“Every small business should run itself” | Lassie with a16z

AI is overhyped in Silicon Valley but underhyped in Iowa. I would actually argue software took things that were stored in paper format and made them available, first on-premises via green-screen computers, but people still had to do the work.

I never forgot what I saw there: a small-business owner who was the number-one-rated dentist on Yelp spending 200 hours a month on paperwork.

The models are trained on so much data, and they’re so large, and yet they actually don’t really know how to do any of this work.

Initially, we were actually the humans in the loop. We kind of automated away our own problems.

The battle between every startup and incumbent comes down to whether the startup gets the distribution before the incumbent gets the innovation.

We come by and say, “Hey, we’ve actually built this agent that can already provide you with tens of hours of labor.” They adopt it very quickly. They see us as someone they can bring in to actually run the practice for them.

There was a great quote from Dr. Quan about you guys, which is that Lassie isn’t replacing humans but freeing them from wearing so many hats.

It’s not like AI is going to take the jobs. In many cases, you can’t find somebody.

How are you prioritizing what you build and who you sell to? Is there a world where Lassie for dentists makes Lassie for physical therapists better?

The end goal here is that

So welcome and thank you for joining us.

Thank you for inviting us. Excited to be here.

Maybe we’ll start with the basics. Steijn, this whole company started with a conversation with you and your own dentist, Dr. Quan. What did he tell you that made you decide to quit your tech job at Robinhood and go process payments for him by hand?

I did not know that my American dream would look like this. I was at Robinhood at the time, and I came to this country to start a company. After 6 years, I moved from Amsterdam to Silicon Valley because I was looking for a hard problem to solve.

My doctor, Dr. Quan, I was a patient there. I saw him twice a year, as you do with a dentist. He knew I was looking for a hard problem, and he said, “Do you want to see how I run my business?” I said, “Absolutely.”

He walked me to the back, and I never forgot what I saw there: a small-business owner who was the number-one-rated doctor on Yelp spending 200 hours a month on paperwork and busywork. He was submitting claims by hand, and he had to stick around himself because he couldn’t find people to bill the patients.

I thought, “Wow, this is fascinating.” This was a couple of years ago. I thought this was a solved problem because, in the 1970s, my mom worked in a hospital, and that’s what she did. She brought bags of cash to the bank and then processed payments by hand. But it wasn’t a solved problem here, so that’s why this piqued my interest.

Though I have to ask: when he gave you this offer, were you in that kind of reclined position? Could he actually process your answer as yes versus no if your mouth was open and drills were in your mouth? How did that go down?

Up until now, I haven’t had cavities, so there was no drilling happening yet. Knock on wood.

Examination.

Exactly. Exactly, yeah. No, he took me aside after that because he knew I was looking for a hard problem. I was roaming around, and after that appointment, he showed me what was going on.

At first, I thought maybe it was just him, right? But that didn’t really make sense to me because he was very well-rated and used all these modern technologies. Then we started talking to other doctors because maybe Dr. Quan was just an anomaly. I also worked for a gastroenterologist in Scranton, Pennsylvania, and we saw the same thing there. I thought, “Wait, you’re doing this all by hand?”

Then we figured out there were hundreds of thousands of these small businesses that literally did this all by hand. That would be quite a fascinating problem to solve. We knew it was going to be a hard problem, but we were looking for that. This was a real pain that people were desperately looking for a solution to.

The Lassie story is so unique to me because you both spent months, if not years, before fully releasing the product—literally in the offices of the customers.

Yeah.

How did you convince them to let you in, look through the heart of the business, and get into the financials?

It’s a little weird, right? It’s like, “Hello, I work at Robinhood in growth, on the referral program. Can I get a job here? And by the way, Frédéric worked at Superhuman on product. Can we do the billing for you and take over the finances?”

I think that was the first sign that we were on to something big, because, to our surprise, all these doctors said yes when we asked them. We first talked to all of them, like Dr. Quan, because everybody likes to talk about their problems. When all these doctors started talking to us for hours, we knew this was a real problem they had.

This wasn’t some vitamin that maybe it would be nice to solve for them. This was something that kept them up at night. It made them almost quit their jobs and say, “I got into this industry because of my passion and craft”—in this case, because they wanted to take care of patients.

At first, these people told us, “There’s no chance you can come work for me because you don’t have any experience running the finances. What about HIPAA and security reasons for having access to all this information?” The first sign to us that we were on to something was that these people said yes.

Dr. Quan said, “Just come and sit here night five. You can do the job.” Dr. Sha was in Scranton, Pennsylvania. He sat us down behind the desk and said, “You can have access to anything you need to have access to, or I can teach you like my son.”

That was the first sign that this was very broken and not a solved problem. I think that triggered our intuition that we might be on to something because this was a real pain that people were desperately looking for a solution to.

From a technical perspective, Frédéric, the business started in 2020, and so much was different then in terms of what was even possible to build. How has your product-building process changed over time? How is what you thought was possible then different from what you think is possible now?

When we started the business, we were always obsessed with automating and putting the business on autopilot. That hasn’t really changed. Back then, the models weren’t that good, though—especially reasoning models, which didn’t really exist in that form.

If you think about what it takes to automate any job, you’re really looking at getting context on the work. In the case of a doctor’s office, you need access to all the historical data, the patient records, and that sort of stuff. Then you need tools to do the work. This is true whether you’re a human in the office or an agent.

