We're going to be talking about what we call the age of agents.
Oh, the age of agents. That's what we coined. I love it.
Agents, everybody. Agents has lost all meaning at this point in time. Everything's an agent. These tools enable you to go faster with less capital. You can build a company that gives you more control. Everything just moves really fast right now. You can have hypergrowth like high-performing companies that are bootstrapped. We raised a million in change post YC and then we got profitable after that and have never raised any primary capital since then. I'd be lying if I was like, man, it would be nice to have just a little more time to think. I see some of these reasoning models, they get to think for like an hour now, and I'm like, man, that sounds pretty great. Can I think for an hour?
Hey everybody, it's Sam Jacobs, CEO of Pavilion. Welcome to another great episode of Topline. I'm joined by my co-host AJ Bruno, who's the CEO of Quotapath. Asif Zaman is not available today, but we do have a very special guest on the show, Wade Foster, who is the CEO and co-founder of Zapier. Now, before we get into everything that Zapier does, I want to set the stage for this episode. We're going to be talking about what we call the age of agents. Now, depending on who you ask, agents are either about to remake how work gets done or they're just the latest layer of hype on top of automation and AI. The reality, as usual, is probably more nuanced. The real question for operators and CEOs isn't whether agents matter, but where they actually work, where they don't, and what changes inside companies as they become more real. And to do this, Wade, I'll give his bio and then he'll introduce himself. But Zapier, if you don't know Zapier, you're living under a rock. They're the most connected AI orchestration platform. Zapier connects teams, tools, AI, and agents across more than 8,000 apps to automate critical workflows with flexibility, scale, and control. Prior to Zapier, Wade worked in marketing at Veterans United and Idea Works. He has degrees in industrial engineering and business administration from the University of Missouri Columbia. Wade, welcome to Topline.
Yeah, thanks for having me, guys.
Wait, don't you love how Sam said we call agents as if Topline coined the term agents? We call these agents.
It was the age of agents, AJ. It was the age of agents.
Oh, the age of agents. That's what we coined. I love it.
Agents, everybody. Agents has lost all meaning at this point in time. Everything's an agent.
I think before we dive into the heart of the conversation, I do think there are probably a few people that haven't heard of Zapier. Some people think it's pronounced Zapier. I'm pretty sure that it's not. But tell us a little bit about the origin of the company, how it emerged and how it's gotten to where it is today because I don't think you started with the description that I read in your bio, which is the most connected AI orchestration platform. But maybe that was the vision all along. So walk us through the origin story of Zapier and then we'll dive into what's going on in the world today.
Yeah, you can see bits of it that most connected bit like that was there at the beginning. In 2011, basically what we observed was that if you were to go to the help site of any popular SaaS product of the time, there would be this type of question that was commonly asked: 'When are you all going to integrate with X?' And you would read the replies, and the replies would be a bunch of customers saying 'plus one, me too, I want this as well.' And eventually a product manager would chime in and say, 'Hey, thanks for the feedback everybody. We're going to go ahead and take a look at this and we'll get back to you.' Which is basically code for 'yeah, probably not going to happen.' And so we thought, hey, we can make a really simple way for people who don't know what APIs are, who don't know how to write code, to build integrations behind these tools. You'd have a trigger for a new event happening, you'd have an action that would happen automatically. And that's kind of where it started, just these very simple integrations. Five years later, we added multiple actions and allowed you to build end-to-end workflows. A lot of the ways you can think of those as a precursor to modern-day agents. Now, these were deterministic workflows that performed a discrete set of tasks in a very predictable way. But it turns out it's pretty darn close to the modern definition of what an agent is. So zaps in some way are kind of like mini deterministic agents.
It turns out Wade actually coined the term agents from this.
Well, we tried to call them zaps, but I think the world wanted them zaps.
How did things change in 2022 when ChatGPT came out and GenAI became part of household conversation? Was that a readjustment of your roadmap or were you saying, 'You know what, this is great. This is just in line with everything we've been doing. If they want to call zaps agents, call them whatever you want as long as they're using Zapier.'
