Retool CEO David Hsu: 130M hours of work automated in 12 months, targeting 10% of US labor by 2030

excited to finally talk to you. I was at your YC demo day at alumni demo day. And no one's going to believe this, but I it's true, so I just have to say it. and and you were the one company that stuck out and I was like that company's going to be amazing and I wasn't I didn't even think of myself as an angel investor.

I should have just been like let me please let me invest. Anyway, it's been fantastic to watch the arc of retool and everything that you've built. So, congratulations and uh thanks for joining the show. Thank you. Great to be here. Huge fan of TVPN. So excited to be here. Great to have you.

Uh can you give us uh a little update on the company? kind of uh define the different eras, what the product is, and where it's going. I know obviously we're in a period of transformation with artificial intelligence. I'm sure we'll go into all of that, but what has the bread and butter been for the last couple years?

Yeah. So, for the first few years of retool, retools like Legos for Code. Yeah. Which is we allow you to build sort of higher level building blocks and you can code faster. You could say you can piece them together and you can build software. Yeah.

And the weird thing about retool was that we always focused on this category that we called internal software which is extremely not sexy because no one ever thinks about internal tools but surprisingly is something like 50 to 60% of all the software in the world is actually internal facing. Yep.

And no one thinks about it actually. And so that's where started the idea behind that and and engineers are not like oh I want to work on the internal tool.

typically like you know you put the engineer that's not cracked you know and be like yeah just focus on this or just someone who's just like grinding away but yeah I mean this was the big we were talking to Joe Eisenthal about this with the the question of like meta training llama and there's a bunch of places where LLMs will will instantiate themselves in consumerf facing products across meta but also they have a massive amount of internal tooling that can benefit from from AI and so uh you know not having to fork over endless boatloads of cash to, you know, to just check is this does this have profanity in it one billion times a second, right?

It's like that could be a very big uh open AI bill if they don't have train that internally. And so yeah, I think the the the the dark clouds over the internal tooling world is something that most people might not be aware of, but you've obviously been living in that. So so talk to us about the growth of the company.

Uh what's the mix? Is this all enterprise driven? Uh, is there small and medium businesses that benefit from Retool? Kind of what's the bread and butter on the Well, if you look if you look at the homepage, it's like and the logos, it's like, oh, so you guys work with every company? Yeah, it's pretty cool.

I mean, we work from companies ranging from the US Army to the US Navy, the state of Utah, all the way to small startups, two person companies, five person companies. So, you don't get put on the American dynamism market maps. I'm gonna start demanding we gota put them on the American dynamism. Some respect on retool.

It is the backbone of the US military. No, seriously. I mean, there's there's internal software everywhere. Um, but yeah, may maybe it'd be good to shift into kind of like the AI moment.

It feels like retool was kind of vibe coding without the vibe coding meme or without the it allowed you to vibe code without the actual instantiation of you know you're not in an IDE but you're effectively vibe coding or building something that's uh very quick. Uh how have you been processing the the AI boom?

When did you first think about implementing LLM and other AI technology into retool and and how has it been going so far? Yeah. So the really cool thing about retool is that once you have these Lego blocks built. Yeah. You can actually give them to AIs to actually use actually which is really interesting.

Y believe is missing now which is today if you look at how many dollars have invested in the US in AI. I think it's around a trillion maybe a trillion and a half or so.

But if you actually look at how much revenue there is from all AI products across all companies, I think it's like 20 30 billion, that's pretty crappy ROI. Have to believe. Terrible. And the question is, is it all a bubble? Can we figure out some use case for this LM beyond just chat? Yep.

And if you look at where we are today in the consumer market, the enterprise market, it pretty much is all just chat. Mhm. Uh I think if you look at that 20 $30 billion of revenue, I want to say something like 80% of it is chat GBT revenue plus cloud revenue. Y which is awesome. Yep. But chat is pretty limited.

I mean is chat going to grow 30x from where we are now? It's hard to say. I think probably no is the answer. And so what we think is really missing is a way to actually leverage and use LLMs to actually go automate labor. Mh. And the weird thing about this is that LLMs are actually plenty smart already.

