Full Transcript for Deel Head of Product on How AI Agents Do the Work of 600 Employees and Added $140M in New Revenue Without Hiring | Aaron Goldsmid, SVP of Product, Deel

You guys recently published a stat saying that you found over $140 million in ARR. That's the equivalent of the work of 600 people that now you can do with agent. How you actually connected those type of investments with actual ROI? It's picking up the change on the floor. Software was the last artisanal craft. That has changed. Aaron Goldsmith, SVP of Product at Deal.

You've been at Deal for almost four years at this point. I know you guys recently announced $1.5 billion in ARR. But where was the company when you joined? I think it was around $250 million, which is an inconceivable set of growth for such a short period of time. There's one other aspect, which I believe is true, is we're also the largest remote company, like full stop.

people have to choose. Either you grow faster at the expense of not making money or you become profitable at the expense of not growing fast enough. How were you able to pull both levers at the same time?

Hey, this is Carlos, CEO at Product School and your host on the Product Podcast. My guest today is Aaron Goldsmith, who runs Product at Deal. He joined about four years ago when the company was around 250 million in ARR. They just announced 1.5 billion, profitable for nearly three years, and they are the largest fully remote company on the planet.

He's been at Amazon, Facebook, and Twilio, and he says the hardest part of his job now is that headcount stopped being a useful proxy for anything. Here's what we'll cover. The bot he built that predicts his own review questions. Training a small model on their own corpus and beating the frontier ones for their use case. how they found a $140 million by picking up change off the floor.

No company is as sticky as it used to be, including his. AI does not replace your job, it replaces the worst part of it. Let's get into it. AI helps us generate ideas, draft plans, and spin up work in minutes. But faster doesn't necessarily mean better. When everyone has their own AI tools, their own outputs, and their own context, you get fragmentation, fractured version control, scattered review processes, decisions happening in parallel, not in concert.

Rebuilding context between people and platforms costs you time and momentum. That's where Mural comes in. Mural gives teams a shared visual canvas to bring work together. You can generate structured starting points right on the canvas instead of moving outputs between tools. Keep all of your context in one place. And compatible AI agents can contribute alongside the team, bringing in context from the tools they already know.

The future of work is still human. Learn more at mural.co slash mural dash AI. Welcome to the product podcast, Aaron. Thank you so much for having me. I can say this like it's an old public radio, long-time listener, first-time caller.

So you've been in this deal for almost four years at this point. I know you guys recently announced $1.5 billion in ARR. But where was the company when you joined? I was trying to think about this. I think it was around $250 million, which is an inconceivable set of growth for such a short period of time, which is something... And I don't mean to say that to toot the company's horn, it's wildly successful, but it's like it still feels like the difference of the $250 million and the $1.5 billion company somehow feels like it's the same company.

And in other ways, it's totally different, right? Like the way that we've scaled and stuff like that and the sophistication of customers and the complexity of the product at that different level of scale. But as an organization, it kind of feels the same. It's also a very short window, isn't it? you know, like you said, about four years. One of the fastest growing companies, also probably one of the largest employers if you include all the people that are being hired through your platform, right? So I'm curious to know what are some of those elements that you think stayed true as you stayed and which ones are the ones that you actually had to change? So I think There's

one other aspect, which I believe is true, is we're also the largest remote company, like full stop, which is another aspect. And in part, it's so important to sort of what deal is, you know, really enabling global work. You can work anywhere being that sort of global platform. You know, I think there's certain things that are in the DNA of a company.

And I think about this a lot also in terms of like competing with other companies. Like what's what is their like genetic makeup? It's like very hard to change. Right. So it's all centered around speed, customer centricity. And I was at Amazon and, you know, it's a new high watermark for being customer focused. right? You could be an average mid-market customer.

You could be on a shared Slack channel with us. Within, you know, a half hour, you'll be talking to myself, most of the C-suite, and we'll probably be pushing code within 12 hours if it was an issue, right? And so that level of access into the organization is incredibly high. I think one of the things that we've had to really learn, but then there's been some bumps along the way, is that, like, as we've scaled, the culture of You know, comes down from Alex and, you know, he is very visible to every level of the organization.

