Wade Foster & Carilu Dietrich • How to Transform Your Org with AI

Welcome Wade Foster, co-founder and CEO of Zapier, and Carolyn Dietrich, CMO, hypergrowth advisor with Took a Blessing Public. Welcome to the show. How are you doing?

Good, thank you. Hi, Goddard and Carrie. Love your work and companies.

Hi, Carol. Hi, Wade. Hey, folks. Excited for your session. Yeah, good to see you.

Awesome. Carol, take it away.

Thanks, guys.

Great. Well, thanks for having us. And Julia, thanks to you and the team for hosting another great conference. So much is moving so quickly, so it's exciting for all of us to keep learning from the best and brightest. And that's one of the things that makes me excited about interviewing Wade today. As many CMOs are, I've been digging deeper and deeper into both AI vendors and specific examples examples of what real-world marketing teams are doing today to use AI, to use agents, to use automations. And Zapier keeps coming up. In fact, just today, maybe an hour ago, I published a blog. I write a blog at www.carilou.com, that means three W's, carilou.com. And I wrote about a specific research process that one really innovative marketer at SupportLogic has rolled out that started as a BDR research project and really informed across the go-to-market funnel, how they, um, understand different prospects, personalize messages, and, and make their team much more effective. So on Zapier amongst, um, several that I've been looking at recently. So Wade, you guys just in the last week have actually rolled out some really big leaps forward in your Agentic capabilities. Do you want to give us, give us the highlights of the news?

Yeah, so last week we launched our AI orchestration platform. It integrates AI agents with over 8,000 apps. Zapier is the most connected platform that does this type of thing. And it allows users to add AI agents directly into those workflows. It enables those systems to intelligently adapt and make decisions rather than just following predefined instructions, which that was sort of the world pre-AI where you'd have this like, if this, then that, that, that, that sort of style logic. You know, alongside that, we've added human-in-the-loop controls via Slack. I think this is a really important mechanism if you're introducing agents into your workforce, especially if you want to do it in a trusted way. You know, oftentimes these— well, these agents are probabilistic. They're not deterministic. And so that means they may not do the thing you totally expect it to do. And so, you know, as we— start to deploy these things in more and more riskier environments, you're going to want to have a mechanism to verify that these things are performing the way that you expect them to. Those are the two big announcements that we've had in the last week.

Well, and there's a third too, right?

Oh yes, our Zapier MCP server. We launched our MCP server earlier this year and the really exciting thing is it's available in Cloud. If you are a user of Claude, which is a product from Anthropic, you can install the MCP server right there and interact with any of these 8,000 different apps directly. It can be simple things like, hey, can you tell me what's on my calendar for today? Okay, hey, I haven't caught up with this person in a while, can you tell me more about them? All the way to like, hey, pull this report from Salesforce or pull this data from HubSpot or do this action in this way. And so it really gives you a lot more, it gives Claude the ability to actually take action for you in the real world versus sort of just tell you the instructions for how to go do that thing.

And break down what an MCP server is for people who aren't as close to the cutting edge as you are or may not have already be deploying.

Sure, so MCP stands for Model Context Protocol. It's a standard way for agents to discover and act on external tools. You can kind of think of it a little bit like, you know, what APIs do for like SaaS servers or SaaS applications, like MCP does for agents. They can talk to each other in a particular way. So that's like the most simplistic way to probably describe it.

And so one of the things that's on my mind is just that there's this proliferation of agents. Salesforce announced their agent strategy last year. It was the basis for their whole Dreamforce. You know, you've integrated now and so you're providing the integration layer, it sounds like, but behind Claude's agent capabilities, OpenAI, of course, is going to roll out agents. Potentially all of our different platforms will roll out agents. You know, one of the advantages of Zapier is that you're really horizontal. You've traditionally been this integration layer that cuts across all different companies. But how do you think that leaders are going to pick and choose their agent? Processes if there's kind of this proliferation of agent capabilities throughout their tech stack?

