Wade Foster on how Zapier scaled profitably without raising another round — and why agentic AI is still too unreliable for complex workflows

guest to talk about building a business without burning a ton of cash because we have Wade Foster from Zapier Zapier here in studio. French pronunciation.

I you know you were in my YC batch in 2012 and and I called it what did I call I called it Zapier like a rapier for a long time and then and then people were telling me that's wrong and so I started calling Zapier and then I learned that was wrong. But it's Zapier like Snapier, right? Yeah, Zapier. It's funny.

Snapier was actually what we called it first. We Yeah, we had to switch to Zapier because there was, you know, uh, copyright issues like and of course it was one Snapier with one P and so we just doubled down on the mistake and went Zapier with one P as well. It certainly hasn't hurt the business.

You've you've been on an absolute tear.

uh we've enjoyed uh talking to your co-founder a ton about uh ARGI and that stuff, but we'd love to get uh I I just into the rest of your world and what's going on with the business and um do you have any advice for uh uh current generation startups who are maybe playing a little bit fast and loose with the accounting?

Um did you ever have gross margin problems? Was there ever a moment where you were like, "Okay, like we have customers that are actually making so many calls to our APIs that we're losing money on them. " Or was that just never an issue? We became profitable in 2014, so basically two years in. And congratulations.

When we raised the uh Yeah. When we raised our seed round out of YC, yeah, we said, "Hey, uh we want this to be the last money we ever get. " Mhm. And so I I like looked at every expense on our P&L and approved everything and was just like vigilant about not spending money if we didn't have to.

And um you know I think in 2010 like that decade we were looked like as crazy. Totally. And now it's very trendy. Like in fact I heard uh this term popularized I don't like six months ago seed strapping. Seedpping where it's like yeah one round and then you know don't raise more. Okay. So that's what we did. Yeah.

Did you believe it at the time? Uh were were there specific uh entrepreneurs that you were looking to as like mentors? Like where did that idea even come from?

Because the meme in Silicon Valley at the time was much more on this on the Reed Hoffman blitz scaling raise a ton of money and every venture capitalist would echo that because they want to put dollars to work.

Uh and so there was definitely like a a mood in Silicon Valley of like hey it's more than okay to raise an A B C D burn burn win. Like did you believe that when you said it or did it come from somewhere like talk to me about the the thinking at the time? We had two examples.

So we founded the company in central Missouri. All three of us worked at Veterans United. This is a company owned by two brothers 50/50. They never raised a dime. I was employee 500 and when I left 10 months later there was a thousand employees there.

So this company scaling like crazy with mortgages during the financial crisis. Wow. And so we see that and are like, well, clearly you can build a successful company this way. Then one of our favorite products, Mailchimp, they were one of like our earliest partners that was most successful, entirely bootstrapped.

And then coming from the Midwest, we're just we just don't know anything about like this Silicon Valley world for how to build companies. Yeah. Yeah. And so it kind of makes us skeptical of anytime, you know, a venture capitalist is Yeah. I remember one said, "No important company has ever been built this way.

" And you know, when you make grandiose statements like that and you have counterfactuals, you're just like, "Well, what other advice are they saying that might just be wrong? " Yeah. Uh and it doesn't mean that all the advice is wrong.

But I think we just ran it through a thicker filter than most people who are like, "Well, I guess we're going to raise an A and a B and a C. " But I think it's important it it doesn't my read and I guess what I've heard over the years is it's not like you said we will never raise another dollar.

It was more so we we hope to just make a great business and not be forced to. And I think the issue when I hear people talking about seedstrapping it's like seedstrapping I don't know if it's like the right goal.

It can be something that you hope to be able to achieve, but if you're raising a $5 million round and you're telling everybody, "Oh yeah, we don't want to raise more money. " It's just like not the right oftentimes not the right thing to be fixated on.

It's like be fixated on like we want to use this money to find a really sound business model that will allow us to scale efficiently and we will tap capital markets as needed is like a better goal than oh we just don't want to raise more money and because I usually feel like that comes from founders that are like you may as well just say I don't want to dilute myself anymore and that's not necessarily the right thing because ultimately if you want to create value you should just make the share price go up and like there's some companies that need to dilute a lot in order to create more value.

