AI Agents, Clearly Explained in 40 Minutes | Wade Foster (Zapier)

Most people, when they say, "I want an agent," you realize they actually just want a workflow, like a, just a deterministic automation. The reliability of these full systems is just not great for more complex tasks. And so the things that are most impressive today are kind of in like this AI workflow or agentic workflow world. Instead of me sitting down and like answering all these emails, what I'm gonna be doing is talking to the agent and saying like, "What new instruction can I provide?" Or, "What new guidance can I give it so that it does a better job?" As cringe as the like CEO memo is around like, "We must use AI," I do think it's important. I think where CEOs often fail is that they just kind of stop there and then they just expect the organization will sort of like magically figure out how to go use these tools. Best thing you can do is just—

Okay, welcome everyone. My guest today is Wade Foster, co-founder and CEO of Zapier. Uh, excited to get him to demo how you can, uh, automate work and save time using Zapier's AI workflows and agents. Uh, talk about what the difference is even between workflows and agents. And also talk about how Wade is making Zapier AI-first and how he evaluates AI fluency with both employees and also people looking for a job. So welcome, Wade.

Yeah, Peter, thanks for having me.

All right. So, you know, everyone keeps talking about AI agents and— but there's actually like a whole spectrum between workflows and agents. Do you mind like kind of walking us through this difference?

Yeah, happy to. So this is what we call like the AI automation spectrum, and it sort of goes over you know, this progression of workflows that are deterministic all the way to workflows that are like pure inference. On the left-hand side, this is, you know, pre-AI. Like, this is what— when we thought of workflow automation, it's kind of how it worked, where you'd have like a new lead comes into your website. I want to send off an SMS via like Twilio, and then I want to add it to my CRM like Salesforce or HubSpot or something like that. It's like taking deterministic data and then shuffling it from one system to the other to like generate an end-to-end workflow. Now on the right is like all the way over here. This is like the entirely new way of doing this where, you know, you've got an agent sort of that sits in the middle of this that has access to tools and knowledge and instructions and it can reason for itself. It can decide things. And, you know, you spit information into it, you kick it off somehow. Mostly what people talk about as agents today, I call them chat agents. Like they're inside of a chatbot. It's a message to the chatbot that kicks off the agent. I think that that's too narrow of a definition. I think it's better if you start to think of like all the different ways you could wake up an agent. You know, you think of all the triggers on Zapier, like a new lead, a new email, a new blah, blah, blah, blah, or like all unique ways you could wake up this agent. The agent goes in and reasons through all this stuff based on the instructions you've given it. And then in the other side, the agent takes action and does something. You know, most of the time today is the chat agent the only action it can take is post a message back to you. But again, I think more powerful agents are ones that are able to go do stuff based on the tools that you give it access to.

Got it.

But then you've got stuff in the middle here. This is where I think you actually see the stuff that is working in production looks a lot more like the stuff in the middle today. You've got AI workflows, which are basically a workflow, but you can just add an AI step in the middle of it. Or what I would call an agentic workflow where you're using sort of multiple agents to make stuff happen. And now the reason this stuff in the middle is what's actually working is because you get more determinism and with determinism you get reliability and you get cost advantages. Whereas this stuff on the right is really exciting. I think we all believe that this is the future of where it's going. The reliability of these full systems is just not great for more complex tasks. And so the things that are most impressive today are kind of in like this AI workflow or agentic workflow world.

Yeah, like it really grinds my gears. I guess the word agent is just like way more sexy than workflow. Workflow sounds like work and agent is like, you know, you like aspire or something. But like, you know, everyone just keeps calling everything agent, like including OpenAI's latest thing, right? Like the— but it looks very much like a workflow that you're putting together with like agent steps.

Yeah. I would call their agent thing like an agent workflow. Like it's this sort of, but again, I don't know that it's re— it's not automation, at least not really in the way I think of it. It is more of a like chat agent builder at the end of the day.

So on the agent side, like, so, you know, it looks really cool, but like, you know, the agent can just like keep spinning and like burn a lot of tokens and like get to an output that you don't want. Like you mentioned, it's not very deterministic. What kind of use cases actually work right now for like a full agent? Is it like coding? And that's pretty much it? Or like, is there any use cases that actually work?