Yeah, I do think that, you know, one of the unique things that I think we are Is that my internet, Julia, or is that Wade's internet?

Okay.

Yeah.

Okay. We were— Julia and I were worried about my internet because I'm traveling, but it seems like we've lost Wade for a moment. So we will wait for him and talk a little bit more. I guess I'll just talk for a second that one of the things that I'm seeing is one of the reasons I'm running into people who are using Zapier is because they are trying to stitch together a lot of different platforms. And so some of the agents within a specific platform are like more optimized for that, that platform. So again, walking through a type of process that I've seen in two instances recently, that's a bit more of the AI. I don't know where I guess some of the AI ends and some of the agents begin. Hey, we've got Wade back.

Hey, sorry about that, everyone.

Yes, no problem. So I was just thinking about some processes that I've seen recently where people are starting in Salesforce and they have a button that like creates an analysis of a sales call or a sales account planning or starts research for sales in their Salesforce instance, but then goes out across a lot of different systems. And it might go out across Perplexity, it might go out across OpenAI, it might do searches on SEC databases. So it's really moving horizontally and moving through a lot of different types of services to bring things back. And then what I'm seeing with a lot of the innovators is that the more stepwise they create their process, if they break it down into smaller steps, they're finding that there's a) less hallucination and b) they're able to optimize different steps with different LLMs. So that they find that there's better results from some of them. So I guess I was saying that my own personal opinion, I think that like some of the horizontal tools might win the agent war because really no one is a single platform company as much as Oracle or Salesforce, you know, wish they would be.

But yeah, I think totally. Like, I think we all have this dream of like, oh, I'll just, delegate this like super high-level task to an agent, it will go figure out everything for us. And I think, you know, maybe we'll get there. But for where we're at today, folks are getting the most benefit by doing exactly what you described, like break the tasks down into very discrete bits and then assign an agent to each one of those and then orchestrate those steps as you go. And then to your point, you can, you know, test and tune each bit along the way. And I think this is where like, you know, we're having— we're seeing a lot of success is helping our customers do that orchestration across all these different tools that they have. Because the reality is, you know, they've got some data over here that has really important access to information. You need to do a web search over here. You know, you maybe do 2 or 3 different web searches, but each of those need to be their own step because the agent starts to hallucinate if you try to combine all those things. And so that ability to break it down into chunks is like a really important skill set, I think, for folks who are wanting to go deploy agents in the future.

What kind of people are you seeing be the innovators in AI and agents within different teams? I know you've always been looking for the builders, some folks with like slightly, slightly technical, I guess, technical confidence. They don't even have to be very technical now in like kind of the drag and drop world. But, but what kind of people within a team should leaders be trying to tap to say, like, can you really push on this for us?

Well, I think there's— it's not— I don't see like a particular role that necessarily is like standing out more than the rest. I mean, if I were to pick one, it's anyone who probably has an ops title. Like those folks probably on the margin tend to be oriented this way. But the folks that really seem to have this mindset seem to be these tinkerers. They're like intrinsically curious. They like to break this stuff down, deconstruct it, put it all together. And, you know, they're finding a lot of success just by like, you know, continuing to tinker with this stuff. And the cool thing about AI is that you get like this knowledge on tap. And so these tinkerers now are augmented by being able to go to ChatGPT or Claude or Gemini or any of these tools and actually build like domain expertise. So they may not necessarily be the expert marketer or the expert sales rep or the expert product person. But they've got the knowledge on tap such that they can then apply their like technical tinkery mindset to actually orchestrate these things. And as a result, they're often performing better than somebody who maybe was like a career, more a career salesperson like that, able to combine these two worlds together in a really interesting way.