I think the exercise to go through is to look at your growth and really understand where is there a bottleneck in your business. What is the long pole in the tent that is holding you back from achieving your goals? And every time we did that exercise, we came back with an answer that wasn't more money.

it was management horsepower or uh you know we need this particular feature to go launch which we can go build like there was just other things that were the thing to go address to help us grow faster and the place where I see folks making mistakes is they get caught up in the noise and the logic will be like well so and so raised a lot of money or what if something what if this happens or what if that happens which are not actually grounded in factual truths of reality.

What you do know is this is where my company is. This is what it'll take to get there. And if my bottleneck is money, great. Go get money.

But if the bottleneck is something else, you as the CEO, like it's incumbent upon you to spend time on the most important thing instead of go chase another thing that actually is is not the most important thing. And so that that's kind of how we always thought about it.

And you know, I think it raising more money never was the most important thing. And so we spent our time differently.

Was there ever a moment where you blinked like you know halfway in or a few years past being profitable where you were like ah like this idea maybe it's worth you know dipping in raising some money getting back on the train or was it just like okay now we've been going for a few years we're good well once you get it going it it actually gets harder to want to raise more money because you start to look at your growth rate and going you know what like why would I raise at this valuation when I can just wait 6 months.

Yeah. And you know my revenue will be up by 50% or 100%. Like the valuation just never makes much sense. Now obviously and this is I think why you see a lot of like really runaway valuations right now because the founder kind of just looks at the math and says well it's there's no there's no sense.

And so a venture capitalist has to come in and be like, "Well, I'll give you 40x or 80x or 100x to get them to go, well, okay, maybe for that I would do it. " Yeah. Yeah.

It it it's very funny that we're seeing in the market right now like like tough margin stuff, crazy burns, crazy valuations, and then simultaneously we have another company in the AI space midjourney, David Holes, that hasn't raised any outside capital and is just growing, growing, growing.

And so it's not like it's something unique to the to the like the fundamental structure of AI like there is a there's clearly a way to do it without raising an incredible amount of money and yet uh not everyone's on that path. Everyone seems to be on the path of like burn burn. I don't know. Yeah.

Well and I think it depends you know like some of the the foundation models obviously have huge capex and uh expenditures and so that's a different beast you know whereas if you're at the application layer your cost structure obviously looks a lot different.

So, you know, this this advice, this way of building obviously has to be filtered to your your context. It doesn't make sense for all companies. Yeah.

Was there something about your particular business or where you fit into the landscape where you it wasn't it wasn't positioned as a product where on on day, you know, 200 AWS is like we're competing seriously. uh or you had like a million competitors immediately.

What like was it less competitive because of the position in the market and kind of like how you how you shaped the product? I like the product felt like I don't know somewhere in between like a developer tool. It was almost like a new market that felt less competitive. But I don't know what your experience was.

It definitely was an odd product. Like we had folks like comment to us like how exactly did you come up with this idea? And I think partly what made it so defensible was integrations middleware is just inherently sloggy. Like there's not a lot of secret hacks to do that.

You know now with AI it's getting easier but it's still not easy. You still just have to like you know just kind of chunk away at this stuff.

Like I would put us plaid others kind of in this category where it's not actually easy for a big company to just decide oh we are going to go compete in this and most of them just don't have the fortitude to kind of keep chipping away at a bunch of like really boring engineering work like they are hiring all these engineers that want to work on more sexy things.

Yeah. Um versus chip away at integration code all day every day. Yeah. Now you kind of have um like such a platform that you're in I feel like you're just like a natural beneficiary of the AI wave um because you're already plugged into so many customers, so many integrations, so many data sources.

There's already so many primitives that AI can take advantage of. Uh how there's two areas that that stand out to me as potentially valuable. I'd love to know which one's more valuable to you and then if there's a third that I'm not thinking of, let me know.

Um, first is just having AI co-pilots help write new integrations, help accelerate your software development, just do more with the team. I doubt you've done like massive AI layoffs or anything like that, but you're probably being more productive. That's maybe one area that is interesting.