But because those like interactions are not as frequent, that culture imprinting takes a little bit longer. So it's about sort of making sure that that sort of that next set of leaders, the leaders below that are the culture carriers around making sure that there's a speed. Customer centricity, a level of ownership. You know, everyone is a player coach, which was before that became the rage.

I am ICPMing work every day. And I think it's great because if you're your hands on the craft, you're not missing anything. Right. Like, you know, when like some law that says this, but like the number of people that you have in that graph, the number of lines increases between them like every time. Right. And so if you're not still hands on, especially as a company scales, process bloat is very easy to happen or things that did not work at your SMB mid market all of a sudden, you know, or you need to do for enterprise or vice versa.

And so by really staying close to it and also like every every strategic strategy. vision insight I've ever had, even if they sort of are prepared to be like macroeconomic disruption, came from being in the weeds, working on something, and just catching it to be like, wait a second, what did that person just say? That was like, really interesting.

They probably didn't know they said it. And it was interesting. But like, it was the impetus for something that could be, you know, you could have hundreds of engineers on the next in a short period of time. big is your team? So I try to think it's like R&D is probably around 1700. And that is almost certainly a wrong number. Like we're hiring an engineer or two a day.

Products probably around 120. Mix up a bunch of things between operations and But, like, the reality is, like, I actually don't pay a lot of attention to the count anymore, right? Because it used to be those were incredibly good proxies for complexity, for output, for scope. But, you know, in the last 18 months, those proxies are, like, just not valid anymore, right? And so I really end up focusing around, like, how leveraged are these people? Let me think about, like, what's the...

Main page at 7:25

I mean, we didn't talk about PM to engineering ratio anymore. It's just, like... Are these people really fully loaded? And do they have the resources to be able to fully execute? That was going to be my follow up on that. It's like, OK, if you are the largest remote company, you have a pretty large team and you are in the weeds. How do you make sure that people are fully loaded and as self-sufficient as possible so you can stay there with them and not spend too much time on the politics and the processes? Sure.

I think most of that is really making sure that you are close enough to the work on a regular basis that you can see. Right. So like almost every product team does every two weeks as a business review. Right. Where it's the engineering product design and sometimes it's ops and sales, whatever that sort of at the at the line level for that team, their leadership, sometimes like one above that.

Sometimes it's my level and it's making sure that they're going through the weeds. Right. Like what's happened in the last few weeks? What's shipped? What are they planning? How are things going? Where are their blockers? And it's not a time for pretty presentations. It's a time for like, this is the real work. And the only time that meaning you can bring as much dirty laundry to that meeting as you have.

Like if things are not going well, it's actually the right time to talk about it. The only time that meeting doesn't go well is when the team, they don't have to agree, but they have to be coming, they have to be aligned, right? Like we have two very different opinions on this topic, but there's no daylight between us on what those opinions are.

We may not agree, but we all know, understand where everything is. Because what it forces to do, it forces especially as teams get bigger and, you know, sort of the guilds, so like engineering teams and product and design become sort of like also identities in themselves, it forces them that they can never diverge for more than two weeks, right? So they can kind of like drift.

And then, but within two weeks, they have to come back and be on the same page. The other thing is that it eliminates most of the pairwise conversations. You know, I'm giving the product folks some set of advice. Someone else is giving the design set of like same amount of engineering. And like, even if we think we're aligned, we're saying it slightly different ways.

And it just creates drift. Right. Because I think a lot about when the organization gets to a certain scale is like you're not in the room for every decision. So you need to make sure that the decisions. I mean, ideally, if you're doing your job well, the decisions are better when you're not there. Right. Like, you know, I hold my litmus test.