I'm seeing a lot of folks in the marketing sphere, the go-to-market sphere in the, in the revops team that are specifically strong or the growth marketing teams where they're a little bit closer to, to kind of some of the process structures that enable the broader sales team or enable the broader marketing team. And some of those are showing innovation. One of the challenges is that agents, especially as you're building them to be simpler for everyday people to create, could be created by anyone in a company. And a lot of CMOs and CEOs are talking to me recently about how difficult it is to get adoption across their entire employee base of AI. You know, we recently saw the Shopify CEO's memo leaked about, you know, we're going to put a pause on hiring unless you can prove that your new hire can't be partially replaced by AI. Duolingo, I know, sent out a big— CEO sent out a big note as well. You had kind of a different approach instead of a stick. It sounds like you guys did a carrot. You did like a global hackathon. You kind of had like a Red Week or something where everyone got deeply involved. Did I read that you guys have 89% adoption of AI across your company?

Yeah. So, there's a little over 700 employees and it's actually over 90% folks use AI daily inside their work. But this is not something that like, snapped overnight. I don't know that our approach was crazy different from what Shopify or Duolingo did. We just started sooner. We did this 2 years ago. For us, the impetus was my co-founders and I had been playing around with GPT-3, GPT-3.5 before ChatGPT came out. We were exposing the team and just saying, hey, check this out. If you like interesting technology, You might think this stuff is fun and there might be interesting ways to deploy it, but it was very suggestive in nature. ChatGPT launched, get even more excited and maybe we're making stronger suggestions. Hey, this is really important. But again, we're not really turning the screws on the company to say, hey, we really got to go do this. The wake-up call for us was in the spring after ChatGPT launched when ChatGPT came out. The thing that's so impressive is that GPT-4 was a great model, but even more so was the rate— like, it came out 6 months later. It was so much better than 3.5. It was cheaper, faster, etc. And the things were improving, but like something that we just had to— we had to be paying attention to. It was like a massive wake-up call for us. And so that's when we said, okay, AI code red inside of Zapier. We're going to create We're going to do a hackathon for an entire— we're going to— whether you're in engineering or if you're in sales or finance or HR, we don't really care. We want you to go just build stuff. And because we want you to build a fluency, we want you to understand these things. Pre the hackathon, we did have to do some prep work. So, you know, we had to go procure all these. We had to make sure these things were abiding by our terms of service, our data privacy agreements, all our compliance regulations sort of things. Then a little bit of education because there wasn't a crazy amount of external education at this point in time. My co-founder Brian was recording Loom videos showing people, here's how you hit the OpenAI API and here's some weird things that you might think it behaves this way, but actually it's new and different, so it actually behaves this way. Expect some of this stuff. That hackathon definitely created organic adoption, spiked pretty quickly. After that, then we really reoriented a handful of key things inside the company. First, we shifted from being very suggestive to a little bit more directive in a couple of key places. Certainly our product roadmaps, we were like, okay, there's a few areas that have to be much more AI-forward. There are certain features and capabilities that have to be baked in. That was straight from the top. Then there was we made it a habit of doing show and tell in our all hands about once a month where we would invite a few key people in the company who'd built something interesting with AI to just show off what you built. Then we did those hackathons about every 4 to 6 months. We just kept that repetition going. As a result, more and more people started to see what was possible and snowball effect happened. In the early days, if you were using AI, you were an oddball. It was like, okay, you're weird, nerdy early adopter. Like, you go have fun, like, have a good time. But once that snowball kicks and all of a sudden the, like, the energy becomes so infectious that folks are like, okay, we got to do it, and now you're kind of the oddball if you aren't. It's sort of like, wait, what? Why aren't you doing this stuff? Like, this stuff is so obviously impressive and good. Um, yeah, you may still be running into bumps, like, this technology is not perfect, uh, at the end of the day, but most people see the potential and the promise such that you just get that flywheel starting. But like I said, it took 2 years to get to this point. This was not a thing where I wrote a memo and it was like, great, you know, everybody's doing this stuff. There was a lot of other touchpoints that had to happen to encourage adoption.