Well, a lot changed, but it took us a minute to figure out what exactly we are dealing with here. I think the ChatGPT moment itself wasn't actually a catalyst for us. We had been dabbling with AI pre-ChatGPT. And I say we, my co-founders and I, mostly them, kind of as ICs. The rest of the company, I think we were like, 'Hey, that's cool. Have fun.' Co-founders doing weird experimental stuff is par for the course. The rest of us have revenue goals to hit and product to ship. ChatGPT comes out and it was like, 'Hey, this is pretty cool.' But it wasn't too different than some of the things that my co-founders had built. And so we were very suggestive about it internally where in Slack we'd post in general, 'Hey, check this out. We think this is cool. Maybe you think this is cool too. It might be useful to think about where you could put this on your roadmap.' That was kind of the vibe. What really was the catalyzing moment for us was GPT-4. GPT-4 launched roughly six months later, and the speed at which we went from 3.5 to 4 was very impressive, and the difference in model quality was pretty large. GPT-4 was a fantastic model, way better than 3.5. We looked at that and said, 'Whoa, if this pace continues, we need to change everything.' And so that was where we got much more intense about shifting the company to taking advantage of AI capabilities. Now, in terms of what specifically has changed, I think there's two pretty important takeaways. One is AI opens up entirely new use cases specifically around working with unstructured data. In the past, if you wanted to build an automation, you had to have it very well structured. You know, think like I've got a lead form that has a name, an email address, a company, a domain, and I'm going to collect it from the lead form and then put it into the CRM as name, email, company, domain. Very structured every step of the way. But if you had a big old email or a big old Gong transcript or just messy like a PDF, if you had stuff like that and your data was kind of 'well, it's in there somewhere,' you basically couldn't do automation around that, at least not in any highly effective way. Now it turns out with LLMs, it's much easier to take advantage of unstructured data. You can say, 'Hey, there are contact details in this invoice. There are product purchases inside this invoice. Extract them and now log them into QuickBooks and make a good invoice out of these things.' So you can work with unstructured data as part of your automations. That's a big one. The second one is creating workflows is now a lot easier. Used to be you'd have to go through a bunch of configuration steps and explain things very precisely in the setup flow. Now you get natural language. The idea of vibe building is a really big concept because it's a lot easier for humans to speak in the way that we've been trained to speak since we were toddlers. The machine can now interpret that, put that back in front of you as a plan and say, 'Hey, what I think I heard you say is you wanted me to go build this, this, and this. I've added a bunch of extra details to it. Do you think that's correct?' And you can say yes or no. It's a lot easier for humans to go build stuff in that form factor versus sitting down and saying, 'I have to come up exactly precisely with what I want to create on day one.' And so I think those are the two big shifts you see when you go interact with Zapier now: you get that vibe building experience and you get the ability to work with unstructured data.
We put a bunch of questions here, but one of the questions you're just talking about with the workflow when ChatGPT came out and it became clear that the start of it was going to be a chat interface and then that has evolved and you all definitely picked that and said, 'Okay, that's what people are familiar with,' where some other competitors have done more of a workflow visualization of it. I know that's something that you all have changed. How this area of the world is evolving so quickly on what customers expect and how they want those workflows to work. How are you all staying on top of that? And are there changes or hazards in time where you're like, 'Okay, this isn't working for what our customers want. We actually have to throw it out and refactor this whole idea and how we're thinking about this.'
Yeah, I think product development is definitely changing right now. You can't be in love with your particular solution to a problem, in part because you might be wrong. That was always true. But I think the real key difference now is you can build prototypes so much faster than you could in the past. There's just so much more throwaway code where you're like, 'Hey, let's try this. What do you think about this? What do you think about that?' And the cost of throwing away code just doesn't feel like that big of a deal. If you talk to your engineers, they probably have something like Cursor or Claude Code running constantly. A good chunk of that code that's being generated doesn't go to production. It's just, 'Let's build a thing and use it one time or maybe never use it.' The cost of that is effectively free now. So you just have a lot more of that going on. The second thing that's shifted in terms of thinking through the roadmap is you just have a lot more ideas in the world, so you have to be kind of nimble and agile and respond to what is working for customers and be ready to react to that. I think the third thing is the model capabilities are emergent. You've got new models coming out every three to six months, and once the model hits, you don't always know exactly what it's capable of on day one. But over the course of weeks and months, you start to figure out, 'Hmm, this model now allows me to build with these new types of experiences.' As a result, this idea of 'here is my three-year vision or my one-year roadmap' — I find the percent of correctness of those documents to be a lot lower these days.