I hold for myself is that like are my leaders asking all the questions in those reviews that like I would or better. And if that happens, it's great because I can sort of step back and like, you guys got this. And that's where I talked about being able to impart that culture down in the organization. And I was actually talking with someone from another company last night.

Actually, I started building this overnight, which is a little bot that took all of our reviews and said,

Here are the questions I'm going to ask. Sadly, I'm wildly predictable, which means I was kind of disappointed that I could be reduced to a soundboard. But apparently it's true. And like the idea is like I want the line level PMs to be like, look, this is these are probably the questions you're going to get asked. Make sure you have answers for them.

What I think is like really cool, because I know, especially as I was coming up as a baby PM, you would you would obsess over this stuff. You're like, what? What am I missing? What am I thinking? You know, and then like you go into me like you after after a review, you just beat yourself up because like you're how did I not think about that? And you're like, as you grow up, you're like, of course these are the questions they're going to ask.

And it just, but being able to sort of like short circuit those cycles is I think going to be really, really cool. Fascinating. It reminds me of how I tried to scale myself at my company first before I hired the chief of staff. And I train that person to kind of be me when I'm not in the room and actually encourage her to make decisions. Even if it's exactly not my decision, I would be like, decision is better than no decision.

And then I can, you know, clean the dirty laundry. But that was helpful. And then now with AI. I'm on fire because I'm like you. I'm very predictable with some of the questions. And I found myself in meetings where people would do meeting for the meeting. And that to me is like the biggest letdown. So now I have my agent collect that information, answer the questions.

So hopefully people can prepare and we can discuss the insights instead of share screen and try to analyze the spreadsheet. No, I mean, and look, inevitably it still happens because of the You're going to find there's just there is just work and insights that happen as a group that you can't do individually. And like as lovely as it is to sort of like have an AI as a thought partner, it's it's still somewhat reductive.

Main page at 11:29

And so like but but it is really powerful because it's like all of a sudden I get to be a lot more places or at least a somewhat reasonable facsimile of me gets to be a bunch of different places. And so it's it helps me scale, which is great. It also allows on the flip side, I can sort of, you know, we spend a ton of effort. I look after our data organization as well.

building semantic layers and our own little AI analysts so like rather than me having to ping someone and one of the things about being a globally remote company is that like at any point in time someone's asleep right like we are you know and so like the person I need to have some information is asleep or vice versa like being able to ask that semantic layer a question about a bit of data or is and being able to short circuit that is amazing because those loops and it used to be, it would be like, oh, until you could schedule a meeting, which became wildly, you know, inefficient.

And it was like, oh, well, until they're on Slack. But now it's like I only need them to be awake to be able to have that information sort of flowing very freely. more about how you've built that semantic layer that is helpful not only to you but other people on your team we've grown very very quickly right you've heard the numbers you've talked about the scale of engineering entropy is inevitable one of our core promises is to be able to deliver an all-in-one platform where it's like you don't necessarily have to take all the pieces but like there is a better together thing right this is the classic you know, G suite office thing, right? Like what, what makes it

wonderful is that like these things just like naturally work together. And in trying to figure out how we could make teams still move as quickly as possible, but also sort of deliver on that process, we, Because there's two ends to that spectrum. One is everyone writes their own thing, which allows them to move wildly quickly. You get a ton of endpoints at the other side.

And on the other side, it's like, oh, everyone does this shared service thing, but you have to always take this dependency. I tried to uncouple those two. Because you want the upside of both, but the downside of neither. And so we kind of let that happen at one layer and then used, you know, like the DBTs of the world to kind of like create sort of aggregation.

So like there was a, there was an L2 that was how all the systems could talk to each other. And then very quickly we realized, but actually like to be able to understand that both for a purpose of our customers to be able to access their own data, but also for us internally, because it's like, it feels like an interesting thing because as you mentioned very quickly, we're going to be one of the largest employers on the planet.

if you look at the folks that are through our ERR platform, et cetera. And so we classically eat our own dog food. In fact, almost every product we launch, we use internally first. We're always our first customer. And so Being able to say like, okay, these are the questions that we need to be asking of our organization when they're like internal operational things or, you know, trying to answer questions for our clients are exactly the same problem.