You know, it's funny because Atlassian really loves hackathons and it was such a core success driver for us, especially in the early days, in part because we would do a lot of hackathons that were cross-departmental. And you'd get more technical people, you know, you get a marketing person with an idea and a support person who was interested in that same kind of category and maybe an engineer and an IT person, and they would all work together and it would both build relationships, it would build the actual hackathon thing, but then it also built these relationships that led people to have collaborators in their ongoing ideas. So I can see how this like ongoing hackathon would give your employees a chance to like be like, I kind of have this idea, but I can't take it all the way by myself. Are there people I can talk to either at this time or, you know, on an ongoing basis because we've started to form these relationships. So I guess I couldn't endorse hackathons more for all sorts of things. Are you planning an agent hackathon? Like now the whole company—

Yeah, we will have an agent hackathon as well too. Yeah, I want to add one other thing with the hackathons too, which is The challenge I think your average employee faces in AI adoption is that you've got metrics to hit, goals to drive, and they have a way of doing things and they're good at that way of doing things. And so there's a little bit of like tunnel vision going on. And so it is beneficial to you to kind of like create some space for them to be like, no, I need you to just go put your hands on the keyboard and go mess around with this stuff and learn. And you get what you just found, which is that once they do, they get excited, they get energized. It's not that they aren't into this. They actually are excited. It's just they need a little space. They need a little bit of that to go make it happen. And so that, that I think is really important. But yeah, agent hackathons, 100%. We're going to go do a lot of these.

Well, and I would say over the last 6 months, as I've gotten deeper and deeper and deeper into AI, it seems to me that there's really this slow down to speed up aspect. You know, when we first— when ChatGPT came out, you were like, give me a marketing plan or, you know, what should the strategy be? And now we can see these people that are including prompts with, you know, dozens of documents, hundreds of lines. You know, I was following one of the Anthropic technical folks who said when she's optimizing the LLM, she goes back and forth 100 times. And like, I used to get frustrated in the beginning when you went back and forth 4 times, you know, and now as I'm seeing people build these like more complex AI automations, a lot of it is documenting things that haven't been documented in the company, like exactly what is the best case process and like what kinds of exact answers do we want and what format. Do you see that same thing happening with the agents? It's really kind of like a slowdown to speed up aspect because what you put in is so critical then to letting it do its work speedily on the back end.

Yeah, my co-founder Brian had this comment he actually made yesterday where, you know, we were debating this around like, you know, particular AI effort we're working on where you know, folks are like, okay, we're going to have to like— this is going to come in at a cost of this other thing. And he was like, folks are going to have to be comfortable with this idea that you actually do kind of have to put 2x the work into this stuff, but you get like 10x the output, you get 100x the output on the other side. And so that's like a weird— that's like a weird way to think about it is like, hey, there is— there's a level of polish and rigor and attention to detail here that is really important. If you put it in, you're going to get outsized returns. On the other side. And so I definitely think that is very much true. And, you know, the folks that are getting stuck are kind of getting stuck with what you said where you're like, oh, I kind of put a basic prompt in and it kind of gave me a basic answer. I don't really see the value, so I'm going to move on. And it's like, well, you kind of got to— you got to develop some fluency with this. You got to have a deftness with these tools. And that just comes from, you know, trial and error. It comes from sitting beside somebody who's done some of this stuff and going, oh, I didn't realize you had to do that, that, that, that. You are on a whole other level than I am.

Mm-hmm. Yeah, it just takes a lot more upfront. So tell us then, as you've rolled out, how are training the agents, how is training your agents going to be different than setting up your Zaps and your process? You know, I think agents, of course, aren't deterministic. But talk us through how people should try to think about approaching their agent projects in a more complex way than just their automated processes.