If you, you know, I use chat a lot. I'm sure you two both use chat quite a bit as well. It's basically AGI at this point. I mean, there's the classic Turing test thing of, you know, blew past that. Way past that. Yeah. And yet AI aren't doing anything yet. So, it feels like there's this big disconnect. Totally.

what we're here to really solve is can we actually allow AIS or LLMs to actually do things in your business. So right now it's just chatting back and forth. Can we actually allow it?

Can we allow the US Navy for example to say hey it's not just using chat GBT to answer some questions but actually chat GBT actually or LMS actually do things in the US Navy whether it's approving orders whether it's approving plans whatever it might be that I think is the next frontier for AI is AI that actually does things and I think we're almost there which is pretty cool that's what we're working on can you exciting help me work through this question of there's like this AGI ASI narrative And then like how does that not destroy every software company?

Because I I I'm seeing MCP servers spun up left and right and my question keeps being like if the AI is so smart, why does it need a server? Why can't it just use HTML and UI like any other human? Right?

um at a certain point is is it going to is there is there a world where I say I want to sell a t-shirt online and instead of spinning up a Shopify store it just writes payment interop code and it doesn't even use Stripe it just builds Stripe from scratch I at a moment's notice if it's so smart and it just works for a billion human hours add it up at at 160 IQ and it just builds me Stripe for this one t-shirt that I want to sell uh like that feels like it align.

That's also maybe not even the best example because of the regulatory component, but we also had Steven on from the co-founder and CEO of Lambda Labs.

His his point of view was like broadly that you're just going to generate the software that you need and you might generate 500 and and I'm sure this is stuff that retools already doing to some degree.

Generate a bunch of different versions of what a tool could look like, rank them, allow you to sort of try different versions of it before landing on.

So there's like this one there's this one world where like having retool primitives that LLMs can interact with is amazing at in terms of like making sure that there's robust performance and everything gets up to speed really quickly at the same time you know in the really long term do we even need these primitives and can we just do everything from scratch what's kind of your long-term view of how this plays out so this is I think a secret that we've actually discovered is in what cases do you want determinism versus what cases do not want determinism.

That's interesting. And we actually just announced yesterday that uh we have automated 130 million hours of work for our customers over the past 12 months. If you divide out the math, that's around 70,000. It's a lot of hours.

Yeah, I think the secret is exactly what both of you just pointed out there, which is in what cases do you want AI and what cases do you not want AI? And uh to give an example, OpenAI has this product called operator. Yeah, it's this agent like thing that does things on your computer. Yep.

And actually for consumer use cases, operator is really good. What up does is basically an LM with one tool and that one tool is use the computer. Yep. And if you want to go buy a uh shirt, if you want to go buy a pair of socks, that agent is really good.

It, you know, what it'll do is you say, "Hey, I want, you know, socks size medium uh a navy blue color. " It will open up the browser. It'll Google, I'll find the sock. It'll buy it. It'll use a credit card. It's done. And that's fully non-deterministic. You know, it's kind of making it up on the spot.

And that's pretty good for a consumera use case. Whereas for a enterprise use case or a federal use case for example, you actually don't want it doing that. And so I'll give you an example. Uh one of our customers, a large company, um actually uses retool for employee onboarding.

And so uh what they say is, hey, every time a new employee starts, you got to go do all these tasks. Maybe you have to go send them a laptop, you got to send them a key fob, you want to do this, you want to do that, whatever. And if you ask operator to go do that, operator says, I got one tool and it's web search.

So, how do I onboard an employee? I'm going to open my browser. I'm going to Google how do I onboard employees? Find a WikiHow article. It reads the WikiHow article. It's like that's not at all what you want. Business has very specific ways of onboarding an employee. Yeah.

And the AI actually calling those specific things because you actually don't want it reinventing the wheel. all the time and a lot of those might be gated and private and maybe not even on the open web and also very high risk from a regulatory perspective even.