So we built that semantic layer to be able to say, hey, how can I be able to ask questions of all of my data, knowing that the surface area is so broad that you're not going to know, like no one's going to know what table it is. Right now, we got lucky because we started doing this before, I would say, like the the most recent sort of evolution of the foundational models or the frontier models, because the the thing that makes that semantic layer also we're really good at explaining the thing that made a semantic layer good to explaining it to a human that is an internal resource or a customer.

was actually really powerful for an LLM, right? Because it's like largely denormalized. It looks like a spreadsheet, right? Versus a complicated data set. And so the quality of response that comes out of it is like incredibly good, which, you know, was a healthy reminder. We did this also with our knowledge base, which is like the inputs that you make, you know, I've been talking about this now for a couple of years, but like, There's like a new version of SEO.

It's like we all, we at one point had to rewrite all of the internet to work for search engines. We now need to think about all of our content, all of our data, all of our knowledge to be consumed by these LLMs. The thing we were like, we customized, we tailored it to that. And actually we've been doing a lot of that, especially with the corpus of global work and all the legislation and all the compliance rules.

And it's really powerful because if you know that you're writing for SEO, the LM, the risk of hallucination goes wildly down because it's like, it's consumable in a way that it expects. And so it's much more predictable outcomes. Well, at the same time, Aaron, I mean, you guys are working in a highly regulated environment across, and I think last time I checked, it was 150 countries, maybe more now.

Yeah. We're running out of entries. We're going to need to create some new ones. There's still a few we're not allowed to work in, which is for the best. They're usually on a sanctions list, but yeah. Yeah, so in those type of high-stakes situations, right, like how do you really ensure that the LLM has no room for error? There's a few things that we're doing, and we started this before LLMs were privacy safe.

Right. Like, you know, where you had a real chance of like it giving you, you know, someone else's data. And so we it was all at the infrastructure level. Right. Because at the end of the day, the LMS non-deterministic, you ask it a question a thousand times. It's not going to give you the same answer a thousand times. Maybe with some of the classifier models, that's going to be a little bit more predictable.

Main page at 16:30

I'm really excited about that, especially for our use cases like the Jets of the world. But it really is about structuring your input and making sure that it's, you know, what is the corpus that you can pull from? Because we have thousands of compliance experts in all of these markets, we've built just organically our knowledge base And we made an enormous effort to put this all like in a standard wiki.

It was like based off of Wikipedia in a very easily accessible way. And it's like, really, it's a source of truth. But because we created this corpus, that is what we focus the LLMs at. And actually now we're taking some small language models and actually training them on our own knowledge base in corpus. And, you know, we're at least as good as Sonnet and Chrome.

quickly approaching Opus in terms of quality on a small language model focused on the corpus of global work and compliance. And the upside of that is like second level responses versus waiting five to 10 to 15, 20 seconds and a fraction of the cost. And so, you know, what we're really seeing is that like the investment in that foundational work is actually allowing us to build a moat in an AI world, right? Because we now have our own model and it is performing really well, but it's also highly, highly reliable because it's being trained on a known good set of data versus it's trained on the universe.

And you're like, but try and only look here. But you can't enforce that and, you know, like pretty please don't do other things. That's a good point. You mentioned Jeff or the Jeffs of the world. And it's so fascinating now, right? Because we have these large language models, small language models, your own proprietary data. So how do you go about orchestrating this and still give you the flexibility to make additional changes as new stuff comes up? I think there's like two ways I think about this.

One is, so I started asking teams, you know, like we love talking about the wins. I was able to make an LM do this. I was making the new model can now do this. One of the things I've been I've I've really focused on asking my teams is like, tell me the thing that didn't work. Tell me the thing that the LM sucks at. Right. Because I think I find the teams that are performing really well, they understand where the edges are of the current model or of this approach.

right? And if you understand that edge, you are now saying this class of problems is really good to solve this way. This class of problems is not good to solve this way, right? And so the thing that I really like about, you know, obviously it's the new hotness. So like the Jebs of the world is that we were taking something that naturally outputs a narrative form.