Yeah, I think the, the main advice I have here is human-in-the-loop steps. So you're going to want to break these things down. We talked about that already. But as you're doing testing, you're going to want to have a verification step, especially early on. And so there's a few ways you can do this. You know, we have our Slack approval steps that work. You can also like have it create a Gmail draft for you instead of send the email. You can have it like log a thing inside of a Zapier table or Google Spreadsheet, and then you can go review it in there and then like flip a field to be like, okay, I've approved this, go ahead. And so I think that's going to be really important to help build that trust. Now, over time, as you start to see this thing perform more consistently in the way you want, then I think you're going to be able to start to remove those verification steps at certain points and say like, okay, I trust this. This is not making, you know, the level of mistakes. Longer term, I think what we're going to eventually see is what I would call statistical quality control. In the past, this idea of quality control has been a side thing in most software companies. But I think it's actually going to come really critical in the same way that Toyota ran the field on American manufacturers in the '80s by getting really good at quality control. I think you're going to see the same thing happen with agents where you're going to implement statistical statistical process control standards to help you really verify, hey, are these things working consistently or not? When it doesn't hit the— when your sample is not verified, you're going to shut the whole thing down, you're going to revamp it, and then you're going to go. But I think we're a bit of a ways from that. That's going to be really important, I think, for enterprise that are trying to deploy these agents doing millions and millions of tasks every single day.

Are you still seeing a lot of hallucinations? Because to your point, I think there's the There's multiple problems that you would encounter in quality control, but certainly hallucinations is one of them that we know about.

Yeah, you know, the models are getting a lot better at this, especially the reasoning models. You know, this would be like, you know, OpenAI has like the O1 series or the O3 series. These reasoning models just behave differently than the pre-trained ones. And so there's less hallucination with them. The second thing that helps with the hallucination is this thing we've been talking about, break it down into smaller tasks. That said, you're still going to have hallucinations. You're still going to have it make up stuff and you're going to have to find ways to like verify some of these things at the end of the day. And so that's where you're going to want to have some sort of baked-in quality control mechanism and you're going to probably have to define like what your fault tolerance is for some of those things, just as you would like, you know, if you delegate something to a employee, you're not expecting that employee to be 100% accurate. But you're expecting to be like pretty accurate. But, you know, you're not going to fire them the first time they make a mistake. You're going to go, oh, actually, like, hey, you made a mistake on this. Let me show you how to do a better job with it over time. And your agents are going to have some of those same characteristics.

You know, one of the reasons I think a lot of companies rely on you for integrations is because if they build their own integration and those vendors change their software, then that underlying process breaks. And so the whole build versus buy, you're kind of between build versus buy. It's like someone could build it all themselves from scratch inside or have like a fully complete, much more expensive system. And you kind of sit between those two saying like, hey, we'll take away this burden of keeping up these connections because you have what, over 8,000 and 30,000 pre-built applications or something already. But you also give people so much autonomy to build their own thing because it's kind of like building blocks that they can use. As you look ahead, you know, you're trying to keep up with the different LLMs, you're keeping up with different integrations, integrations, that seems like where it's going to break down for a lot of people who try to build their own. How do you guys stay across so many different vectors to keep all of these things working? I mean, obviously you've been doing it for years, but LLM adds another layer, I think.

Yeah, you know, this is what we, we bang our heads against the wall on this so you all don't have to, like, is the end result. But I think there's ingredients that are really impactful. One, You know, we have just tons of monitoring set up for all of these situations. So we're just spotting this stuff early to be able to fix a lot of this stuff. With LLMs, that monitoring has gotten better. Those fixes have gotten faster. You know, it allows us to just move at a pace that we couldn't before. The second thing we'd benefit a lot from is just our scale. You know, we have a relationship with these 8,000 vendors we work with. And so when they go make updates, when they go make changes, they're thinking about how's that going to impact the relationship we have with Zapier and all those people that are consuming and using this stuff. And so because we have that trust, people know that we're being used. There's a level of care that these folks give to, to us. And, you know, that wasn't— that didn't always— that wasn't always the case, but that just has happened as we've grown and gotten more popular over time. So and then the third thing is like we just have like really good engineering teams that are really good at doing this stuff. There's not like a secret sauce to it. It's just something we spent, you know, almost 14 years getting good at.