Well, certainly like coding agents are definitely the most popular ones that we see. But there are others that are starting to work. I think the real challenge that folks run into is the best agents are small. And there you really focus them on a particular task that you want to go achieve. And, you know, it takes a certain amount of creativity and work to like really get that agent dialed in and be effective. Would it help if I show off maybe an agent I built with Zapier Agents?

Yeah, yeah, please. Yeah.

Yeah. Okay. Here's a good example of something. So this one is an email categorization agent that I work on with my EA. So as you can imagine, I get a fair amount of email every single day, but most of that email is either not actionable or not high importance to me. And this helps us sift through that really quickly. And so we've basically built like a series of steps and it's all just text. You can see we're just giving natural language instructions to this agent around what we want it to do. And so we've kind of like created a bunch of categories. So there's these ones in red. These are action that is required by Wade Foster. And so you can see what some of these are. This is, you know, exec-level communications, board, strategic partnerships, escalated customer support issues, hiring decisions, things like that. Then you got another category, which is things that my EA should be doing. So this is meeting requests, scheduling requests, travel requests, Uber, like HR notifications, all that sort of stuff. And so we've done this for like a whole bunch of different categorizations. The next thing we do is for things that are from customers, we want to get more information about what's going on there. So we have it go reach out to our HubSpot account. We find out details about the company, um, you know, we do web searches to like better understand like what's going on in this account. Um, the third thing is we've got a bunch of like informational categories. So these are emails that, you know, like most of the emails actually are this stuff. It's like marketing, promotional, spam, uh, random, just like internal updates. Uh, and then lastly, we go and apply labels to all this stuff to like better categorize things. So So you have like the action labels, we got customer labels, we got information labels, all this sort of stuff. And at the very end, we start to triage these things. So we'll archive stuff we don't need to archive. You know, Courtney will go snag all the stuff that EA needs to snag. And once you're done with this, once you actually set up an agent like this, you find that your email inbox actually starts looking way, way better. In fact, you know, at the end of the day, when you look across like, I don't know, 100+ emails I might get, it turns out less than 10 actually really need my attention and matter. And so this agent will just use its— like, take these instructions, and then based on these instructions, it will iterate over every single email I get and come up with its answer to this question and help me triage through my email inbox.

That's amazing. You probably crafted this prompt over time, right? As you saw more annoying things pop up in your inbox. Is that how it worked?

Spot on, right? Like, if you were to start working working on this, you would probably just give it, you know, you'd just be like, hey, help me triage my inbox. And like, here's some things that I noticed. But then, you know, if you put a little effort into it over the course of like a couple days or a couple weeks, you'll start to be able to build it up where the agent is able to tackle bigger and bigger and bigger, uh, uh, or just like more emails for you at the end of the day. And that's where I think agent building is probably a little different than building workflows. Is that with a workflow, you kind of know what it's going to do. It's very deterministic. And so you just say, I want to do this, I want to do this, I want to do this. And then you're like, okay, great, I'm done. And you might like, if you change your process, you might come in and like add additions to it and things like that, which is pretty common. But with an agent, the first version of it, you'd go, okay, it's, it's helpful, but it's not as good as it could be. And so you find yourself doing this like repeated tuning of the agent over time. And in fact, I think this is where like a lot of knowledge work is going to go is instead of me sitting down and like answering all these emails, what I'm going to be doing is talking to the agent and saying like, what, what new instruction can I provide or what new guidance can I give it so that it does a better job at answering the emails or triaging the emails or whatever it is that I have to do. And you'll notice this one is just really doing categorization. There's another email that does like responses and there's another agent that does this. And so you have these agents that are pretty narrow because you want them focused in on you know, a concrete job that it can actually go do. The bigger the task you provide it, the more I find it has, like, it starts to get confused, or the reliability goes down. But the more you can constrain the task, the more likely it is you're going to get something you're pretty satisfied with.