And then the other is just AI as a product that can be drawn on in the midst of moving data around. So if I have one connection over here on Zapier and then I want to have an LLM transform that data and then send it to another integration that feels like a really like really amazing use case.

Um have those both been equally beneficial? Has one been has one stood out to you as uh as particularly great or is there something else that I'm not thinking of? They're they're both really valuable for different reasons. Yeah.

So the first this idea of like a co-pilot or text to workflow is really valuable for helping people build new types of workflows. Sure. Inherently like these noode workflow editors have a fair amount of configuration involved with them. You know, we always invested a lot of in design and making that experience good.

Yep. But it's still a lot to configure. And co-pilot just gets closer to how humans think about this stuff. Most humans are not great systems thinkers. They don't break the task down bit by bit by bit.

And so as a result, a co-pilot can be a really good way to attract folks into the market that maybe struggled in the past, but it doesn't fundamentally change the use cases that folks can do, which is where the second one gets really interesting.

And I think the this is where we've had some really interesting learnings because we have launched Zapier agents which is a purely agentic experience and we have our classic workflow editor where you can now inject AI steps.

Y now both have been successful but the latter has been far more successful and I think it's because when you look at the um and the latter sorry to clarify the latter is in in this case is workflows workflows AI steps. Yes. Exactly. Mhm.

And I think what the reason why is because agents inherently just don't quite have the reliability yet to get you to solve like sophisticated use cases.

You know, you think about it where it's like, well, if you're asking it to do anything complex, call it like 10 tasks in a row, 20 tasks in a row, but the the error rate is 10% every time.

Well, by the time it, you know, rips through 10 or 20 tasks, like it's messed up on a handful of these things and it's way off in left field.

Y whereas within a workflow you can put it in a harness and say hey mostly I want this thing to be deterministic throughout all these parts of the steps but right here I need to to summarize something or I need it to analyze the sentiment of this or I need it to generate an email.

Uh, and you can put the agent part, uh, the AI part in the spot where it does the thing best and then you wrap it back in, fold it back into a deterministic workflow. And that has opened up, um, a huge amount of usage.

We have like a quarter billion, more than a quarter billion tasks that have run through Zapier with AI like that alone. And I think that's the honestly I think it's kind of a sleepy category. Yeah. When did when did you guys realize what a good position you were in here?

Because you basically spent before agentic workflows were like a term that people were throwing around every single day.

You guys spent a decade basically building out all the really really hardcore integrations with hundreds or thousands of different services and getting just simple workflows dialed so that you can yeah you can just slot in, you know, reasoning or or slot in LLMs at different at different steps.

But um have you been surprised at at at how many people are coming in to kind of like compete in this space? Cuz it feels like I've just seen a lot of pitches over the last 12 months that are I just look at and be like, "Okay, this is like this is Zapier but just like, you know, pretending that Zapier doesn't exist.

" Like but AI first, right? Yeah. Yeah. Yeah. AI first. And it's like, well, you're it's Zapier, but you can buy the stock in the private markets. Well, it's it's Yeah, it's Zapier, but Zapier, but we're doing a series B. There you go. Zapier, we sell software. We're Zapier, but we sell stock. Yes, exactly.

There you go. Um, yeah, you know, I I do think it kind of surprised us. um that we were as well positioned as we were. You know, we had started to dabble with generative AI a few months before Chat GPT had come out and um had started to see some customers building workflows on our private developer platform.

we have a developer platform where folks can build their own stuff. And so we'd started to see some of that happen. And as uh the models advanced, in particular when GPT4 came out, we were like very unsure if workflow was durable.

Like we started to think like actually this like pure agentic experience is probably going to be the future. Now I still actually believe that it will be the future. I just don't think it's going to get there as fast as um we we think it will. Yeah.