Like it just creates text like it writes it writes back to you and trying to compute you know use deterministic software to do that right like obviously the coding agents were creating deterministic software but you know creating something where the output looks more consumable is rather than trying to like shoehorn you know we're going to give it sort of computational data into this modality where narrative form comes out and then try and recast it into computational data Like the amount of error propagation and single loss that happens along that is like just wild, right? It was like an unnatural, it was a square peg in a round hole.

But now that we're starting to see things that like output as computational data or there's other models that like are working on tabular data or stuff like this, it's sort of speaking the right language, right? It's the right tool for the right place. Yes, you can technically use... the, uh, wrench as a hammer or the screwdriver as a hammer, but it is like, sometimes you just want the hammer, right.

Or, you know, so it's allowing, allowing us to sort of like the, the tools that are fit for tasks, especially as folks get wildly excited. And, you know, there's like, I, you see this just, you just need to go to X on any day. Like every, every company is talking about how, you know, look at all the things that I was able to do with the new model.

And I'm like, you know, it's, it's, it's, I just joke like founders being like, I can now do 300% of our code using an LLM. Like, that's more code than you the math doesn't work that's about that's not how percentages work um and but like understanding where that edge is is what makes those teams incredibly comfortable because there's so much noise in the system about what these things in theory can do but like actually aren't great at yet and the thing is that they may be great at tomorrow or something actually we're also finding cases where it was good at something and now it's not There are regressions.

Main page at 20:57

There is drift. But understanding where that efficient frontier is is incredibly powerful at making sure teams are using the tools in the right way. It seems like not only you, but your team is allocating a lot of their time into using those LLMs, understanding the trade-offs. You guys recently published a stat saying that you found over $140 million in ARR.

thanks to this type of AI optimizations. And that's the equivalent of the work of 600 people that now you can do with agents. I'm very curious to know those specifics on how you actually connected those type of investments with actual ROI. It's picking up the change on the floor. Right. And, you know, the thing that existed in the before time, right, was your writing software, writing automation was an incredibly expensive thing.

Right. It was like software was the last artisanal craft. Right. Like everything was handcrafted. You could you could look at a piece of code and you can know who wrote it. Like it was it was like, you know, people that were like built wagon wheels. And that has changed. And so what happened is when it was still that artisanal craft, it was still a very, very precious commodity.

There were just so many aspects of the business that were still manual. Right. And they didn't need to be. And it wasn't because there wasn't a desire to. It was because there wasn't an awareness that someone is like manually moving this data from this system to this system or like reformatting the spreadsheet every month or go down that list.

Right. I have to go log into this one place and click this one button for, you know, because it doesn't have an API. And it just sort of fell below that. We hadn't gotten to it yet. But what this enabled us to do is first, how deep that list of things we were able to hit just got way, way deeper. But then the second unlock is we took it out of the hands of the engineers, right? Like how can people build these automations and these workflows for themselves and not necessarily inherently technical resources in a way that they feel really confident about? that one, their intention was captured and two, they could rely on the output and it was as good as them, right? And I

mean, and how would you do this if it was just humans, right? Like if you had an intern and you're like, hey, I have this thing I have to do every week. come shadow me, watch me do it, then show me how you do it. And then like, I'll check the results. And that's, we built off exactly how humans interact, right? So it's, that is how, and the product's called Akai, but like, that's how we built it.

So that one, how far down that list of like little things that were just wildly inefficient. And that includes the go-to-market function, right? Like includes the partnership function was like, I go back to, you know, it was like the marginal user theory, like who is the person who should be able to complete this task, but is like the most likely person to be able to complete it, but doesn't.