We started building the context layer and the tools. It’s just that the intelligence layer wasn’t that intelligent. For the first job, it wasn’t that necessary because the most basic kind of automation didn’t require that much reasoning.

As the models got really good, we had this huge tailwind because we already had all this context built and all the tools built. As the models got better, we could just replace our intelligence, and the product would get smarter over time.

In that way, we got a little lucky, but I think the core vision hasn’t really changed at all. It was always about automating the work, not building tools for people that they would have to use.

I feel like, as the models have improved, we’re increasingly seeing software do the job of labor—which, Alex, I would say you were the first to argue, and famously argue, would be the case. I’m curious how you think about that when you look at companies and how it played into the thesis around Lassie.

I’ve given this whole presentation on the origin of software. It was basically: take a filing cabinet and put it into a database. I picked the t = 0 moment for that with the SABRE system, because airlines would just keep reservations in filing cabinets.

The SABRE system was a joint project between IBM and American Airlines. That’s why Saber is spelled with two A’s: Saber.

Then this took hold everywhere else. There were HR filing cabinets, and that became something like PeopleSoft. There were legal filing cabinets, and that became all of these LexisNexis products. There were accounting filing cabinets, and that became QuickBooks and then NetSuite.

That was the origin of software. Software took things that were stored in paper format and made them available, first on-premises via green-screen computers, because it was a lot more efficient to book an airline ticket and change it if you didn’t have to use an eraser anymore, starting with SABRE. But people still had to do the work.

I would actually argue that the world didn’t get that much more efficient with software, because all that software did was take HR—did PeopleSoft and Workday make HR efficient? Did they make HR departments more efficient? I don’t think so, because the same number of people worked in HR for the exact same size company in 1950 as probably in 2000.

Instead of using filing cabinets guarded by Steijn and Frédéric, making sure that nobody breaks into the HR files...

Now you have an IT department and a CISO to make sure nobody hacks into the IT files or the HR filing cabinet. So nothing really got more efficient. I’m somewhat exaggerating for effect here.

But what you can now do with software is edit the filing cabinet, right? It’s no longer just dumb storage; it’s actually the smart implementation of changes against those things. If it’s HR, let’s do a background check. Or let’s do onboarding, or let’s explain the benefits to this person.

If it’s accounting, what do you do with the financial statements? Imagine I’m a dentist and I see I have all these overdue invoices, and I can look them up in QuickBooks. What do I do? I might want to call and say, “Please pay me.” That’s what the filing cabinet should be doing, and not just giving you the information.

It turns out that the work is orders of magnitude bigger than the storage of information that the work is done on. That’s really been the thesis, and you need the technology to catch up so it can actually do it.

Mm-hmm.

Because in 2023, next-word prediction—which is basically what AI is—wasn’t really good at going and doing these things. It wasn’t good enough to say, “I’m going to run my practice,” or, “I’m going to do background checks. I know I’m going to have statistical inference play out, and that’s how I’m going to do a background check and make sure that Frédéric didn’t commit any crimes before I hire him for my company.” No.

But now things have gotten good enough, and that just massively expands the market size. If you think about fintech, fintech massively expanded the size of many nonfinancial markets because now you could bundle financial products with nonfinancial products. My favorite example of this is Toast, and I know you and I have talked about this a bunch.

It was a great collaboration.

But Toast could have existed in 1985. Everybody had an IBM PC. They worked pretty well. MS-DOS worked pretty well. Why wasn’t there a restaurant software company in 1985? Number 1, it was too hard to use. But number 2, you have this capex issue: could you get a big restaurant that grosses $5 million a year to spend $100,000 on an MS-DOS software product for keeping reservations, paying waitstaff, and having a little menu that showed up for the cook so they could make your hamburger more quickly or something? Nobody would pay $100,000 for that.

But if you bundle in payment processing, you’re effectively charging $100,000 for that, right? Because maybe you get a 2% vig. So fintech made the market much, much bigger for software because of this bundling effect.

And that pales in comparison to software now doing the job of labor. Fintech made it a little bit bigger, but now, instead of just being a dumb pipe for data or dumb storage of data—and instead of just charging incrementally more by bundling in financial processing—now we can do work. We can charge for work, and we can charge for work in a way that is cheaper than humans and better than humans.

I think both of those sell the opportunity short because, in many cases, you can’t even find a human. The best and funnest story of the Lassie introduction—our announcement that we made together with you, or your announcement, your amazing video—is this.

My first dentist, hopefully he’s listening to this podcast, was Ronald Sloop. He was my parents’ first friend when they moved to Florida. I’m from Florida. He retired as a dentist.

Was he a Dutch guy? He sounds very Dutch, though.

No, he was an Ashkenazi Jew from somewhere in Poland or Ukraine—wherever my family’s from, too. He’s probably 75 or 80 years old right now. But part of why he retired was that he lost his key woman who did the books and everything else. He said, “I can’t deal with this anymore. I quit.”

And then he sold his practice to his junior practitioner, and now he’s out of the dentistry business.

This is why he’s now doing nothing with his life—just playing golf or something in Southern California. He moved there from Florida, apparently.

It’s just too hard to hire the person. So it’s not like AI is going to take the jobs; in many cases, you can’t find somebody. This is the part that people don’t realize.