That example you just shared is a full agent, right? Like, what, why did you decide to build it as an agent versus like a workflow or like, you know, a GenTech workflow?

Well, so that one I don't think you could do as a workflow because I would have to come up with, um, a deterministic rule for all of those categories. And I would have to be going— I'd have to know, okay, every person that reaches out to hiring, they might include this term. It's like, must contain this keyword or must contain that keyword. And so, you know, a deterministic workflow doesn't have the brains that you can give an LLM, where an LLM you can just describe like, emails like this, I would like you to do this action with it. And so it can extrapolate based on the instructions I've given it. And so that's really powerful when I can use that type of capability for this particular task.

Got it. Yeah, I feel like this is like, I mean, this is basically, you're basically building an AI product, right? You're trying to update the prompt and then you're, you're not writing maybe like, you know, super metric-focused evals, but you're doing evals by just looking at your inbox and see if it's doing a good job or not.

Well, and that's how I think a lot of like personal agents are going to start. But like for people who are building agentic products, they are gonna have like a bunch of evals and things like that because that's where you get, you know, uh, that's where you get like the, if to really, really feel great. But for me in my own personal inbox, like I don't need that. I just need it to like mostly be good enough, you know? Got it.

And, and in that prompt you probably have links to like Gmail or like, um, uh, HubSpot or wherever you're pulling the information from. Like, like is that pretty easy to set up in Zapier or, or is that you just connect a bunch of accounts?

Yeah, basically you just say, hey, I want to give it access to this tool. Zapier has access to 8,000 different apps and within it there's like specific tools you can do. And so you effectively just, you know, click a button, authenticate it in, and then you have access. The agent now has access to it. So it's pretty straightforward to, to do tool, tool access inside Zapier.

Nice. Okay. This episode is brought to you by Vellum. Vellum is the fastest way to build AI agents for production just by chatting using plain English. Simply describe what you want and watch complex AI workflows assemble in real time. Vellum also provides built-in debugging so you can see exactly what your agent is doing, plus enterprise-grade monitoring and compliance. PMs, operation teams, and engineers are using Vellum to go from idea to shipped agent in minutes. Check out the link in the description and use the code AIAgents25 to get 25% off your first 3 months. Okay, now back to the episode. Do you have any advice? Like, let's say, let's say I want to build the same email agent that you have. But I don't want to like have it completely screw up my inbox. How do you— is there a test run or something? Or like—

well, yeah. So you, you know, you— the way that one works is it triggers off of new emails. So every new email is coming in, it'll add a label to it. So you want to just start with small tasks and you, you know, adding a label is a pretty small way you can start. You know, a next thing you could graduate to is, okay, I actually want you to draft responses. But still drafting responses, you know, it can— it's not going to mess anything up for you. You just get a chance to see, okay, well, this is what the AI would have sent, but it's not sending anything. The worst thing you have is maybe an inbox that's a little cluttered with a bunch of drafts that you don't like. Like, that's the worst output that you have. Of course, the scary one is I want you to go send an email automatically. And so that's where you have to, you know, it's worth starting smaller and then saying, okay, now that I've seen you do, you know, a bunch of drafts, I'm finding that you're really good at these types of emails. And so go ahead and send it. Like, don't, don't create a draft for these. I want you to send this type of email.

Got it. Yeah. It's almost like building trust with a new employee over time, right? You wanna give us some small tasks first and then, yeah, it's pretty interesting.

Yeah. There's a whole Andy Grove, like task relevant maturity thing. And like, I think this applies to agents as well too, where it's like, I'm gonna start small and as you prove yourself, I'll give you increasingly harder tasks.

I love that book. Yeah, that's a very good point. Yeah. Let's kind of switch gears a little bit and talk about like APIs versus like MCP. So, so like Zapier has kind of built a huge business out of just like connecting a bunch of APIs together. Right. And building workflows on top of that.

Yeah.

But, but now these AI products are using these MCPs and are these two like complementary or they're like different or?

I think they're complementary. Yeah. You know, an API is you know, a very specific request that you give it, you know, you give it a very specific set of inputs and you expect a very specific set of outputs in return.

Yeah.