Like, what's that one thing that you need to fix? You know, the button, you change the name of the button, all of a sudden they're successful, right? Like, we were able to do that for a ton of operational things. And if we think about what is our day, it's all operational stuff. All of my, you know, a lot of my day is operational, a lot of your day is operational.

And so how much of that backlist of things can get? And the answer is like, there is a lot there. And while any one of those things may not feel like a big deal, if it was a big deal, we probably already would have automated it. We probably would already built systems, but they compound. Right. And so like the pennies add up and all of a sudden they become big numbers.

But the second part of that is that it allows us to continue to scale at a different efficiency. Right. So like the old scaling function is, you know, you would you would see the complexity would go geometrically against a linear scale or against a geometric scale. You scale exponentially in terms of complexity. But now that looks to be more linear because you have so much compute resources that are being able to sort of like keep the complexity down.

Resonate with what you are saying, because I remember when we sent our first AI manifesto and of course wanted everybody to use AI, we found the first bottleneck to be the engineering capacity. Because ultimately, people weren't fully self-sufficient. So at the end of the day, engineers would get more and more requests. And it was only until we made, we unlocked non-engineers to actually ship.

And I'm not just talking about code. That was huge. And yes, getting to the bottom of that backlog, even though I'm sure there's some bad ideas there, to me, it was worth to test because the risk of not launching that is much higher than actually keeping it in the backlog forever. Yes, and these were things that, honestly, they weren't even on the backlog because no one even thought to, like, mention them.

Oh, I don't want to bother you with that. It's just this thing I have to do, right? Like, all of a sudden, it was incredibly powerful. And to your point, like, yeah, there's some really terrible ideas. I didn't, you know, like, there's no good, there's no bad ideas. There actually are bad ideas. But, like, what's great is the cost of failing fast has gotten so low, right? You can have bad ideas.

You can go, that's not a good idea. We should make that go away and then make it go away. Right. Like it isn't that you need to be precious about it. But in those bad ideas were a bunch of real diamonds that you're like no one ever would have thought, you know, they don't want to raise their hand and mention stuff like that. But because you can just do this right, like it's it allows a class of people that wouldn't otherwise have raised their hand, advocated for this thing, make some really cool stuff because everything is now safe enough to try.

That leads me to the next point. You know how in the past, a lot of people would default to buy versus build just because you wouldn't have to maintain stuff. But now, given the cost of building is lower, we're seeing teams defaulting more to building over buying, right? So ultimately, the tech debt there is going to be higher. So as you deploy more agents, as you build more stuff, how do you think about that? So this is actually where those who do not learn from history are doomed to repeat it.

The engineering problems that we are living today are the same engineering problems that like someone making a printing press was figuring out, right? Like the cost of maintenance has always been true, right? Like it looks different now and our ability to sort of like, and you know, so like regular, like why don't we go in a weekend, build this other system that maybe hundreds or thousands or millions of engineering hours have been put through and experience and yada, yada, yada, which would be the buy case, right? Like, and you can get the happy path and like, yeah, in a weekend, I can build the happy path of any enterprise software ever existed.

My favorite business guru, Michael Tyson, also known as the boxer, which is every fool has a plan until you get punched in the face, right? Like the reality of this is hitting us now. It hits everybody. This is not unique to us. And this is not unique to this moment, which is you build a complex machine, it breaks. And so, you know, I didn't really appreciate this, but as a product leader, I've spent the last, I don't want to count the number of years, but like a couple of decades being able to think about how to frame problems, how to structure problems, how to communicate ideas to other product people, to engineers, to be able to build them.

And what has been wildly fun for me is I'm like, oh, wait, like this is now a superpower because like I can explain something to an L.M. better than a lot of other folks. And I was genuinely surprised when I found this, right? But like the work around structuring the question, the work around thinking about maintainability, the work around all of this stuff is just as true as it was before, you know, when we were building factories, right? Like when we're building actual machines.

