system of record would be something like Toast, which vertically runs a restaurant. A horizontal one would be something like Zendesk, which does customer-support software in the cloud for every different type of company.

Building on the Toast example, you start with the cloud wave and move to the financial-services wave. It was initially a hypothesis that these vertical SaaS companies would make a lot more from financial services. Fast-forward to today, and 80% of Toast revenue is payments, insurance, and all sorts of financial services, versus software.

One of my favorite examples is Mindbody, which runs fitness-studio software. It does scheduling for employees and has a CRM. They also make a lot of money from financial services, but they still need a lot of people.

You’re never going to replace the yoga instructor, but you’ve got your financial back office and people answering the phone to answer very basic questions. All of that can start to be done with AI.

The most bullish version of that is: You then don’t need to hire people to do the tasks that aren’t human-facing and that AI can do better. Is that going to increase software revenue 2×? It could potentially increase it 10×, depending on how the customer views it and how much they’re willing to let their software budget bleed into their labor budget.

I think part of the challenge, or the potential for disruption, is that the pricing model may need to change pretty significantly. You talk a bit about this in your piece, Alex. Zendesk today is charging on a per-seat basis, but if you’re actually eating away at some of the labor, you can charge for the output of work.

How does that create the potential both for increased ACV and for disruption among a lot of these larger incumbent players, given their existing pricing structures? Can we talk about the example from Zendesk? What’s the difference between the software component and the human component?

It shows how stark the difference is. Most companies, like Salesforce, charge per seat. Zendesk, at the time I wrote my piece—they might have changed their pricing a little bit—was charging $115 per seat per month.

Imagine that you have 1,000 people working in a call center, or really more of an email-based support center, and they use Zendesk. Zendesk then profits as you hire more people. At $115 per seat per month, 1,000 seats is $115,000 a month, or about $1.4 million a year, in software spending on Zendesk.

Zendesk has about $2 billion in revenue—something in that order of magnitude of annual recurring revenue—from all these seats that pay every single month. Of course, they want their customers to grow seats.

Now, assume that each seat represents a person. How much is that person paid? How much does their healthcare cost? How much is their stipend for commuting, and how much do the yoga benefits and all these other things that the company throws in cost?

Maybe it’s $50,000 a year per person. What’s $50,000 times 1,000? That’s $50 million. So you’re spending $50 million on people and $1.4 million a year on software. Which one is bigger?

This is the concern on an intermediate-term basis. Zendesk is actually very lucky because it was a public company and got taken private. Two big private-equity firms bought it, and they’re actually working on this right now. They’re saying, “Uh-oh. If we make AI really good, then the customer that has 1,000 seats might cut down to 10 seats.”

There are 2 forms of AI tools—I mean, there are more than 2, but the common examples we talk about are autopilot and copilot.

Copilot is a productivity enhancer. I’m trying to figure out how to answer Angela’s query. It has all these questions. I just got here yesterday. I don’t even know where the bathroom is. What do I do? The copilot gives me something like, “Hey, we think this is the right answer.” It makes Angela much more productive in her job, and that’s great.

Autopilot is: “Angela quit yesterday, and we need somebody else to answer emails because it’s Black Friday. What do we do?” You throw the tool directly at the customer and have it answer the questions.

That’s the big danger. Copilot is a danger for revenue as well, because why do I have 1,000 support reps? I have 1,000 customer-support reps because each one can answer only 10 questions a day, and I get 10,000 queries a day. It’s just basic math.

If, with copilot, each rep can answer 100 questions a day, I only need 100 reps. Now Zendesk has lost 90% of its revenue. If autopilot becomes a thing and it works very well, then I need nobody, and therefore I sell no seats if I’m Zendesk.

It’s a real defense-and-offense problem for these companies to figure out what the hell to do. If they play offense, they could maybe 10× their revenue, because it’s $50 million for people versus $1.4 million for software. I’d rather get $50 million if I’m Zendesk than $1.4 million.

They’re probably not going to be able to take all $50 million, because ultimately the cost of delivering these services is very low. The moat is not that high. It’s higher for companies that have the system of record, I would argue, because all of your past correspondence with all of your customers is in Zendesk.

If I’m Salesforce, every communication I’ve ever had with any of my customers, my pipeline, and everything else is in Salesforce. It’s hard to yank that stuff out because everybody’s using it.

It’s not a binary thing where tomorrow we’re all on autopilot. We’re going to see a lot of these copilot tools, and they’ll run on the systems of record. But even as they’re running on the systems of record, I need fewer seats.