And so it's like just a— it's a very concrete deterministic thing that you do. MCPs, you know, what's the best way to describe it? It's— it lets the agent reason about like what exactly it's going to go call. And so an MCP might actually wrap around an API but it might have access to like multiple API inputs and it might like take in unstructured data as the input, but then format like decide I want to go format it a particular way to hit this and like return a call. And so, you know, MCPs sort of act as that like agent to agent like interface where they can sort of communicate with each other in less structured ways at the end of the day.

So they kind of similar to your determinism versus inference line, right? Like, like API is pretty deterministic.

Yeah, I think they end up complementing each other because of that. Like, I think if you want reliability, cost advantages, you're always going to opt for APIs when you're designing these things out because, you know, you're going to be able to like do exactly what you want the exact way you want it at a high scale, low cost effort. Whereas MCP is— it solves a different problem. It's like you're in these use cases where you may not know exactly what you're trying to, like what the input is coming in, but you know it's gonna be roughly in this shape and you need somebody to reason about it. Now you're gonna spend more tokens, you're gonna have reliability challenges on it, but it's gonna tackle a use case that you couldn't do with an API at the end of the day. And so it opens up these like whole new catalog of things. Now there's obviously an overlap where there's some things you could do with an API, you could do with MCP. It's just sort of like, your choice, right, in terms of how you want to go about it. But I do think they are complementary in their usage.

Yeah, I have like a rule when it comes to building products, like just like do the simple thing first and like, you know, if you can solve your problem with like a workflow or like an API, then just, just, just do that, you know?

Yeah. Well, I think one of the things or one, it's funny you say that because one of the things I do think we're learning is that most people when they say, I want an agent, you realize they actually just want a workflow, like a, just a deterministic automation. And I think what's been exciting about AI for us is that AI has unlocked a lot of people's imaginations where they're like, oh my God, I could do this thing. It's like, oh, you, you always could have done that. You just didn't realize it. But the fact that AI exists has people's minds turning in a way that it maybe wasn't pre-AI.

Yeah. Yeah, exactly. Exactly. Yeah, people were just doing a bunch of work manually. Like, how many jobs out there are you just doing the same job, the same steps every, every day?

Totally.

There's so many jobs.

Yeah, yeah. Nobody stepped back to answer like, hmm, could there be a better way here?

Yeah, exactly. I actually have a really good problem for product building called the 90% rule or something, where, where, um, it's basically like, I think the point is basically like, you know, AI can take you 90% of the way there, but the last 10% maybe still requires some human work. Right? Can you talk a little bit more about that?

Yeah. So I heard this other framework that from someone who talked about like AI is really good at the like middle parts of the work. And so you still need like humans to kind of like kick this off and like figure that out. And you still need humans at the other side to like review it. But there's a whole bunch of like stuff in the middle that usually once you get the inputs well designed, you can like hand it off to an AI or an agent and it can kind of you know, do all the stuff that you want. And then at the end, it's going to spit an output out. And you still kind of want to like go take a look at that and make sure that it's like hitting your standards and your goals and all that sort of stuff at the end of the day. And so I do think, you know, in the future, we're going to have just a lot of jobs where people are reviewing the outputs of these agents and like saying, oh yeah, that looks good, I like it. Or, you know, coming back and like tuning the agent to try and deliver better outputs at the end of the day.

Got it. Yeah. Kind of similar to the email prompt that you were tuning over time, right?

Yeah. Like a great example of this is, you know, we have a lot of workflows that are there to assist our like sales reps. And, you know, I think a lot of folks are trying to figure out how can I take sales out of the process? And I found that that like it doesn't quite work yet. But if you can actually equip the sales rep with all this context, now they can walk into the meeting like really well informed, but they also see the places that the AI's like, ah, that's not, that's not quite how I would talk about it. It's not exactly right, but it's good enough. It's good enough that I can go as a human know, like take that context and do something way more meaningful with it.