And you know, over the course of the next year, we just kept seeing a gent like AI use cases on the traditional workflow builder grow and grow and grow and grow. And then we started to realize, huh, like maybe what we need to do with our product roadmap is actually help people surf the wave.

like help them go from just old school classic deterministic workflows and move steadily into purely agentic ones uh and allow them to kind of mix and match these experiences along the way. Talk to take it back to gross margins for me.

um how important is offering the full suite of paro frontier models within those uh deterministic steps and is there a world where you want to offer even smaller and more focused and cheaper models?

Um we we talked to one company that's training uh uh transformer-based models but just for sentiment analysis just for uh for censoring bad words just for uh JSON to CSV or something like stuff that could almost be done deterministically in many cases but just like you need a little bit of uh of AI there and they train them on GPUs or on on like gaming graphics cards and they run like super super cheap.

And I'm wondering like uh if you're if you're in a high if you've built out uh uh if you build out Zapier workflows that are high volume, the cost per inference feels relative there.

And I it feels important and I feel like you would have a good insight into like how some of these like bigger models can kind of blow up on you if you're not careful. Yeah. Well, this is also where it's nice if you can encode your workflow into something that's deterministic.

you get reliability, you get cost advantages, you get speed, you get a lot of things that are really really nice. Uh and then uh you know insert AI only where you need it. Now the way we've thought about this is we really want to abstract as much of this away from the customer as possible.

you know, we sell predominantly to folks in GTM orgs and GNA ors and um folks that are not paying attention to the subtle differences between, you know, which model is best for coding, which model is best for writing, which model is best for this, that or the other.

So, we're trying to abstract away a lot of those choices. And that extends into how we price for this stuff as well, too.

So, we want to make it a little bit more like you just buy it, we charge you for tasks, doesn't matter if it's an AI task or another task, we'll just sort of it all the math will kind of all work out for you in the end.

Um, but I imagine if you were selling, do you guys have a lot of do you guys have a lot of learnings from like pricing on on pricing based on on out outcomes versus traditional SAS?

Because that's again so many of the ideas that VCs have gotten excited about with agentic workflows and just agents broadly I feel like are ideas that you guys have explored. You've basically always been consumptionbased. Correct. We've always been consumption based. Yeah.

And I think the challenge that and sorry to interrupt but the other way to look at that is like uh you've been pricing based on work done. Yeah.

is is like the the sexier way to describe that at least 202 uh we do a task like you pay for tasks and I think the challenge that exists for horizontal products is the way in which folks use the software they they get variable value from it and high usage doesn't necessarily correlate with high value and low usage doesn't ne necessarily correlate correlate with low value.

And so that's where, you know, if you're building horizontal usagebased software, you run into a tra a a trap where you're like, hm, how ex like where along this spectrum are we going to price? Yeah.

Are we going to pick one credit system where everything is going to cost one and we're trying to figure out well is it a premium thing or a is it a more commoditized thing or do we risk making our pricing more complex and start to say we're going to have a credit system that's tied to different value but now your customer is having to do a bunch of different calculations in their head figuring this out.

This is where I think vertical agents have an advantage right now where they can say hey we are hyperfixated on one use case and as a result we can maximize the value that we are going to sell for.

We're going to say hey you know if you're in customer support we're going to charge you if we close the ticket and you can do very direct math to say well today I have a support rep that I got to pay this much money every single year to answer this many tickets.

if I can turn that into an agent that they can answer those questions just as well, then I'm willing to pay, you know, this this is sort of like the maximum I'm willing to pay for this.

So, that's where these vertical software do have like pricing advantage, whereas if you're charging horizontal, you you have a much trickier challenge because you have so many variable use cases at the end of the day. Yeah.

Are there new uh new markets that you're noticing open up that are uniquely available to you because of AI?

I think the thing that has surprised me is the c certain like more traditional or like industries have been fairly quick to jump on this and I think my hypothesis is partially that their use cases are well suited for this.

So if you have if you're working a lot with uh like PDFs and like paper documentation, the SAS stack was like kind of tough for you because it had to be structured and there wasn't tooling that could turn all this like unstructured data into structured data for you. However, LLM are uniquely good at that.

And so as a result, these companies have seemed seemingly jumped on this stuff pretty quick.