Now, the thing is, is that the number of people, it's become a lot easier, right? So now everyone can build stuff, which is amazing because like so many great things are being built that wouldn't otherwise it's not a precious resource it's not it is no longer this artisanal craft but those folks are now learning to do this for the first time and so there is going to be a set of people that were you know we the citizen developer whatever you want to call it these days that are going to learn all of the lessons that we've learned in you know decades if not centuries of engineering right like Building that's easy, maintaining its heart.

And what is tricky there is the sunk cost fallacy. I build this thing. This is my thing. I'm going to keep iterating. I'm going to keep making it work. But sometimes it's like, you know what? This thing did what it was supposed to do. And you learn so much from it. And now it's time to do it again. Like start from scratch. Which I think is going to be really interesting in terms of like, the one way to solve tech debt is the rewrite.

And one thing about the rewrite is it seemed it used to be wildly expensive. And so it was like the worst idea humanly possible. Now it's actually, wait a second, maybe not that hard to do anymore. But like, the thing is, is how do you carry the knowledge forward? How do you carry, right, like all of those enterprise software things that you would buy? Yes, wildly bloated in terms of software, in terms of cost, in terms of all of these things.

but in them is also millions of insights. And so how do you take the insights forward into a build and a rebuild and a rebuild, but lose the tech debt, right? Like, so it's like, how do you get the goods, you know, get the cream and lose the rest of it? And I want to wrap this up Talking about profit, because you guys have been profitable for almost three years at this point, and that's rare, especially at a growth rate that you are growing as well, right? So typically, and you know this, people have to choose.

Main page at 30:02

Either you grow faster at the expense of not making money, or you become profitable at the expense of not growing fast enough, right? So how were you able to kind of pull both levers at the same time? We have the benefit of... truly incredible product market fit, which is not just you and us, it's just the work we're doing and the tailwind of, you know, a generational moment of time in terms of people thinking about global work, you know, between COVID, between expansion, between all of these things.

But the answer is like when we make when when when companies are often made with that sort of diametric choice, we can either grow faster or we can be profitable. They go one end of the pendulum or the other. right? Like you can grow, but how do we grow this profitably? Like, even if not in this moment, will this, like, what is the unit economics of this? So always keeping that in mind.

Like, even if you're like, look, this thing is wildly inefficient right now, but will it be, right? And because we have a big services component and an operational component in addition to software, software, you're like, oh, it's good. Once we get to scale, it'll be, you know, it'll be profitable, right? Like just belief, right? You know, and because it was largely true, but we also rarely did the math.

We always did the math. The second part is, If you start with a customer, it's really hard not to be profitable, right? Because like we're not building something in anticipation of a customer. We're finding a customer and building it with them. So you have your first buyer and it is incredible market validation. Now, this is a lot harder to do at the enterprise level, like SMB and mid-market.

It works great because those are people who are like willing to go on that journey with you. And so if you build along with your customers, what you end up getting is something that you know people are going to buy. And so they can be profitable. Now, then it's about cycle speed. And this has been like deal speed is one of our core tenants.

And so when you put those things together, you grow profitably. One, because you know that you've done enough of the math, even if it's inefficient at this moment, that it will be efficient. You can get conviction around that. And then you're growing with your customers. So this is never in search of revenue. There's always revenue to it.

And then you just say, okay, how fast can we turn that cycle? But we often have customers for things before we start building them. Your business, your product is very sticky, right? It's really the switching cost is relatively high at the same time. And I'm just wearing the CEO hat now. Yeah. When a company is trying to be efficient as well and leverage more agents than humans in some situations, what are the ways for you to still find product market fit so you are not just heavily dependent on number of humans hired for a company? Absolutely.

So two things. One, no company is as sticky as it used to be, right? The switching cost has gone down across 100% of things. So it is lovely, but I think You know, it's like really easy to get lost on the fact that like, oh, we're sticky because that's not going to be as true as it is in a month from now and six months from now and 12 months from now as it wasn't.