Yeah. And, and like if you're building like an AI product on this stuff, like I, I, I feel like the really good ones let you customize the last 10%, right? Like, like I guess a really dumb example is like, You know, when I generate an image using Gemini or ChatGPT, like an infographic, it's never perfect. And then I can't actually edit the components. So I have to just keep trying to regenerate. But yeah, if you can customize the last 10%, that's like super useful.

Yeah. I love, like my, some of my favorite AI products do that where they allow me to go fiddle with the underlying prompt. And so I don't have to like accept their like out of the box defaults, which generally are like decent, but I'm like, I have. I want to make it stylistically more me, you know, at the end of the day. And those are so powerful when I get to come in and fiddle with that stuff.

Yeah. Do you have an example of that?

Well, one example is there's a new voice-to-text tool called Monolog that does this where, you know, it has— when I start the recording, it actually takes a picture of my screen. So it knows like, oh, you're typing into ChatGPT or typing into Slack or you're typing into a Google Doc. And so based on that, I can customize the prompt and be like, oh, this is how my voice and tone sort of works inside of Slack, or this is how it works in a Google Doc, or this is how it works in Superhuman.

Oh, really? Wow.

Yeah, they have some like out-of-the-box one where they know like, oh, this is what a typical email format looks like, or this is what a typical Slack format looks like. But, you know, I have my own style and I want it to sound like me at the end of the day. And so it's nice that they allow me to kind of come in and fiddle with that stuff.

I've been using Whisper Flow a lot, but yeah, I gotta try Monolog. It's from Every, right?

Every— yeah, so I— yeah, it was pretty quick. Like, I, I've tried a lot of the voice-to-text tools and Monolog has like— I don't know, they just— there's just— there's some just nice design touches where it just made it stick for me. And I do find that I think there's going to be a lot more software like that where, you know, functionally it's not all that different, but there's just things that feel— yeah, it's just like it's not a feature. It's just it feels more like what I want to use.

I think the feature is just like, you know, you can, you can make it yours. Like that's, that's the feature, right? You can just customize stuff to make it yours. Yeah, totally. It's like, it's like, um, the Zoom meeting notes feature versus Granola. Like the Zoom meeting notes feature, like they don't let me, they let me pick between predefined templates, but they don't let me just make my own template.

Like just, yeah, totally. And Granola is another great product like that where you just, you know, it works great out of the box, but then the more you use it, the more you get to like make it yours at the end of the day. And it becomes this thing where you're like, I'm just not doing meetings with that granola anymore.

Okay, so, um, your other point was like, there's not going to be like this uber agent that does everything. We're gonna have a bunch of specialized agents, right? And then there's this thing called the, the orchestration layer. Can you tell a bit more about that, like the orchestration layer?

So, you know, I think what we're observing is that the, the more general the agent becomes, like, it can certainly be impressive. Like ChatGPT is a great example. This is like a big, massive, general model that can do lots of general things well. But if you want a tool to do something like one thing very well, it's much better if you can build an agent specifically for that. And to do that, you usually want to give it like the smallest amount of information, the smallest amount of context, the smallest amount of tools that still allows it to complete the job. Because if you played around with these things, you'll realize like the more tools you give it access to, the more context you give it, the more likely it is to get confused, the more likely it is to like burn through a bunch of tokens. And so it doesn't actually make it better. There's sort of this like Goldilocks zone that you want to be in that enables you to build great agents. And, you know, if that's the world we're going to be in, the end result is that if you want to do more complex tasks, the best way to achieve that is to build agents for each part of the end task and then orchestrate along them to figure out how to get these agents to hand off context from one to the other. And now that might be in a deterministic workflow where you're going to say this agent does its part, then it hands it to this agent, then it hands it to this agent, or it might actually be in a more agentic experience where you're going to allow, you know, the first agent to figure out, okay, do I need to send it to this agent or that agent next and allow it to sort of reason its way through it. But I do think we're heading to a world where these systems are going to be composed of a bunch of smaller agents that each do a piece of the overall task.

Yeah. And is that what Zapier is also investing, like the orchestration layer?