Imagine that there is something that every human on Earth would pay $1 for, but the cost of manufacturing that thing is $100. You just have a market failure, and I kind of call this everything to the right of the supply-and-demand equilibrium point on an Econ 101 graph. It’s like, wow, everybody would have somebody—why isn’t there a Dutch receptionist at every dentist in America?

Because, you know, there might be a guy who only speaks Dutch who shows up. Why not hire somebody who speaks Dutch? Well, because there’s only a 1-in-100 chance that a Steijn who only speaks Dutch shows up at that office. You’re going to have to pay that person €40,000. It just doesn’t make sense.

But if it were free, or if it cost $1, then every dental receptionist would have a Dutch counterpart, right? Stuff like that. So that’s where the market just expands massively once you throw in labor. You have this tiny, tiny market for software—which, by the way, is not that tiny. It’s like $1 trillion.

Concentrically around that, you have financial transactions. That’s even bigger. That’s why Visa has a very big market cap.

What can exist for businesses? But then you go to the concentric circle around that, and it’s really orders of magnitude bigger.

Yeah.

There are about 160,000 dental practices in the US alone, and they spend roughly $200,000 a year on administrative costs. The interesting part is that, because we already serve hundreds, what Alex shared is something we come across literally every day: it’s the doctor themselves, with their Harvard degree, who sits there until midnight, and it makes them not like their job anymore.

I think that’s very interesting because these small-business owners mainly want to spend time on their patients. They don’t really want to spend time on the administration, let alone working with Betty—and in this case it’s Wilder. They can’t find Betty.

So we come by and say, “Hey, we’ve actually built this agent that can already provide you with tens of hours of labor,” and they ask their friends or the people in their study club if this is real. Or they Google it and see that this is real. They then adopt it very quickly. It’s super interesting to see.

And indeed, why that’s an interesting business is because this comes out of the P&L, on the labor budget. We’re already charging 5 figures for this first agent that only does 30 hours of labor a month, and there are 200 hours of labor to be done for Dr. Sloop. That’s really interesting to see: they see us as someone they bring in to actually run the practice for them.

Which is also, on the other hand, complicated to do, because if all of a sudden the requirement for software becomes, “Hey, this is not a tool that I give Dr. Sloop and then Dr. Sloop is still on the line”—in fact, one could argue a lot of AI companies are still like that. There is a human in the loop, and ultimately, the software engineers decide what gets deployed.

We can’t do that, so we needed to build an agent. That’s why it took us years: it errs on the side of correctness, because if you take over a job—reconciling all the insurance payments, interacting with the patient to bill them—it needs to work. That was technically hard to do.

That also makes this super interesting from a technology perspective, because all of a sudden you need to build autonomous systems that run on their own and don’t have a human in the loop. It runs the business for Dr. Sloop, which makes it technically super interesting.

Yeah, there was a great quote from Dr. Quan about you guys, which is that Lassie isn’t replacing humans but freeing them from wearing so many hats. Your launch video had a clip of him talking about how he can actually coach his kids’ soccer teams now and go to their games, which is amazing.

I think there’s a gap between wanting that and being willing to adopt AI and actually having it run payments in a practice—in fact, doing it so well that most of your growth is word of mouth, with dentists recommending it to other dentists.

They do.

How did you approach the technical build process for that? What was it like getting the product to—I think you guys are at 98% automation?

Walk us through that journey.

Yeah, I think a big part of it was actually spending time in offices doing the work ourselves. I don't think we could have built a product that works as well as it does if we didn't know how to do the job.

I think another part, which we already talked about, is that a huge difference between SMBs in general and enterprises is that, in SMBs, there's nobody to use the tools. You can build a tool, but there's nobody sitting there who's going to use it. And so—

Loop at night.

Yeah.

Yeah. [laughter] It needs to go into the tool.

From the very beginning, we focused on initially being the humans in the loop. We took over all of the work and said, “We'll just do this work for you,” and we automated away our own problems.

At some point, we got to a high enough level of automation that we felt comfortable handing it back over to the office. Now, we obviously learn when we can't do something for some reason, which is pretty rare. But if we don't know how to do a certain case, we learn from what the staff tells us.

We want to get to a sufficient level of automation across any product before we sell it. For us, I think that's 95-plus, but not necessarily 100. I don't think we're going to wait until we get to 100 with one product and then do the next one.

We think about the business more as a whole: How much of the business can we automate, and how much of the labor can we do with software? As soon as we can take over a job, we take it over and move on to the next one. Over time, we'll just learn the long tail of cases.

Yeah.

For a business, it also doesn't matter that much. Let's make Dr. Sloop famous on this podcast.

He's going to learn.

Yeah. If Dr. Sloop said he could become a customer, maybe we should talk about someone else who's still active.

We could reactivate him.

Yeah, he might come out. That's how he sounded.

Yeah. Dang. That's the Series B story. We got Dr. Mac out of retirement.

We had a dental shortage, and now we don't. [laughter] We had him come out of retirement.

We got 100 million more people in the States who get good dental care.

For a business owner, it’s fine if there’s a tiny sliver left of claims that need to be touched every week. They live in a nightmare world where they have to update 200 ledgers a week because that’s how many patients they see, and then go to an insurance portal, update a system of record, and check against the bank account 200 times a week. If you instead need to do that a handful of times rather than 200 times, it saves 10 or 20 hours. If there are a handful of claims that you need to file yourself, but the majority is on autopilot, it makes it tremendously easier to run a business.