Yeah, I mean, it feels like in this land of AI and agentic, your build is never quite done because you're going to keep optimizing as you have new information to feed your agent, new systems you want to connect to, new outcomes you're trying to drive. We have only 3 minutes left. So if people are just getting started with agents, what is your advice to them to help them be successful and understand what's going on in the agentic world?

Yeah, I think one of the easiest ways to start is to set up a Zap and instead of setting up like a full-blown, you know, probabilistic agent, just add like an OpenAI step in the middle of your Zap. Oh, here's a really simple use case for a marketer. You got a form on your website. Okay, great. Send that over to— send the email address and the company domain over to OpenAI and say, hey, tell me what you know about this stuff and then add it to Salesforce or add it to HubSpot. That's not like a full-blown agent, but it starts to give you a taste of like what is possible with these things. And it's super simple to get going. And then over time you can start to figure out like, oh, how do I combine more and more of these things to actually get like a full-blown agentic system that has a lot more autonomy to go, you know, do a task at the end of the day. So that's probably my recommendation for a simple starting spot.

That's great. Thank you so much. I really enjoyed this. I don't see the other people on our, on our line yet. Oh, there's Erin. She must be joining. And I just wanted to again mention that I've, I've written up a case study just this morning on a great Zapier example. And last week did a prioritization list of a number of different places where I'm seeing the most processes being deployed successfully in marketing. So I'll put a couple of those links there into the chat. Thank you so much for having us, Julia. Or do you want us to do one more question while we wait for whomever is next?

Thank you so much, Carolu and Wade Foster. Such a treat. I shared your link with the community. And yeah, I guess we had a lot of concerns about transitioning to Agentic AI in our CXO. Lend. So what's your advice to all VPs and CXOs who are considering the next step, but they are not quite sure where to start?

I think, I think the best thing is to start. Yeah. Like, you know, it's so easy to be fearful of it, but to me, that is a surefire way to not reap any of the rewards. Whereas like the upside is so incredible here that you have to find a way to get to start dabbling. So you can, you know, mess around with Zapier by adding like a little bit of AI into the mix of your workflows. You know, sign up for ChatGPT or Claude and start playing around with it for these different areas. And as you create space for others and your employees, run hackathons to do some of these things. You don't have to deploy everything to production. Find the best ideas and go, oh, okay, that's where we're going to start. We're going to start doing it for this, this, and that. And as you experiment, you're going to start to, you know, just walk down the transformation path to reinventing your organization. I find that this sort of like fear-based approach is a surefire way that you will probably get—

And I would just add that you don't have to do it alone. So I've done some experimenting by myself, but I also hired this guy who's a, like, great process optimizer to, like, work on a specific AI process for me that I couldn't do myself. And then we worked on it together and optimized it, and then I took it over and started using it. And I'd also say one of the more innovative examples I've seen is of a sales leader who's, who's done an automation inside of Salesforce where it automates— I think I talked about at the beginning— an evaluation of the sales rep sales call and their account planning and everything that he would have done for each sales rep every couple days, but they can do it on demand immediately. And he just partnered with someone else that was technical on his team and they hacked it out together in like a long weekend and a couple days. And so I think that having a partner where you're talking through how you're thinking about it and can plug at it together also makes things move a lot more quickly. And in the end, it's not the AI that's the intelligence, right? You need the people who really understand the business process and know what good looks like. To both create what's good, you know, create the automation and then, and then test and evaluate the output because every, all of these systems, these, the AI processes and agents need to be continued to be optimized and they need to be optimized by people that know and understand. And sometimes that's done really well in a team instead of isolated by yourself.

Super helpful. Thank you so much again, Wade Foster and Kara Lou.