100%, right? Like, I think this is where I, yeah, I can't take credit for like our original workflow builder being built for this purpose. It was sort of just a happy accident where it was like, oh, it turns out, you know, this like workflow engine that we built over the course of the last 14 years is really good at connecting agents together as well too. And so, you know, we're able to like build on top of this corpus of knowledge that we've had for over a decade now. And it actually works exceptionally well for this use case.

That's awesome. That's awesome. You know, like, there's all those people on Twitter saying like, you know, when OpenAI launched their product, they're like, oh, you know, rip Zapier, rip, and, and, and the end, and like rip, rip, rip, all this like workflow. And, and, um, what's your response to these people, man? These people are so cringe, dude. Like, how do you—

You know, I think you also saw a similar response, which was from the automation community, the people who actually use these products. And their response was, well, as soon as you logged in and used AgentKit, you could tell these are not the same thing at all. And so, you know, we're in this odd period of time where there's a lot of influencers and content creators out there that their job is to create hot takes and their job is to create clicks and, you know, get stuff. And if you say somebody is dead or somebody is, you know, I don't know, you, you can have that thing happen. And of course, AgentKit looks— and it's a, it's a, it's a canvas, it's a workflow builder. So it looks like Zapier. Even though it performs a very different function. You know, what we're— AgentKit is able— Agent Builder is able to do is help you build better chat agents, like help you extract like a better response than you would get from the general purpose model at the end of the day. What it doesn't do is wake up for any trigger, you know, that you might have. It doesn't allow you to orchestrate across like a wide variety of tools unless of course you're using Zapier's MCP server along with it. and it doesn't allow you to work, you know, sort of model agnostically. And so, you know, I think there's, you know, anyone who's like working in automation, like pretty quickly recognized that, hey, you know, Agent Builder isn't, you know, Zapier or Anytown or Make or any of these other tools. Like it's a pretty, it solves a fundamentally different problem.

Yeah. Yeah. Cause they want to make ChatGPT the app store, right? That's kind of what they're trying to do with like, and get more engagement with ChatGPT.

Totally. Totally. And it's going to help you do better, better stuff inside of OpenAI and better stuff inside of ChatGPT.

Got it. Got it. Well, I mean, if you're an influencer listening to this, like, you might gain some followers, but you've lost credibility with Wade Foster if you do this. So you should not.

Well, I mean, this is where I think it's advantageous to actually go use the tools. And you can, you can tell who is using the tools and who isn't, who is just regurgitating headlines. Now, the irony is like some of these tools are what's causing that. They're like extracting trending topics, regurgitating them back and allowing them to publish them in a slightly different way. Yeah. So there's a bunch of workflow slop that does that kind of stuff, which I'm like, nah, that's not the way. That's not how we're going to do this stuff in the future.

Got it.

Got it.

Okay. So last topic, Wade. So I think Zapier is one of the most AI-forward companies out there. I think you went from what, zero to, I think it's like 97% or maybe 100% now internal AI adoption.

Effectively everyone. Yeah.

Yeah, okay. So, um, let's talk about, let's talk about like how did you actually get everyone to use AI? Is it just like the CEO says so, or how'd you get everyone to do it?

Well, I do think the CEO says so is a good step. So like, I think, you know, as cringe as the like CEO memo is around like we're, we must use AI, I do think it's important. Like you have to say it because if you don't say it, how will people know that this is important and you think that this is the future? So I think you have to do it. I think where CEOs often fail is that's, they just kind of stop there and then they just expect the organization will sort of like magically figure out how to go use these tools. And as powerful as these tools are, they're still just tools. They're— it requires like a lot of organizational change to actually go effectively deploy them. And you know, what has worked for us and what we've seen work for a lot of our customers is a couple quick hitters. One, it's really great if you can do like hackathons or boot camps. And provide space and time for this because now everyone in your organization has dedicated set aside time to like put their hands on the keyboard and make it happen. I do think this is important to do it for everyone in your company, not just engineering too. Like when I say hackathon, I mean like literally get sales, get accounting, get marketing, get, you know, whatever it is to like be playing around with this stuff. And they'll realize that, you know, if you're using a tool like Zapier, oh, it's actually not as hard as I thought it was to go build my first agent or build my first workflow. The second thing is part of those hackathons that's really impactful is have someone do some judging. You do some judging, it creates some accountability. So everyone's going to go, you know, put their best effort in and show up and do something. But even more so than judging, you want to have those demos because it promotes knowledge sharing. So now everybody gets to learn from each other. They get to see, oh, what did you do with that tool? How did that work? Oh, you prompted in that particular way. And I think there's so much So much of these like tips and tricks that people have figured out, but they're not common knowledge yet, that this is the way in which you can get the most out of these tools. And demos promote tons of knowledge sharing. So if you do the hackathon, you do demos, you do that knowledge sharing, the next thing you can do is just as part of all hands, just have somebody come in and do show and tell with something they're building with an agent. And, you know, mix it up. Have someone from marketing, have someone sales, have someone from engineering. Just keep showing off new ways And it starts to create a culture where people see the possibility, see the benefit, and it drives more curiosity through the organization. And I bet if you do this for 6 months, for a year, you will see your company go from, you know, low AI adoption to most people using it within a year because it's just so obvious that once you start building this way, that there are so many valid and awesome use cases.