I often compare it with how we are served as tech companies. It’s the third company I’m building, and it is a lot easier and with a lot smaller team than it used to be 10 or 15 years ago, because there are great tools for us that we can use. Are all these tools completely running our finances autonomously yet, or our payroll or HR? No. But they do save a tremendous amount of time.

So I think that’s how we approach this build as well: we didn’t want a human in the loop because we want software that scales and can be implemented quickly. But it’s fine if it does, in this case, for the first agent, 98% of the work, and then there’s a sliver left. And the interesting thing is that we have thousands of staffers who are basically giving us input on how to make that appeal that the agent currently cannot do.

We come from the consumer world, right? Frédéric worked at Superhuman. I worked at Robinhood. So we have a very high bar for shipping stuff. We don’t release it before it’s actually good and hands-off and works. Then these staffers help us get it to an even higher percentage.

From an implementation and onboarding perspective, you guys integrate with existing practice-management systems for the most part, versus making practices switch a bunch of software to adopt Lassie. Alex, you've written and talked a lot about startups getting distribution before an incumbent can innovate. I'm curious about your thoughts on that in the AI era, and I would also love to hear from you guys how you thought about which path to take there.

Yeah, I had this epiphany when I was building my company: Holy crap, there really aren't many good outcomes. If you build something—I call this the TiVo problem—and TiVo and ReplayTV both invented the digital video recorder so you could pause live television, which is an amazing innovation but a terrible company because you really have very few outcomes that are good.

You either end up selling to one of the big guys, like Comcast or Time Warner Cable, but they're not going to pay you that much. Partially, that's because if Comcast bought TiVo, all of the competitors to Comcast would say, “Well, we're not going to allow this to work.” You have what I call a control discount versus a control premium. That's option 1.

Option 2 is that they copy what you've done many years later, and much crappier, if that's a word, because they have all the customers. Or maybe number 3: You do a licensing deal with them, and they take all the economics because they have all the customers.

Hence, my recognition was that a lot of startups should do the boring thing. They should build the raw pipes, just own the customer, and then build the fun feature on top. That's something I still stand by.

This is why, maybe 4 years into TrialPay, I realized what I should build was this thing called Stripe. This wasn't revisionist history, because Stripe had 5 people. It was, “Wow, we should do boring payment processing, which is a commodity business, because if we do that, we own the customer.” Chronologically, you get them first. It's a very, very boring thing, but then we had this other thing, which in my case was offer-based payments, which was very lucrative.

But you can only do that if you control the pipe, in the same way that you can only build a digital video recorder if you have digital video to record.

So how does this change with AI? It changes with AI because, in many cases, these are new categories—there is no incumbent. For most categories, imagine that I say, “I have a great idea. I'm going to do background checks for new employees as part of onboarding, and I'm going to integrate with Workday,” Workday being the biggest HR information system.

That's a great idea. However, it's such an obviously great idea that Workday might copy it, and they own all the customers. That's where the battle between a startup and an incumbent is: The incumbent might win because of its ability to add things.

I feel like that is actually magnified in the AI era. Why are big companies not good at replicating small companies? There are lots of reasons, but one of them is that they hire very bad engineers and have lots of process. Now AI kind of makes a bad engineer into a pretty good engineer, so that excuse goes away a little bit.

This is the cool thing about a lot of the AI software companies, or AI that does the work. Who is the giant-ass incumbent in dental software? You and I know the answer, but it's not the same thing as, “Here's Workday.” It's a tech company; they already have software, and they can add something to it.

I remember, actually, this is a cool story. There was a company—I think it was called X1. Microsoft Outlook had really bad search, and this company, which was funded by Idealab and all these VCs, had search for your Outlook email, which was so good.

It's like, you know who I think is going to do this eventually? I think Microsoft. That company unfortunately went to zero—or actually, Yahoo bought it a long time ago. So I kind of think the same rules apply. The battle between every startup and incumbent comes down to whether the startup gets the distribution before the incumbent gets the innovation.

With AI, one change is that the incumbent can get the innovation much more quickly. But the other is that there are a lot of categories where there never was an incumbent software company, because the only job to be done was actual human labor. That's really exciting, because now you don't have to worry about, “Oh, shoot, these guys are going to come in and eat my lunch.” Who does that?

There are a lot of industries that just don't have an incumbent software solution. For the industries that do have an incumbent software solution, the risk is very high that they will start releasing AI features. It's a really interesting way that the market is playing out right now, because if you look at the public markets, the public markets are saying, in many cases, “Oh, you're a software company? Software is dead. Software sucks. Software is zero.”

Oh, you’re an AI. It’s the opposite of what VCs are saying. It’s like, “Oh my God, you do AI stuff,” but AI is software. Software is AI; the two are the same. But I don’t think everybody’s come to that realization yet.

If you’re doing Workday but AI, or NetSuite but AI, the pure-play AI thing is software at its core. The pure-play software thing is smoking crack or not showing up to work if they’re not working on implementing AI features, because that’s what their customers are demanding of them.

Yeah. And we see exactly that: there isn’t an incumbent.

That kind of does this job or can do this job quickly.

Well, the incumbent was named Betty, and she quit 2 weeks ago. That’s the incumbent.

Yeah. Or a billing agency.

It’s Dr. Sloop’s old assistant. That’s the incumbent.