That’s why Salesforce is a $200 billion-plus public company. If they don’t do this right, they could lose all of their revenue, or most of it. If they do it really well, they could 10× the revenue. Where is it going to go?

As early-stage VCs, we’re very excited about this. All of this is the question, right? It’s more about finding the next entrepreneurs versus the fate of the incumbents.

It’s an interesting moment in time for enterprising young founders to reinvent the model and charge radically differently, because many of these incumbents aren’t going to evolve. They’re not going to change their per-seat pricing, and they risk disruption.

Is that the wedge? If you’re a startup trying to figure out how to enter the market, especially when you have companies like Salesforce that have the system of record, all the data, and all the customers, is the wedge to say, “I’m going to undercut them and charge a tenth of the price,” even though I still have great margins because I’m entering the labor part of the equation?

I think so. I wrote a piece recently called “The Messy Inbox Problem,” which is my way of describing a wedge strategy that we’re seeing across lots of different industries.

There’s a class of founders who have started building software products to solve what was historically judgment-intensive work. In lots of industries, there is some sort of human administrator whose job is to extract information from a wave of unstructured information, whether it’s emails, faxes, or transcribed phone calls, and then put that information into one of these downstream systems of record. It could be an EMR, an ERP, or a CRM system.

Historically, that work lived upstream of any of that software. It was the human’s job, and software couldn’t do that, as Alex described.

Now we’re seeing companies wedge in and replace that messy inbox problem with software, and slowly begin to eat away at all the downstream workflows. Over time, I think the thesis is that while that initial wedge is highly differentiated against the human, it really is an opportunity to eat away at everything else and become the new AI-native system of record.

We have a company called Tennr that is doing this in a healthcare context. The problem they’re solving specifically is patient referral. You go to your general practitioner, and they refer you to a specialist. It could be a dermatologist or an imaging center.

Today, they’re often faxing your medical records, and it’s somebody’s job to go physically to the fax machine, take that physical fax, and re-enter the information into the EMR system. Tennr has trained a model against, I think, 4 million healthcare-specific documents and can now extract all that information about the patient programmatically.

They’ve effectively begun to solve this patient-intake problem, and they’re able to reduce about 90% of the administrative costs of that intake before the patient actually sees the clinician. They wedged in with the messy inbox problem, and over time they’re now eating away at things like scheduling, eligibility, and benefits.

Over time, we’ll see if they become the core AI-native system of record.

Something you’re pointing to there is also the defensibility of it all. You can get the wedge, maybe through pricing. How do you actually protect your customers?

I think one place people jump to is, “You need your own models, or you need some sort of proprietary data.” Is that the defensibility of the future, or how do you think about the distinction between differentiation and defensibility?

I think AI is an incredible catalyst for differentiation. Solving the messy inbox problem with software is 1,000 times better than a human doing it. It’s not even on the same playing field. It’s a super-differentiated way to wedge in and own the downstream workflow.

Is that wedge product alone defensible? I would argue no. Today, it feels like magic to the providers they’re working with, but that capability is going to become commoditized over time. They may have an advantage because they’ve trained a model for now, but I think that’s ephemeral.

The defensibility comes from owning all of the downstream workflows, deeply integrating into every other system, and effectively owning that core, end-to-end workflow.

The hot take would be that moats still matter, and a lot of the moats in software today are the same as they’ve always been. Becoming a system of record, having a network effect, becoming a platform, having virality baked into your product, and deeply embedding yourself into the existing system so it’s hard to rip out—these are all the moats that we would always have looked for in software. I think they’re still true today.

I agree with all of that. The other way to think about this is: Why did software start with airlines?

Everybody traveled—not everybody, but a lot of people traveled. Airplane tickets were very expensive, and software was a pittance for them. It made so much sense versus having throngs of filing cabinets and gofers. It made sense to pay hundreds of thousands of dollars in 1960s money to buy some giant IBM mainframe.

This is why I brought up the financial-services example. Software for restaurants did not make sense. It just wasn’t a problem to be solved, and the market wasn’t big enough. You made the market big enough once you threw in payment processing, insurance, and all these other things that they were paying for anyway.

You can either try to come up with a wedge, as David mentioned, and then figure out how you expand that wedge, make it defensible, and become the system of record.

The other thing you can do is find things like restaurants in 1980 that had no software, needed no software, and wouldn’t pay for software, but whose labor budgets were enormous. Sometimes you don’t know what these things are.

What is the incumbent software product for compliance officers at banks and financial institutions? Excel, Word, Microsoft Edge—a browser—looking for bad things. What does a compliance officer do? It’s now in the news a lot because of this de-banking issue.