The second thing that's been interesting is we've watched um particularly media companies be relatively fast at this which also surprised me in some ways I felt like they were faster at it than certain tech companies we've spent time with this and I have to wonder like how much of this is because they saw so much disruption in their business model during the internet craze that they were like we are not going to let that happen to us again.

Whereas technology companies have kind of had it pretty cushy. Yeah. And so like, yeah, we're we're doing the AI thing. It's fun when you're doing the disrupting, but getting disrupted. Yeah. Not so exciting. And so they're like, act two, we're going to be on it this time around. Yeah.

Uh where where do you stand on uh on estimating the impacts of AI on the job market? because I feel like a decade ago or around the time when you were getting Zapier off the ground uh Zapier, sorry, we said it both ways.

Um I'm sure people were saying like uh automating business processes will lead to with software will lead to job loss. And did people say that? I guess. Yeah. I mean people said that with the internet. They said you're not going to need sales reps anymore.

Everybody's just going to be able to go and you know you're just going to be able to go to a website. Um, and I just, yeah, I'm curious like, well, there's two data points that we've been pointing at.

One is, uh, the Wall Street Journal had an article today saying that, uh, new grad unemployment is, uh, is not tracking as well as it should, and there might be a real effect of of AI not driving companies to hire less new grads.

And then but then on the flip side, there was this report that came out that said that, you know, 95% of AI tests at um at Fortune 500 companies did not make it past the test phase. And so AI as an installation into these large companies is not going as well.

So there's kind of evidence on both sides, but I' I'd love to know what you think. Yeah, you know, I when I think about this, I my my take is that capitalism is largely undefeated. Mhm. Uh, and you know, it's it's it's easy for us to spot the disruption.

We can see the jobs that are disappearing easier than it is for us to see the new opportunities that are being created. Sure. But history will tell us that there is always new jobs created. Y and this is what we see play out in our customer base.

Um, I talked to small businesses over and over where they're like, I I I never could have hired somebody to do this work and now because I'm doing it with automation, I'm actually getting more done.

I'm generating more revenues and now I'm reinvesting that and when they go to reinvest that, yeah, some of it's being reinvested in AI and more automation, but some of it's being rest invested in people. So, that's kind of where I think this plays out.

And the second thing I think will happen is there's this meme that goes around where it's like, well, all these software companies are just going to have crazy high margins now. Like, we just don't need that many people. But again, capitalism is undefeated.

If someone looks out and sees like, "Wow, you're generating so much value. How about I just come in? I'm willing to tolerate. " Or if everybody everybody has high margins and they're like, "We should just spend three times as much on user acquisition and just get more scale. " Exactly. Exactly.

And well, how are you going to get user acquisition? Well, some of that's going to be invested in advertising. Some of that's going to be invested in people. And like I think at the end of the day, it nets out where, hey, we're all we're all going to still have jobs. Yeah.

But what jobs we do have, that's where it's going to shift. And I think those reports are kind of telling that story where new grads, the requirements are shifting.

what's what's needed inside the work force is is changing and so there's kind of a shift going on there and that's always what's really challenging in these moments is time is the skill sets required to be successful they're not static yeah what about uh job creation around the implementation uh like I'm sure there's been a cottage industry around Zapier for a long time now of people that just help integrate Zapier in companies um and we we were talking probably uh at the beginning of this year thinking that what what an amaz you know in the same way there used to be a big opportunity to just help companies run Google ads or help companies run Facebook ads and that there's still opportunities there but it feels like right now taking Zapier into a company and just saying like I'm going to save you guys uh you know $250,000 a year if you pay me 20 grand to like set this up properly.

It just feels like that is extremely repeatable and there's so many businesses that as for as many that are coming on and and and using Zap year or other platforms there's there's probably a multiple of those that just don't aren't even fully aware that these tools exist yet or how to implement them.

We so 100% we're seeing the rise of the AI automation engineer. We have some of these internally where for our enterprise customers if you want us to do it for you, we will come do that for you. And to your point, we're seeing this cottage industry starting to blossom.

And it's particularly effective when you can combine that skill set with domain expertise. Yeah. So if you have somebody in HR who has this capability, they're able to work magic on this.