Or a version of that that is somewhere overseas or in the States. That’s indeed exactly what you’re competing with, and that’s what intrigued us so much about these small businesses, because there’s no major player there.

If you show up with software that they have never seen before, which is now possible, they will adopt it and you can grow. That’s very defensible. Maybe we can talk about it later, but the schlep you have to do to actually perform the labor and talk to all these systems is significant. You need to build an ontology to make sure that everybody in this whole ecosystem is on the same page about an insurance claim and a patient payment, because all these different systems have a slightly different definition of those things.

That also makes it harder to build, because there is no incumbent. You need to stitch a lot of things together, but that makes it extra defensible. What we had to do was first figure out all the read-and-write integrations between all these systems that Betty—the AI version—needs access to.

Then you need to figure out what the data model is that can be used across all these systems. On top of that, you need to build agents that you can’t really build without actually doing the work. There are years of work that you need to do to get that going, which makes it very defensible.

The go-to-market side is similar, right? You need to knock on millions of doors and say—The interesting thing we just talked about is that they’re not necessarily skeptical about AI. Dr. Sloop is like, “I wish this was there.” The question is, how do you get ahold of Dr. Sloop?

Dr. Sloop isn’t done at 7 p.m. There’s an emergency patient who calls, and then he goes back into the office to treat that patient. He has dinner with the kids and then opens the computer, only to realize that the supplies need to be ordered because Betty left, so he needs to do this himself right now.

For us, and for any company selling to SMBs, the interesting puzzle—and why this is such interesting go-to-market work—is how you adopt and spread AI in small businesses. That’s a super-interesting puzzle to solve. That’s the crux of it: how do you get ahold of busy, nontechnical owners?

That adoption. Well, I imagine the other part of the question for you—and I’d love to hear your thoughts on this, and I’m sure our audience would love to hear our thoughts on this—is that this is not like you download the AI app from the App Store and then you’re all done.

Right? How do you actually do the onboarding?

Yeah.

And how much of that can you automate? Because that’s part of what makes a business that is selling these things work or not work, right?

Because if you have to send your own Betty—

To every single office in the country. Yeah.

It would be amazing if it were just, “Download the Lassie app.” These systems and processes are very manual.

And I think AI is overhyped in Silicon Valley but underhyped in Iowa.

There are a lot of people in Iowa. How do you solve that last-mile distribution problem for Lassie?

Yeah. I think it’s a super-interesting problem. Let’s assume you can knock on all these doors and get them to try it. Then how do you get this adopted? Because the same argument still stands: they’re very busy and they’re not technical. Good luck setting up AI in their business.

I think that’s also why our consumer backgrounds come in handy here. At Robinhood or Superhuman, you get a time window of 48 hours. If the thing doesn’t work and you don’t provide core product value, you’re out, right? These doctors are very much the same. Not only are they hard to reach and hard to work with, but if it doesn’t work in a couple of months, you’re out the door.

It really needs to be plugged in and then do the job. A lot of the work went into not only building this agent, but also figuring out how to set people up on that agent such that almost all the friction is gone. That was a lot of work, and it’s already at a point where it’s almost self-serve.

There are a few more things left, but it’s literally at the point where a doctor in Iowa—let’s not use Sloop again—says, “Yes, I want this.” They then go to an almost Stripe-like checkout or Replit-like onboarding flow, where they hook up the bank account of the practice in the product, link the system of record, and link all the insurance portals that claims come in from.

It confirms business information: are these the doctors who actually work in your practice? Then, under the hood, it configures things. There are 1 or 2 pieces left, but I think we’re months out from having an agent that you can almost set up self-serve.

That was, and is, a big part of bringing this technology to people in Iowa: can you build a consumer-like onboarding flow where a lot of the complexity is abstracted away and happens under the hood?

Maybe not many people know this, but I think I’m very impressed by Robinhood. I’m a little biased because I work there. In order to set up an account for a user without a human in the loop—which is what Robinhood pioneered—back in the day, there was Charles Schwab, and then you would say, “Hello, I want to open an account.” Before that, you had to go into the office.

One of the many things that Robinhood pioneered is that you could almost self-serve your way onto an account. KYC gets done, your bank gets linked, and a lot of product work went under the hood to make that happen.

We ran into a lot of similar situations: how do you connect all these insurance portals and kick off the process such that digital claims are coming in reliably? How do you reliably link a bank account and all these systems of record? A big part of this was figuring that piece out.

Yeah. I guess to that point, there’s more you can build and are building for dental practices. Then there are all these other types of health care practices that could use Lassie, and there’s the broader universe of small businesses that could use you guys.

How are you prioritizing what you build and who you sell to? Is there a world where Lassie for dentists makes Lassie for physical therapists better?

The master plan.

Yes.

Are we at that part of the episode?

Yes.

Yeah. I think there are 3 steps. The end goal here is that every small business should run itself, right? The busy work is done by agents. We want to build an agent for the business that will then interface with the personal agent of a consumer, which will highly likely then interface with an agent at the insurance company or other parties that the business needs to interface and interact with.

Step 1 is, to Alex’s point, there are 160,000 dental practices in the U.S. alone and $200,000 in labor that Dr. Sloop and others can’t find. You’re looking at $1 billion in recurring revenue as a market. That’s step 1.