Main page at 32:46

You know, we're way less and every software is less sticky than it was 12 months ago. Right. And then it goes back to how do we solve this if we didn't have LMS? Because the thing about LMS, they like largely replicate human behavior or they approximate it. Right. And so if we solve that problem in the same way, like how would we solve this before we had as even before we had software? Right.

Almost none of those business processes have changed. Right. Like the process of like figuring out how someone needs to get paid, how much someone gets paid. Right. Like or how much tax you need to do or how do you apply for this visa or go down the list. Those are largely static and they've been static for probably 100 years. And the macro process is like generally the same.

Like there's like a bunch of information you have to get. You have to fill out a bunch of forms. You submit the forms to a government. They give you a visa. Right. I'm just picking one that's kind of universal. But what we now have is each one of those micro steps used to have to be done by a human or used to be done by an expert human and then used to be able to be done by a human with good instructions.

Then we're able to do software and now we're able to sort of like you know, have agents. And so, like, the thing is that you still, the end-to-end flow is almost not changed across it, right? Like, think about advertising, right? Like, here's the product I want to sell. Here's the creative I want to use. Here's the people I want to target.

Go watch Mad Men. Nothing's changed, right? Like, but what we have is we have new toys and new tools to automate that and speed it up. And so as long as you can keep conviction that I can build software that can do this thing a human does, and now it becomes way easier because an agent can do that much faster, right? Like, because there was a set of problems that were easy to automate, you know, move this file from here to here, right? Match this ad to this person.

But where we got stuck is that next class of problems, which was this requires a little bit of judgment, just a little bit, not a lot, just like a little bit. Well, like the format is a little bit off or like it could go left or it could go right. And so that's where those agents have been able to sort of unlock that next set of those sort of macro processes.

Now, we still firmly believe there's a human in the loop, like for what we do. And I think for what a lot of and the, you know, the compliance, the, you know, Someone getting paid the right amount is like not something you get to screw around with. And, you know, complicated medical leaves, complicated terminations, complicated all these things.

They're still going to be human there. Right. But what we want is that human is only working on the part that like is legitimately necessary for them. That requires their expertise, which is great because all those humans now are high leverage. They're being valued for their expertise, not for their ability to move stuff down an assembly line.

That's what I'm seeing with a lot of the real AI-nated businesses that are positioning themselves around now... this tool does the work for you or for the most part instead of I just have a hammer but I still have to go and do the heavy lifting the way that I like to talk about it is this doesn't it's not going to replace your job it's going to replace the worst part of your job like what's the what's the detritus of your day that you like do not derive any value for you do not add any value to but still need to get done that is what like give the agents that right like I wouldn't necessarily say, like, have it write your, like, three-year strategy and just take that as,

like, gospel. That would be an objectively terrible idea. Not to say that you can't, like, have it craft some of it and be as a thought partner, but, like, have it, like, you know, approve the expenses. What value do I add in, like, oh, yeah, this was the right amount for this hotel? I'm like, sure. Like, I'm responsible about it. Right. Like, you know, yes, if something's crazy, send it to me.

But for the most part, like these are these it's that level of judgment. Oh, well, you know, they were 10 cents above the limit. Well, you know, like trying to do something in New York and trying to do something in Lisbon are not the same. So, like, OK, we create one limit, 10 cents. It's fine. Right. Like that judgment stuff, which which relied on humans because computers were shit at doing.

computers are now like actually really good at doing because we have LLMs. The set of detritus that we couldn't do historically is now wildly achievable. I love that because I think, and I keep hearing a lot of people talking about judgment as the ultimate differentiator. And to be honest, sometimes it comes across as an excuse not to try and not to put LLMs to do some stuff that do not require that much judgment.

I mean, yes. And the answer is that there's a lot of stuff that requires a little judgment that probably occupies 80% of our day that can go away. Yeah. Aaron, it's been awesome to have you on the pod, man. Thank you so much for sharing how you're actually building from the trenches. Thank you so much. It was an honor and a pleasure. And I think the work you guys are doing is amazing.

So I'm glad that can be one small little part of it.