I found this in the Bureau of Labor Statistics: The 4th-fastest-growing job in America is compliance officer. There isn’t an incumbent software product that every bank and financial-services company on Earth uses.

They’re all hiring compliance officers, and it takes a long time to train them. What if account openings go down? I don’t need as many. What if account openings go up? I need more.

In some cases, it takes a month to open a business bank account because the compliance officer, or compliance officers, are backlogged. What if you deliver that via software?

There is no incumbent that can add an AI module in the same way that NetSuite was an incumbent for accountants and financial officers at companies, where it can add a module for AI work collecting money.

You find these other areas where there really isn’t an incumbent, or where the incumbent is Microsoft Excel. Sometimes these are just so bizarre. You wouldn’t think about a compliance officer until you saw some random report from the Bureau of Labor Statistics that showed that manicurist is number 1 and compliance officer is number 4. Manicurist is a hard one to replace with AI.

You talk to banks and ask, “What software do you use?” The answer is Excel again. There’s a giant labor budget and not enough people using software. That’s one way you don’t have to worry about the incumbent layering something in. It’s still a wedge, but you could probably turn it into a system of record.

Why were no venture-backed companies, or even non-venture-backed companies, built in this space? It’s because you couldn’t charge that much money. It just wasn’t a big market, in the same way that there was no restaurant-software market in 1980. It’s the exact same reason.

But there’s enough budget there to fill this very, very pressing need.

One of the most interesting parts of our job is that every time there’s a new technology shift, we have to challenge every investment thesis that we thought was not going to work. So many of them are now going to work.

If we come back to financial services, there are a lot of pretty terrible systems of record where smart people have tried to get them ripped and replaced, and it just wasn’t going to happen.

My new conclusion with AI is not that they would never do it. It’s that the replacements were 2× better; they weren’t 10× better.

If we come back to compliance, a lot of it is in Excel. You’ve probably read about the $4 billion fine by TD Bank related to transaction monitoring. They had an old transaction-monitoring system—Mantas was one of them—and they should probably get one that throws up fewer false alerts.

They’re trying to clear a backlog of several tens of thousands of alerts, and they can’t hire enough compliance people. So now an interesting wedge is: We’ll provide you with all of these agents, and, by the way, we also have a much better transaction-monitoring system that will actually fix the problem.

This labor-plus-software bundle also helps the sales process and helps the defensibility, because you’re really solving the major problem: better software, and the fact that they can’t hire the people.

That’s such an important point, because you’ve all pointed out different areas where, quite frankly, the labor is not there. As we’re talking about software disrupting labor, the natural question is: What happens to all these jobs?

Maybe we can talk about both that and the flip side of what new jobs are created. Even if we talk about the previous arcs, we saw product managers, UX designers, and social-media managers. Those were all remnants of the previous era.

How do we think this will shape up in terms of new jobs, and how will existing jobs change?

It’s always hard to prophesy these things. In 1789, it would have been hard to say what these farmers would be doing post-tractor.

Sitting in a room talking with these electronic microphones.

Exactly. Talking with these electronic microphones, in this amazing fire-in-the-sky room that we have in the basement. Our ancestors would be proud.

What was a nurse? Medicine was bloodletting, leeches, and prayer. It’s obviously changed a lot.

The 1 thing I think AI cannot do—and, if anything, AI commoditizes this—is build a relationship with somebody over golf. I think the in-person things that only humans can do, that skill set might go up in value tremendously.

At the far extreme, I talked to somebody who believes that, in the distant future, there will only be 2 jobs: You either tell a computer what to do, or you’re told by a computer what to do.

There’s a whole set of things where people could be a lot more productive in whatever job they’re doing, or do part-time work, when they have this little coach by their side saying, “Do this. Do that.”

But I think the human-connection piece is almost the most important. If I think about every other era of a new communications tool, imagine having the first telephone. Alexander Graham Bell invents the telephone, and nobody has a telephone. How do you scale this network?

You get 1, and the only person who calls you is your mother saying, “Why don’t you call me more often?” Then the first telemarketer shows up and takes advantage of the fact that you have a phone. Rather than going by horse and buggy to your house, they can try selling you something over this old-fashioned telephone.

That was a very advantageous place for that first telemarketer to be.

The history of Sears, Roebuck is really fascinating. Even though they had the Sears Tower and these giant stores, it really was the first giant mail-order catalog. They were the ones who figured out how to use the US Postal Service.

Faxes came out, and people started sending unsolicited faxes. The reason I bring all this up is that you can imagine a world where AI is selling you everything and pushing everything at you. Right now, it’s a novelty and it works really well, but once it becomes so mainstream that everybody’s doing it, it’s like that Yogi Berra expression: “It’s so crowded, nobody goes there anymore.”