Then likely we will pick another type of doctor’s office that is not well served, has a big TAM, and needs a consumer-like product. As we discussed, it’s not only that you get the AI to work 95%-accurate-ish; it needs to really work, and the onboarding needs to be as simple as onboarding on Coinbase or Stripe.

The last part is that we’ve then trained AI agents to run the small business. All small businesses, at an abstract level, have a system of record that they need to read and write into. They all have customers. In doctor’s offices, they happen to be patients, but it’s interacting and transacting around payments. You need to book appointments.

I think the end goal is that if we serve all the doctor’s offices, we can help all the small businesses across the world, because we’re just the best at building AI agents for salt-of-the-earth people—or people in Iowa and Paducah, Kentucky, and hopefully in Amsterdam, where I’m from, and Hamburg, Germany, where Frédéric is from—who can start using them as well.

That’s the end goal. But, again, similar to Superhuman and Robinhood, we’re working with laser focus on getting one thing really, really right and then scaling it from there.

Yeah.

But the end goal is to help them all. Amazing. I love that as a master plan. It’s a good one—a big one. You both have been part of scaling many important companies in the past: Robinhood, Coinbase, and Superhuman, among others. Building a company in 2026 is a whole brave new world. It’s so different from ever before. What are the biggest things you’ve carried over? You mentioned some of them already. And what have you had to unlearn, or what do you think is new about being founders right now?

There’s a bunch of stuff that we’re applying now that I learned at Superhuman. I think, for one, it’s focusing on the right ICP and being really strict about who you onboard to basically guarantee that they’re going to have a great experience.

I think, in our case, it’s particularly important because if we onboard the wrong practice and say we can’t actually automate that much of their work, then now we’re stuck with this customer that we claimed we were going to automate a bunch of labor for. We can’t do it. Are we going to do it? Are we going to offboard them? It’s particularly painful—maybe more painful than in an old-world product.

There’s also—we talked a bunch about the onboarding already—but we’ve really obsessed over getting to the core product value really quickly. Our onboarding is a little bit like a story or a playbook, or like a movie. We have set points and checkpoints that we want to reach in certain time frames, and we make sure it happens every time. We measure that, of course.

I think one thing that’s really different about product building from back then to now is that, if you think about the products that you built in the past, I think it was much more about functionality, or the ability for the user to do something. Now, we think really only about what kind of labor we can automate, what work we can do, and where we can save time.

If you’re thinking about patient billing as an example, previously you would have built the ability to send a statement and the ability to receive a patient payment. But if you’re looking at where the actual work of patient billing goes today, it’s really figuring out whether the statement we’re going to send the patient has the correct amount. Once the statement is sent, the patient calls and asks, “Why do I owe this amount? Do I really have to pay this? I thought this would be covered.”

If you’re just looking at where the time goes, it actually goes into customer communication or some other, fuzzier part of the work. When we’re thinking about shipping a product like that, we’re really thinking, “Once we deliver patient billing, the office should not have to spend any more time on patient billing.” That’s super different from giving them a tool that they then need to use, which doesn’t really save them a lot of time.

And neither of you, I think, were dental experts before you started the company.

That’s safe to say.

You go twice a year. That’s pretty good. I think for the average—

You’re supposed to, right?

Yeah, of course.

When you’re hiring, are you looking for expertise in dental? What are the characteristics of team members that you want to hire?

Yeah, which is now mainly on our mind, right? Because we found a really great product-market fit in a large market where you can go after the labor that they can’t find.

We currently have 2 takes on that. One, not much has changed—a contrarian take, maybe. You still need people with a steep slope who are very ambitious and driven and have skills that are in the top 5 percentile, either in engineering or in selling, right?

I think that remains the same. Getting hold of steep-slope people is the same exercise as before, though you could do it in a more AI-native way. I think the skills are the same.

Maybe the only thing is the AI fluency of a person. I think we see a pretty clear division across the board. Do you believe that the way you code will change completely as a result of building a company in this era? When building out a finance department, do you think that’s going to be completely different from before? So, in that way, we interview specifically for that.

We want to build a 2026 version of a big organization where we ship twice as much as others and move 4 times as fast. We should not only have AI adopted in these businesses, but the big puzzle for us is how to build a team that also incorporates AI into all these functions.

I think, on the one hand, nothing has changed, because the bar is still the bar. I used to be a track-and-field runner. I almost became a professional track-and-field runner, but I took a different path in life. My friends went on to the Olympics.

Olympic running is still the same, right? You need to train twice a day and push it to the edge, and that’s not given to everyone mentally and physically. That, I think, has not changed in company building. I think what you can do and the output you can generate is just 4 or 5 times greater.

That’s the big experiment that we’re doing together, right? It’s fun to see how quickly we can get to all the dentists in America and build a really good product, such that 200 hours are gone. How quickly can we then bring it to another vertical and help all small businesses? I think that is the big, interesting experiment that we’re going to do over the coming years.

If it’s so much easier and everybody can hire Betty, right? Just materialize a Betty.

Yeah.

How does that change? It could actually work out where it’s much easier to start and run a small business, but then, paradoxically, it’s much harder to be one?

Because you do have—if you think about moats in the AI era in general—

We often talk about it with respect to software companies.

Yeah.

It’s so easy to replicate XYZ software. “I did it on Replit,” or “I did it on Lovable,” or “I did it on Claude.” You hear this left and right all the time. It’s much, much harder to say, “I’m going to go replicate Dr. Sloop’s practice.”