Same thing in legal, finance, marketing, sales, if they have this systems thinking, this ability to break down tasks to to use the tools, they're often able to work wonders inside the stack.

And it's been pretty exciting because we've this this person has kind of existed for all of our this is who we've served for all of our 15 years, but in the last couple years they're starting to get more um credit inside these organizations.

They often were like the underappreciated person who just kept all the plumbing working. Um but now like they're like the automation engineer like I'm making magic happen.

Um, can you give me your take on MCP and how it just explain it and how it fits in and is it a tailwind or a headwind, an alternative, something that you're just benefiting from immediately? Um, how does it all fit into your landscape? Yeah.

So, MCP model context protocol, it's it's a protocol much like HTTP or or REST or something like this that allows uh it focuses on how agentto agent communication, letting agents talk to each other. Um, I think right now there's obviously a lot of hype around it, whether it's a tailwind or a headwind or both.

Um, you know, we're still figuring it out. Like I think there's a lot that has to be figured out with the use cases and whatnot. The I think the most important thing is to actually get the the use cases really dialed in.

Uh, and the place where we're seeing it be most effective is a in like Claude, uh we're seeing a lot of folks that set up uh Claude projects. They give it access to various tools and then they go deploy that tool like an internal tool for their organization.

So, for example, if you wanted to create like a uh a sales briefing tool for your sales reps, you might say, "Hey, I'm going to set up a project. I'm gonna give it a list of in a a template for how I want the brief to be created, and I'm gonna give it access to HubSpot. I'm going to give it access to Zoom info.

I'm going to give it access to like deep research through Chat GPT or Perplexity or something like that. And now my sales reps can just come in and say, "Hey, I'm talking to, you know, uh, Sam Alman tomorrow. Like, what do we know about our deal with OpenAI? " And it goes and does deep research.

It fetes everything you have from your internal CRM and it generates a brief for you. So, you come in way more prepared.

So those types of like internal tools in the past those were like kind of hard to build and MCP is like making it really easy for folks to build those kind of like chat style uh internal tools really effectively. Um that's the use case that we're seeing pop up most frequently.

But I I have to imagine that as time goes on, we're going to see a a lot more like we it very much feels like we're in the very early innings and we haven't, you know, seen these like new user experience paradigms like launched yet or popularized yet. Yeah. Yeah.

It always it always just felt like to me with uh just the just on a on the basis of inference cost if you're hitting an MCP server a million times, you would rather just have your agent like write an API integration and just go direct to the database for most of those things.

But there's probably specific use cases where it makes sense, especially if you're like dancing around a lot of different Well, I think that's I think you're correct there.

Like at the end of the day, there's going to be some use cases where you want, you know, an API call and deterministic code because it's faster, cheaper, more reliable. And other cases, yeah, we saw this with CHP pretty early. It was like, yeah, it's pretty good at math. It can memorize a lot of math.

But like, why not just give it Python and then it can get the math 100% right. Like, I don't care. I just want the right answer. You don't need to like stunt on me by memorizing every number of pi. You can just go look that up on Google if you want. Anyway, that's fine. That worked out pretty good in the past.

Yeah, exactly. Uh, let's stand on the shoulders of giants as much as possible. Uh, anyway, this was fantastic. Thank you so much for hopping on. We will talk to you soon. Great to chat. Thanks, guys. Have a good rest of your day. Cheers, Wade. In the meantime, let me tell you about Bezel.

Your bezel concier is available now to source you any watch on the planet. Seriously, any watch. You can go to getbzzle. com. And we have Dan Wang in the reream waiting room coming in to the TVPN Ultra Dome. He is releasing a new book today, I believe. Um, it's already being reviewed.

Gotta ask him how many pages it has. Is that the question you want to ask? And hit the gong. Okay. Yes. Yes. Yes. Definitely. So, uh, the the what everybody's been waiting to hear. How many pages is your new book? Yes. Yes. Yes. He's a research fellow at the Hoover Institute.

The book is Breakneck, China's Quest to Engineer the Future. I saw a uh a video of him going viral talking about the difference between uh uh China's government