Yeah, I think there's a lot of things that are just going out of date so much faster, and we're okay moving on to the next thing and trying new things. You talked about the experimentation for Zapier navigating this idea of workflow automation into where there were few competitors into something much larger and agent tooling where there's so much noise and so much building. You all were the very tip of the spear and I think in a lot of ways still are, but at the same time you now have this huge landscape of competitors. How have you navigated that transition? I think the follow-on to that, something you mentioned and talked about on your podcast with Amanda in one mind, was just this idea of moving from PLG to enterprise. I know Zapier has become more of an enterprise solution. I think a couple years ago you released your enterprise version. So how have those two things collided and isn't it fun to be a CEO right now?
Yeah, the way I talk about it to the team is our first decade, we were kind of working in a space that was pretty unsexy, kind of overlooked and forgotten. No one was doing integration and workflow. It was just not all that interesting. We thought it was interesting. We thought automation, efficiency, productivity was great. It felt obvious to us that more companies should care about it. But in practice, it didn't actually play out that way, which always kind of surprised me. The 2010s was the era of blitz scaling. It's like, 'Hey, you should scale up really fast, throw a bunch of bodies at this problem, and you'll figure out how to optimize it later.' So in practice, you just didn't find organizations, particularly in tech, cared as much about efficiency and automation as maybe what we felt they should. But on the good news, no one was competing in our space. So we kind of had this blue ocean to go after. Post-ChatGPT, those traits have inverted. We certainly are not in a one-of-a-kind in the market anymore. There are a lot of people attacking this space from all sorts of different angles, including OpenAI themselves. I mean, everybody kind of has a version of what they're thinking about. But the flip side is the market potential has also increased a hundredfold, a thousandfold, maybe more, because everybody sees the possibility of what you can do with automation, with AI, with focusing on efficiency. In the grand scheme of things, I'd rather be in the market that's enormous and have some competition to keep you on your toes than be in the sleepy other area. So I find that pretty exciting. As far as the PLG enterprise thing, I actually think we've benefited quite a bit from the AI era because it's given a reason for enterprises who may have written us off to come back and look at us again and say, 'Hey, maybe Zapier actually has figured some stuff out here.' I think that's a good thing because we are quite a bit different from a product perspective than we were even just a couple years ago. There's a lot of companies that if the AI era had not come around, they might have just been like, 'I looked at it, it wasn't for me then.' If you're one of those people, you should check us out again because it is a very different product experience now.
I also think that for me, to the point of the intro, it makes perfect sense strategically that Zapier would be building autonomous agents at this point because again, to your point, these structured deterministic workflows are the predecessor to agents. Now you have 10 years of a trusted brand to credentialize your right to win that market, which I think is a big thing in a world with thousands and thousands of competitors. What is your right? Why do you deserve to be there? Have you felt that uplift? Have you become one of these stories where all of a sudden growth dramatically accelerated as a consequence of the AI wind in your sails?
Yeah, I think especially when you look at the enterprise side of the business, it's exactly what you said, Sam. There's 10 plus years of trust, reliability. These integrations are rock solid now. We're operating at a scale that makes it work a lot better for these organizations. I think also we can talk about the challenges of navigating this era in a way that's more relatable. When AI hit, we were a 700-person company. We had to reinvent our own operations. We had to figure out these things from first principles in a way that maybe a brand new startup doesn't have the same problems at this stage, but we have them. So we can come in and say, 'Hey, here's how we thought about solving employee onboarding or offboarding. Here's how we thought about reinventing our revops stack or changing how we work with sales data.' A lot of our customers look to our own operations, not just the product, but the product plus our expertise in these areas. That helps quite a bit.
So diving into the specific topic of agents, I'm reminded of Jason Lemkin who says they used to have 20 or 30 people working at SaaStr. Now they have three people but eight full-time agents, I think, and he's an investor in half of them. So we don't know what the reality is of how agents are actually working in practice and in production, where they're not working, what it truly means, is there a common definition. So when you see use cases that are really successful and then use cases where there's still a lot of hype, where do you think they succeed and where do they still break down?