Yeah.

But one of the things that makes a small business somewhat defensible is that it’s actually an accumulation of people who are required to deliver the end product.

Yeah.

And it’s like—if you know who Yogi Berra is—you know, the famous Yankees baseball player who said all these things that make no sense. They’re quoted often, though.

Quoted often, though. Quoted often, right?

One of my favorite ones is, “It’s so crowded, nobody goes here anymore.”

Yeah.

Yeah, it doesn’t make sense. But I guess my question is: How do you think small businesses change if it’s easier to run a small business and start one? Theoretically, that could erode the margin of a small business such that it’s so crowded, nobody goes here anymore. This one moat that exists—just materializing people to deliver a product—is now so much easier, but therefore it’s actually harder.

Yeah, I think our current take on that is that this assumes there’s a cap on demand. If you just look at the dental practice—but the same applies to trying to find a good plumber—there’s just twice as much demand as can currently be supplied.

I think this is a great opportunity for everybody to have great dental care and go twice a year, and I think the same can be said about primary care doctors. I grew up in the Netherlands. I had a very different primary care experience from what most people have in America because it’s really hard to find a good primary care doctor who is in your community, knows you and your family, and takes care of you.

We see this as an opportunity to create more capacity. Now there are 160,000 dentists, or hundreds of thousands of dentists. What if America had half a million dentists, or could see twice as many patients with maybe the same amount of dentists? Plumbing is the same.

I think there are a lot of these small businesses where they wish they could bake more pies, but they’re constrained by labor. What this unlocks is that people have twice as much time for the craft, and that is an exciting future. All of a sudden, you’re going to see a better world. I think that’s currently our view, which is super exciting about this technology.

I think that’s also, to your point, why, when we launched and finally told the world what we’d been up to, a lot of people talked about it as an optimistic example of how this new technology can be used. There are a lot of questions about what’s going to happen to the world, but nobody really can be against cleaning up busywork for small-business owners who should be baking pies, polishing nails, cleaning teeth, or drilling.

Depending on who you are, of course, but I think that's so interesting about this. Yeah.

Yeah. So you started the business in 2020, and that was arguably pre-AI, or pre-what we think of as this generative AI revolution. I would call it the BC/AD divide of November 2022, when ChatGPT launched publicly. What are the remaining problems to solve? You had reasoning models and all of these things that have built on top of the original revolution of four years ago, call it. What are the hardest problems to solve? When we talk about software that does the job of labor, what cannot be done right now? What do you feel we still need more technical advances to get there?

Sometimes it's a 90/10 thing, where the last 10% is really hard, but you can't be a feature-complete solution until you've done that. I'm just curious, from a technical lens, what are the things where you would say, “Okay?” And it's not a Lassie-specific question; it's just more about the technology and what it enables at large. Where does work still need to be done?

Where it's just not quite good enough, and what do you think the curve of that looks like? It's not a question of AGI for small business; what do you need, and where are we on that curve, if you had to estimate?

I think one thing that's interesting is that the models are trained on so much data, and they're so large, and yet they actually don't really know how to do any of this work. They don't have the workflows encoded in any way. For example, we're working on a product now where we have to collect all of these SOPs and documents about how you're supposed to bill insurance claims to certain payers and all this kind of stuff.

To some extent, humans would do the same, but there's also a big amount of human knowledge encoded in, say, these office managers. They just know how to do this work, which is weirdly not that accessible on the internet.

I think we have a big advantage there because we have all of this historical data out of their ERPs that we can look at and infer some of these workflows from. That's something we notice a lot. Actually, when we started using some of the latest reasoning models, we assumed, “Oh, they probably just know how to do this work, because why would they not? They're trained on all of this data.” But it turns out that they don't know all the intricacies of most of these workflows.

And maybe one final question, between technical and non-technical, is: I've been thinking about this a lot—how does the world change when the marginal cost of arguing goes to zero?

Right? So I was thinking about this because Cigna, who I have for my health insurance, will only send paper checks. I thought, “Oh, I must have missed the whole online enrollment.” Nope. They don't have one. Why don't they have one? They're kind of hoping that you might lose the check or might not deposit it. It's just this intentional delay. And it's the same thing for a lot of insurance: deny, deny, deny. They want to deny.

On the other side, pretend that when my house burned down, I had a Picasso in there. Both parties are trying to cheat each other. This isn't new to the insurance agency or the insurance industry, and both sides are doing it, right?

But now that everybody has this superpowered thing that costs effectively nothing to go argue in perpetuity, I can argue with you and then you can argue with me. How does that change the business dynamic of things like insurance payments and collections? I'm sure you've thought about this a lot.

Because in many cases, the counterparty that the dentist is dealing with is the insurance carrier.

Yeah. Right.

A big part of the accounts receivable, or revenue, comes from insurance companies, right?

Right?

Yeah, I think it's quite interesting because it's pretty well regulated. If a dentist does a crown and you provide them with the correct narrative—Frédéric was already talking about it—there's just a document at Cigna that specifies, “If you give me this X-ray and you give me this narrative, then we will cover that.”

But for a human, it's quite hard to follow those rules because, in dentistry, there are 50, 60, maybe 100 common billing codes. It's a lot to remember for Mildred or Betty, the staffer. In this case, our agent has all the documentation. It has access to the relevant X-ray, it knows how to build a treatment plan, and it will then submit that to the insurance company. The insurance company just needs to follow the rules, right? And then they will pay that out.