So you don't actually want the AI reinventing it all the time. Sure. So that's kind of what we mean by the building blocks is you almost give AI these building blocks, these tools, these MCP servers and have AIs call them as opposed to AI reinventing it all the time. Yeah.

Because you don't actually want AI rewriting your security policy or how do I send laptops to employees every day. Instead, you want it to say, "Hey, oh, let me think. There's a big storm happening in the southeast. " So, for me to get the laptop in on time, I have to ship by express.

That's something you want the AI reinventing and reasoning about, but you don't actually want the AI reinventing, you know, do I use FedEx today or do I use UPS? What do I feel? You know, you know, you have a contract with FedEx, you want to use the FedEx one. So, Sure. Sure. Sure. Yeah.

And that's something that like a good office manager, good person in HR who's onboarding would actually think about and reason about and you could inject that with a reasoning LLM. That makes a ton of sense. Talk to me about the evolution of the business model. Um, consumption versus seatbased pricing.

And then going into the future, are we going to be looking at outcomebased pricing like what we're seeing Mark Beni off talk about in in terms of Salesforce where it's more like resolutionbased? Uh, you want a job done, our agents, our tools can get that done and you're going to pay for results. Yeah.

So this is a fun this is I think a dirty secret too is that I think outcomebased pricing is basically designed to rip customers off. Okay. The reason why I believe that is when Mark Beinov tries to charge you per outcome. He's like well what's the value of what I deliver and great pay me that.

Whereas in reality it costs a lot less for him to actually go deliver that value. That's software though we want 90% margins. Come on. What do you get against high margins? Well, for I hear you. Yeah. No, it it creates a But but isn't couldn't you argue with SAS?

The SAS vendors basically trying to charge as much as possible without the person saying, "Oh, we're going to build this ourselves or oh, that's that's actually, you know, we can just hire two more people to do this.

" You know, it's always this dance between you want to be you no company should be trying to capture all the value that they create, right? Yeah. because no one would buy a product. Yeah. Yeah.

So, the way that we're pricing is pretty interesting, which is that we price based on inputs, which is we just say our agents work for $3 an hour and that's it. If you want a smarter agent, you can hire a smarter agent. So, actually that $3 agent is a Deepseek agent. So, it's a I guess a Chinese agent for $3 an hour.

Or you could go buy 03, which is a uh quite smart, probably the smartest right now reasoning agent. I think that's something like maybe $120 an hour or something. So it basically depends on what kind of task do you want to do.

If you want to do a simple task actually deepseek is quite good at that at 3 getting a really good deal compared to hiring human labor. Whereas if you want something you know really knowledgeworky or something really nuance done maybe 03 is better actually.

But I think the innovation here is that we are charging on a uh per runtime hour basis. And what we've discovered is that an hour of runtime for 03 for example is actually worth something like 40 50 hours of human labor. You tried at least research can do in an hour. I mean I probably couldn't even do it honestly.

So that I think is really cool. And then what happens is you actually look at how much the customer pays per task actually completed.

If you compare you per hour our pricing to Salesforce's for example part ticket resolve pricing our pricing ends up something like 95% cheaper which is really really cool like if AWS charged you not on compute or storage if they charged you on how much money does your app make let's charge 90% of that that's a total ripoff right so that's kind of how we think we're almost you know we're trying to be almost like an a AWS of labor if you will yeah um how do you evaluate As the business has evolved, I'm I'm sure people early on it was pretty easy to communicate even to investors.

Hey, people spend a lot of time and energy and engineering resources on internal tools. We can be a big company just doing this.

Now, I imagine as you look at the business evolving, you're like threatening a lot of different categories of software because you're saying we're going to enable teams to more easily make the decision. Do we buy this or build it ourselves? and like maybe retool is like some middle ground.

I don't know if that's the right way to think of it. Um but how do you evaluate kind of the scale of the opportunity now? So the internal goal that we have and it's pretty early but our internal goal is to go automate the equivalent of if not actually 10% of the labor in the US by 2030 is our goal.

And the idea is that Let's go. I love an ambitious goal. I love it. I love it. And here's the TAM. Yeah. US labor 10%. 10%. Yeah. And actually the cool thing is we're on track actually. That's amazing.