What is working today is what I would call agentic workflows. An agentic workflow is pretty similar to a deterministic workflow where you still have deterministic nodes, kind of like how Zapier works historically, but then in the middle of it you're inserting either an LLM step or an agentic step that handles some unstructured data, maybe makes a decision for you, helps summarize some information, categorize some details, and then it extends back into a deterministic workflow. The reason this tends to work well is that LLMs do really well with lots of context. With deterministic workflows, you can feed it precise context and then target what you want the agent or the LLM to do to a very specific task. When you're able to do that, you get higher reliability for the entire flow. What isn't working as well is 'Hey agent, go reinvent how I do onboarding.' These tasks are too big, too abstract. The more you chain along various tasks that you want out of the agent, the more it drifts from what a human might consider the correct application. Now, agents are getting better and they're getting better fast, but in the enterprise setting at least, what you really need is to break these things down into their smallest actionable components and then tie those into a mix of deterministic workflows and agentic workflows. That's really what gives you the advantages today.
Are there specific horizontal functional areas where you've seen great success, or is it really that an agent can be successful in HR, in finance, in sales as long as you've done the work of breaking it up into its component parts?
Yeah, we've seen success in any function. I don't think there is 'oh, this function it doesn't apply over here or only this function can take advantage of it.' Obviously there are areas like coding agents that are getting a lot of discussion today because they are quite good. Customer support is another area that gets a lot of disproportionate talk because the success is coming faster. But I don't think that our evidence is limited. We have evidence in basically every function and vertical where you can apply AI and be successful.
Wait, it's clear you're working to evangelize this as much as you can. I've heard a lot about Zapier and making the team AI literate across the organization. I saw the interview that your assistant had with Claire Vo on the Lenny's podcast recently. Have you seen the Wade GPT by the way? Do you have access to the Wade GPT to be able to see how your team talks to you?
I'm kind of curious if you incentivize or how do you ensure your team stays on top of AI trends? Are there different ways and tactics that you get your employees to help evangelize in the same manner?
Yeah, early on it was a little more like pushing a boulder up a hill. We had to get people to pay attention to this new way of working, this new technology. Also, some of those early models had more rough spots, right? Pre-reasoning, you had more hallucinations. It was easier for a smart person to look at it and go, 'This isn't good enough.' As much as you could say, 'Just wait till it gets better,' you could see how a smart person might be disillusioned with the dialogue around AI. But we kept doing regular hackathons. We kept doing regular show-and-tell internally. We implemented a few constraints. We said, 'Hey, we're not going to greenlight new hires or new software spend unless either me or my CFO have a chance to review why you think you should go get them.' We still will have new hires and new software spend. It's just we're putting a little bit of friction in there to make sure there's a forcing function to go and try and use Zapier, go and try and use AI first and foremost. That recipe marinated for about a year, and out the other side we got a team that is much more eager to push on AI now. So now instead of us pushing that boulder up a hill, it's more trying to chase the good ideas inside the company and trying to keep up with the good ideas and say, 'Hey, how do we actually do more of that?' That's the big shift that has happened for us. It's less of 'I have to go convince you to take AI seriously.' Now everyone sees it because once you use it and you get some success out of it, you don't really want to go back. It's not like you have to convince somebody, 'Oh, this thing that saved me a bunch of time and hassle, I really don't want to use that thing. Give me the old boring hard way.'
Well, I find that's true for anyone of course, but it's just ensuring that there's uniformity between the teams and collaboration. Sometimes I find if someone finds something that they feel like they want to keep a secret and it's making them 10 times more productive, you're like, 'Wow, you're killing it.' It's harder to get them to open up. So that's why I asked about the incentivizing question of just how do we get our teams to continue to not just be curious about it, but then when they find something that's working, communicate to the org about what they're doing and why they're doing it. It definitely makes sense to bubble up to you of like, 'Okay, I need to check this out. I want to be hands-on keyboard with this,' but then making sure cross collaboration in the teams is still happening too.