The other thing is that, because we are very deep in this industry, there is an interesting dynamic. It took me a while to understand this, but Cigna also has an incentive to have good doctors in network who are happy with them. Every year, they need to go to an employer and say, “Do you want to renew your dental plan with us?” Then the employer is going to look at whether there are good dentists in the area who are covered, whom Alex and Olivia can go to.

If the answer is no, because there's competition in this market, they will go to another insurance company. So there's actually an incentive to keep Dr. Sloop and Dr. Quan in network, because otherwise the employers will not renew the plan with them.

For us, it's quite interesting. I also think these agents are so well equipped to do this work because there's very clear documentation, to Frédéric's point. We can talk about this for hours, maybe another time, but we're also literally digitizing the file cabinet. It's really interesting that it's really hard to find these files; they're there, and once we have them, we know how to do this.

But all these payments you talked about—a lot of doctors across America still get paid on paper, too. They get $100,000 worth of checks deposited on their desk. And I did this, right? So for Dr. Quan, I literally knock, knock, knock, open the door, sit on my bar stool, and there's the mailman, in this case, who gives me a stack. I kid you not: I open all these envelopes, deposit $100,000 in checks into the doctor's bank account, and then I have to go to work on all these itemized invoices attached to the checks.

The interesting thing is that even if we had the models, you couldn't really do this until a couple of years ago because that was the status quo. The file cabinet was literally the file cabinet. In the file cabinet were the checks and the itemized invoices that you needed to handle as part of the job to keep the doctor's office running.

Then the federal government stepped in and said, “This has to stop. You cannot do paper checks anymore for much longer.” So they mandated this industry to offer direct deposit as an option to a doctor. If a doctor says, “I want to flip the switch,” you need to do that. Similarly, with these itemized invoices, you need to create a digital file format for that. And we're also riding that tailwind.

You'll see that a lot of these small businesses—I think the stats are that 70% are still paid on paper—are going through this massive digitization revolution right now because there's a federal inflection point that's regulatory. I think that, combined with these models being so good, makes this super interesting. Even if you had the models 5 years ago, the payments were still on paper, and we're digitizing that in the meantime.

That's also what we talked about earlier: it is not one click and then kaboom, because otherwise the whole country would be on digital payments.

Why does Dr. Sloop not do this well? Dr. Sloop needs to spend 50 hours figuring out with Cigna and Delta Dental and all the other insurance companies how to flip the switch. It then comes into the bank account. Do you want the staffer to have access to that bank account? Likely not, because the rent payments aren't there, along with the other data that you don't want your staff to see. So that's why the status quo is the way it is.

Then we built this massive engine that basically says, “Hey, give us the business information that we need—the tax ID number and some other nonsense—and we will go do that conversion to digital payments,” which is also a hard product problem to solve, but we've kind of productized that as well. So, yeah, it's super interesting to think about what other industries still have that because that's, I think, an even more interesting mode. It's like, okay, you can apply these models, but bringing them to Main Street is super hard. And then, in addition to that, how do you convert the data that you need that's currently living in a file cabinet to a digital file format such that you can actually automate the work?

To your earlier point about digital file cabinets not being that much more efficient, part of the reason all of these payments are still on paper is that the staff basically prefers doing the work by hand on paper. If you just digitize everything, now you have a PDF instead of a piece of paper in front of you, but you actually need tools to use that PDF, and it's preferable to just work off a sheet. So there wasn't really much of an incentive to digitize. Now that's obviously different.

So, obviously, there are a lot of dental practices that want, or even desperately need, products like Lassie, but they are distributed, they're all over the country, and there are a lot of them to reach. How do you think about reaching them? How do you bring agents to these mainstream American businesses?

Yeah, this is a very different playbook from where, currently, I think the cutting edge is. It's like, you have these models good enough, apply them in enterprises, do a few steak dinners, and then sign a contract, and you have $10 million in ARR booked, right? We literally need to go find thousands, tens of thousands, hundreds of thousands of small businesses.

So it's a super interesting problem to solve because we've built this agent that's really good. What we're doing right now is literally mapping out where all these dentists are, in this case. Then, for the next small-business type in the country, who's the owner? What systems are they on? Are there any intent signals that we can find? They're looking for a job because they say it on Indeed. Then get in touch with these people with a message that resonates with them and cuts through the noise.

I think that's especially true with our customer type, which is very different, right? If you were to sell to someone like me, you find me in Clay, enrich it with some Apollo data, and look me up on LinkedIn. Then you know, okay, this is the guy who's going to buy my HR system. We can't really do that because Dr. Sloop is not in that database. He's often not on LinkedIn. So this is a completely different playbook that we're developing here.

I think that's another very compelling thing that we're figuring out: What does the go-to-market playbook look like to adopt AI at many, many businesses? So, yeah, it's quite an exciting and untapped opportunity.

Thank you both so much for coming to chat with us today. This was awesome. We are very, very excited for the future of Lassie. Anyone who’s listening who might be interested in working with the Lassie team to build something generational here, check it out at lassie.ai. And you guys are very actively hiring from what I understand.

Oh, yeah.

Amazing. Great. Well, thank you guys again.

Thank you so much. Thanks for hosting us.