Yeah.

The other thing that is nice about doing it in this way is it takes the fear away. Right now there's so much fear about what will AI do? Will it take my job? And the moment you create opportunity and space for people to put their hands on the keyboard, the fear fades to the background because they see what's possible. They go, oh, these things are really awesome and it's not as good at this. Like, I'm actually still required for this part and this part and this part. And it just makes it a much more practical, like, effort versus, you know, I don't know, I'm— someone's gonna— some CEO or some executive is going to try and automate my job. And it's like, well, they might want to do that, but like, can they actually do that? And also, wouldn't it be better if you were the one that figures that out? Because if you figure that out, you're going to figure out the next thing and the thing after that. So hands on keyboards is such a great way for that fear to sort of recede back a little bit.

Yeah. And like, you know, I mean, it's going to be a fact that if you use AI to make your job easier, you're probably going to be better at your job than people would don't use AI at all. So might as well learn.

100%. And it's, you know, I think it like this is going to be the new skill set that everyone is going to, you know, need to learn to be in the modern workforce. And what better way to help your employee base upskill than to like provide the opportunity for them to figure that out inside your company? And so that's what I think is like such a great way to go about this is to do the hackathons, to do the knowledge sharing, to do the demos, all that sort of stuff.

But how do you make that part of like, cuz people are worried about, you know, getting promoted and stuff. Right. And like, you know, a lot of people are like, well, I'm, I'm not sure about Zapier, but like at a lot of companies people are just in like back-to-back 30-minute meetings and they're like, okay, like how do you give people time to tinker to stuff and actually feel like they're actually, you know, it's worth doing?

Yeah. Well, gosh, I think that is a trickier question. Obviously you gotta find a way to get people outta back-to-back meetings. Like if you're in back-to-back meetings all day long, you're probably, you're probably going to have a tough time getting work done even pre-AI, because you, if you want to go build things, people have to have time and space to actually go build them. And so I think you have, you don't have an AI problem, you have a just how does our company operate problem. And so I would probably focus on solving that problem first versus, you know, figuring out, you know, how do I get AI in the company? You can just— AI is just the current thing that, you know, you want to go promote.

Got it, got it. Okay, so, so it's about just like giving people time to build, like, even— yeah, makes sense. How about for, how about for, uh, when you're interviewing new job candidates, like for, uh, let's take like non-engineering roles, let's just take like marketing or PM or something, like how do you know how AI fluent— do you give them like a test or something? I don't know how AI fluent—