You can imagine the need for actual human connectivity going up dramatically. It’s followed this pattern where, once something gets so crowded, somebody doing something different—in this case, the old-fashioned way—might be more valuable.

One of the ways we think about it is that all of us have some percentage of our job that’s rote tasks that could be automated with AI. We strongly believe that at least every white-collar job is going to have a copilot. Some might be fully agentic, going back to our Level 1 compliance reviewers.

If you imagine all of us not doing any menial tasks and focusing on the human connection and the most creative parts, having all of our days to spend on that, what might be enabled? I think that’s a pretty exciting way to think about it.

As we think about the companies that can be created in this wave, I’m curious: It does feel fundamentally different. As the 3 of you are assessing companies, are there a new layer of metrics that you pay attention to?

Using a previous wave, we got social media, and all of a sudden we were thinking about daily active users. That was a key metric people started to pay attention to. In this new wave, are there new metrics? Are there the same metrics that matter? Is it too early to tell?

I think it’s actually the exact same metrics. It’s not, “It’s AI, so therefore future profits don’t matter.” It’s the present value of future profits, and that really comes down to: How many customers do you have? Do you retain those customers? How much gross profit do you make per customer? How much overhead do you have?

I don’t think any of that changes. The reason social networks were interesting is that we knew customers retained. But would people pay for it? Would it make money? There was an open question, and therefore there was this alpha of, “Wow, we call this the smile curve.”

It’s very rare. Obviously, 100% of people use the product on Day 0 because Day 0 is when they installed it. Then people stop using it on Day 1, Day 2, and Day 3. Normally, most products just have exponential decay. By Day 200, of the 100 people who downloaded it on Day 0, 0 people use it.

What’s interesting are things like Uber or Facebook, where 100% of people use it on Day 0, then it drops off on Day 1, Day 2, and Day 3, and then it picks back up and plateaus at maybe 50%, 70%, or 90% of the original starting group. That’s so rare.

But then the question was, “Will Facebook ever make money?” People thought, “It won’t make money because it’s free.” But Facebook figured out that advertising was very valuable.

I think the majority of what we’re seeing right now is monetized via subscription, so it’s very clear how they make money. The DAU thing is just as useful today as it was before, but the money part is almost automatic.

The thing that was unique about the internet era was, “Get big and then monetize later.” We’re not seeing as many of those, but I don’t think any of the fundamental isms of evaluating a business have really changed.

The only thing that’s more dangerous is that, since AI can now write software, it’s much easier to spin these things up. To build something and scale it out, the reason Friendster failed was that its servers couldn’t stay up. MySpace should have been the winner, but it couldn’t hire good engineers.

There are all these different reasons that aren’t relevant today because the technology stack is so different. But again, it’s the present value of future profits, and that’s unchanged.

One thing that’s changed—not a business metric—is the potential market size. In the US, there’s something called the North American Industry Classification System, or NAICS. There are 600 of them, and they classify industries: How many companies are there, and what’s their labor budget?

There’s a whole host of industries where, if you looked at them before, you’d say, “There are 1,000 potential buyers. Maybe they’ll pay $1,000 a month for my software service. That’s a $12 million market.” That’s really not interesting if I’m going to build a venture-backed business.

Now, if you think you can layer in AI and replace some of the labor budgets, those markets get dramatically bigger. The different pockets where software can be built—these niche markets that weren’t that interesting—are now potentially very interesting.

I think the other dimension we’re seeing in pitches is: Are you selling software into the incumbent industry, or are you building the full-stack version?

Alex wrote a bit about this in “AI at the Gate,” which is sort of the evolution of private equity in an AI context. Think about an area like professional services. In legal, for example, the challenge a lot of law firms have is that they charge on a per-hour basis.

If AI can do what used to take 3 hours in 3 seconds, where does the revenue go? We’re seeing some people pitch the full-stack, AI-native law firm, which might have a totally different cost structure to Cravath or one of these big firms.

There are other areas within professional services that are much more aligned to benefit from that efficiency. We have a company that we haven’t announced yet—I won’t mention the name—that’s solving a lot of the workflow challenges in plaintiff law.

They operate in both employment and personal injury. In that model, unlike on a per-hour basis, they’re charging on a contingency model, meaning they don’t get paid unless there’s an outcome or a settlement of the case.

In personal injury, as an example, for every 100 leads that these lawyers get, they take 1 case. There’s a ton of that messy-inbox problem: sifting through medical records or employment documents and essentially valuing or quantifying the value of each case they’ll take on, because any case they take is essentially an investment of their labor.