Over the last few months, if you draw that now, yeah, 2030 is a long time away, but you draw the line out, it comes to I think% actually. So that's incredible. It's pretty cool. But 130 million hours automated is no joke. So yeah, serious. Yeah, that is absolutely wild. Well, well, 15 minutes was not enough time.

We have three more minutes. We have three more minutes. Our next guest at 120. I have one more question. Um, uh, talk to me about the knockout dragout fight happening in the Foundation model space. You mentioned Deep Seek at $3 an hour uh, versus 03 at $120 an hour. How do the other Foundation labs compare?

What do you do to benchmark them internally? Are the benchmarks cooked? What's the take on Meta and Llama? There seem to be a ton of energy there in terms of your use case.

Having an open- source model that you can inference for cheap or at cost seems incredible and yet they've it seems like they they're a little bit behind on the reasoning models. Uh are you optimistic there? Uh are you kind of model agnostic? Um what's your overall take on the foundation model space?

It's pretty cool to see this play out especially with our customers because yes so for us we're theoretically agnostic you can use whatever model you want. Sure. But it's really cool seeing what customers prefer actually. Yeah.

And uh so for example with a you know large federal customer you might think oh you know actually onrem is really important and so they actually prefer to use llama you know or something like that. Sure. Not the case. Actually you know open a selling a lot of open AI.

Turns out actually there's a lot of traction even in sort of big federal agencies for something like open AI and they actually have contracts already and so they plug it into retool and just works. That makes sense.

You know I think I would have thought if you asked three years ago is the federal government happy giving all their data to you know LLM. No. Hell no. It seems very unlikely but it's happening which is really cool.

But you also see this obviously happening in other countries too where like whether it's Saudi Arabia or Europe with mistrial for example people are getting kind of nervous about sending data to other countries LLMs and so it's going to be cool to see how the world plays out.

Uh it's surpris it's very much not a meritocracy is maybe one way of putting it which is we thought that when we build our agents product with an eval framework people would just say let's see who does best who does it for cheap business let's go and actually that's not the case which is I don't know how I feel about that honestly is that because like there are SDRs that are buying steak dinners for people and swinging them over or is it more like there are qualitative me metrics that matter more like the nature of the contract matter matters more than just benchmark price, etc.

Like the like the the quantitative metrics. There's that. But maybe another way of putting it is LLMs have been stickier than I thought, which is maybe the branding is so important or something like that. Yeah. Everyone just says, "Oh, it's just one line of code to swap out.

" But if you're used to the certain like what does Underpathy call it? Like spiky intelligence.

There's like certain models that spike in certain ways and you have expected behaviors and they Yeah, they might all hallucinate at the same 3% rate, but if they hallucinate in a specific way, you build around that, that type of thing. Okay, exactly. Yeah, very cool. Cool. But excited to see how the space develops.

So, well, thanks for coming on. This is fantastic. We'd love to have you back and and just chat about AI and and automating 10% of labor. Good luck. Give us give us access to the internal tool that's just the autom the labor automation tracker. I'd love to just follow along.

we can put it up as a ticker on on the screen. It's just like progress, you know. No, I I I love the the ambition and and excited to follow it. Come back on again soon. Yeah. Thanks so much for stopping by. We'll talk, David. Bye. Uh let's talk about sales tax AGI. Numero, put your sales tax on autopilot.

Spend less than five minutes per month on sales tax compliance. Go to numeralhq. com to get started. If you're selling anything, software, hardware, e-commerce, whatever you're selling, get on Numero. makes it really easy.

Monitors where you're selling, how much you're selling, whether you need whether you're approaching that threshold. Start filing and does it for you automatically. Cuts it down to five minutes per month. That's brilliant. Also, get on public. Public. com. Investing for those who take it seriously.

Multiasset investing, industryleading yields. They're trusted by millions, folks. Double kill. Double kill. I love a double kill. Thank you to public. Anyway, uh our next guest is from Pallet. We have Shashant in the studio. Let's bring him in and hear the latest update. How you doing? Doing