What this company is doing is programmatically helping solve that intake challenge—that messy inbox problem—to automatically qualify the value of those cases. It then works as a copilot for the lawyer to draft a medical chronology, create a demand letter, file a complaint, and walk through the entire pre-litigation and litigation process.

That essentially allows the lawyer to take on 3× or 4× the number of cases. The value that the software is delivering to the practice isn’t just reducing labor costs. One way to do it is that you have fewer lawyers and the same amount of revenue.

In this case, I think what’s going to happen is that it will significantly grow these practices. They’re actually passing the cost of that software to the end client in the form of a technology expense, which they often had done historically.

The value that the software is delivering to each of these firms is highly aligned with the impact it’s having on the business. The more clients you can take on, the more people who can pay for the software, and so on. On a per-firm basis, there’s a significant revenue-expansion opportunity.

That’s an interesting tension you’ll see across industries. Does AI help by reducing cost? Is it better to build a full-stack version or sell the software in? I think there will be successes in both dimensions, but it’s something we’re seeing more of.

As you’re talking about the cost being passed along to the end user or buyer, is that just net deflationary as this permeates across the system?

I know it would take time, but eventually, if you see more competition and more people creating these AI-based labor products, people will compete on price. All of a sudden, taking on a new case is no longer $5,000; it’s $500.

You said that in the previous era the open question was, “Can we make money? Can we monetize?” Is that an open question—that over time this just becomes deflationary and firms can’t charge as much?

I think technology, if it’s done right, is always deflationary because you get productivity gains. So, for sure.

I think the defensibility point is the one we struggle with a lot. It’s so easy to build one of these things. The number-one use case is almost recursive: Which profession is using AI tools the most? It’s probably the tech people who actually build tools. That’s where things like Cursor have gotten so popular.

But what if there are 50 companies that end up doing the exact same thing? That’s the hard part.

I can’t see a scenario where prices are more expensive than humans, or where prices don’t just keep going down significantly. That’s the history of technology in a nutshell.

The 100-megabyte hard drive in 1960 weighed literally tons and was probably $1 million or something crazy. Now it’s comical. I bought some for Cyber Monday—these little microSDs. A terabyte was $10. It’s just incredible.

That is an inexorable process for technology costs in general. You also get new use cases. I like this rightward shift of the supply-and-demand curve. This is a really interesting use case where there just wasn’t demand because there was supply.

There’s a lot of supply to do something for $2,000 an hour. If I want to file a trademark with the leading trademark attorney, the trademark market or patent market might be very small because it costs too much. But now, if it only costs $5, maybe everybody does it.

Translation is another fascinating example. It doesn’t make sense for a small company to translate its introductory video into 45,000 different languages that have ever existed. Why would I translate it into Ancient Greek? But why not? It’s free.

You have all these other things that expand the market because the cost has dropped so precipitously.

Where do you guys want to see more builders applying themselves? You’re obviously seeing a lot of companies and a lot of people excited. The incumbents are clearly excited about getting in on this wave. Is there an area where you’d like to see more attention being put?

Obscure is good. We love it when somebody walks in with a decade or a career of obscurity. They served some weird job, or they were in an industry that nobody has ever heard of—the farming industry, the mining industry, or whatever industry—and they have an insight that somebody else doesn’t.

They understand the potential of AI. It’s also important to know that the technology is not ready for autopilot for a lot of these things. The use cases are too complicated, and integrating the different pipes is too complicated.

If you overshoot early, there are going to be a lot of failures, as there inevitably are in every technology revolution—not because the idea is bad, but because the technology isn’t good enough to be 100 times better.

I think it’s about finding obscure use cases where, at least for now, the technology is good enough for the use case at hand.

I’d say also that there are many industries like this, but financial services and insurance have a host of old systems—30-plus-year-old systems of record—that can now be made 10× better by incorporating AI labor and redoing them in a workflow.

Deep knowledge of those areas is a big opportunity. We have transaction monitoring in Sardine, a mortgage-loan-origination system in Vesta, and servicing. We have a couple of companies in insurance as well.

Entrepreneurs who really understand those spaces and can bring AI thinking there have a big opportunity.

I think we’ll continue to see lots of entrepreneurs wedge in with the messy inbox problem across many niche vertical industries. We’re also still on the lookout for horizontal software—AI-native versions selling into sales teams, marketing, product management, analytics, and CFOs.

In those categories, you often do have large incumbent software companies. That’s the tension: You have to understand the market structure and how likely it is for that incumbent to change its pricing model and build more AI-native features.

But I think there will be generationally defining companies built in an AI-native way in horizontal software as well.