well, we're still learning honestly how to do this best, but, um, there's a few things that we've evolved through. So yeah, first it was just, hey, tell me what you're doing with AI. And, you know, people would sort of self-report. And early on, you could sort of just get a sense of like, you know, who's done something impressive versus who hasn't. Now we're starting to ask people to actually— we do have a test where I'm like, hey, show us something. You know, it might be, you know, if you're in a PM role, we might give them a task and a prompt and say, hey, you know, you can use AI, you can use whatever tools in an hour. Let's check back in and see how far you've gotten and just see like what they've done and what they've been able to achieve, you know, with, you know, a simple prompt. And you can do that for, you know, other functions as well. Like you can do the same thing for marketing where you might say, hey, I'm going to give you like a document that maybe it contains like a campaign idea, maybe it contains some like, you know, customer data that you're going to have to analyze and say, hey, I want you to come up with a campaign to like target these folks. Like, you know, we'll check back in an hour. You can use any tool you want. Show us what you did. And you just kind of got to watch people work and see like, you know, how they use the tools, how they use their creativity with the tools. And that's probably what's working best right now. But I'm not going to pretend that this is like, oh, yes, this is the best way to value people on their sort of AI fluency at the end of the day. I think there's just a lot of learning that's going on right now for a lot of companies on how to go do this well.

I really like that because a lot of companies are actually doing anti— like, even they want people to use AI at their job, they're doing anti-AI in the interview process. They're like, hey, you gotta be face to face, you can't use any tools, and you guys just show me how you think on a whiteboard.

Yeah, and I mean, like, I get where that instinct comes from. You know, I think one of my concerns is that, yeah, I see this with like documents that come my way where all of a sudden when ChatGPT came out, like the documents all of a sudden just got like much more polished, but the substance of it maybe didn't change all that much. And so you start to worry, oh, are we outsourcing our thinking to the AI? Yeah. And that's the part that is like kind of worrisome, I think, as a CEO is you're like, you can outsource the work to AI, but you can't outsource the accountability. To it. You still need to really understand what is this AI doing. Is it actually solving the problem that I need? Is it going to deliver the results I care about? And at least for right now, that's where humans do seem to have a little bit of a leg up.

Yeah, but I, I will say that, like, maybe I got lazy too, but I will say, like, just being able to talk to AI, like, for, like, strategy or whatever I'm doing, like, it just feels like I can actually unlock my thinking a little bit. Or maybe this is an excuse, But like just going back and forth with it, I, I, I come up with new, new things, you know?

Well, yes. And that, but that's different, right? Like, so you're using it in a smart way. Like you're going back and forth with it. You're providing context. You're probably comparing the output of one model to another, et cetera.

That's right.

And that's very different than somebody who just uploads the prompt you give it and say, hey, can you write a strategy for how Zapier might expand into XYZ market? And then they turn that in as their homework. Like, that's just a very, like, that's, you know, two people generating the same out— like, like getting to the same output, but very different ways. And you're probably getting a much better output than the, the, the latter there.

Yeah. Yeah, exactly. That, that makes a lot of sense. Cool. All right. Well, um, I think this is an awesome conversation. We, we cover a lot. Um, and, um, if I'm just like a, you know, random tech employee or someone, uh, who wanna get started learning about, you know, agent workflows or starting build, building my own agent. Where, where in Zapier can I go? Or what, like, how can I get started with this stuff?

Well, sign up for Zapier. The very first thing you'll see when you get in is a big old box and you can just type your ideas into it. And if you don't have an idea, you can ask it for ideas. And so you can see what types of automation, see what types of agents it might go build.

Awesome. Yeah. And just like use it to solve your own problems, right? Like if you have like a super boring, monotonous task that you do over and over again, that's probably a good idea. To automate.

Yeah, look at your schedule. Like, just go look at it and be like, hey, I spend a lot of time interviewing. I wonder what things I could do to make my interview process work better. Or like, hey, I spend a lot of time talking to customers. Huh, I wonder what things I could do there to sort of make my life a little bit easier. So it's just helpful to look at the things that are in your calendar that you're just doing a lot of. And, um, you know, you can start to come up with like little bits where you're like, I could automate this piece of it, this piece of it. And before you know it, you'll start to realize like, oh, maybe I could actually chain these things together and do something way more impressive. But it's better to just start with a small thing where it's like, hey, could I just generate like a prep doc for this interview for myself? You know, so you start with those like narrower tasks and then work your way up to something a little more impressive.

That's awesome. I think that's a great way to close the thing. So yeah, just spend, spend like, you know, 30 minutes this weekend looking at your calendar, think about what you can automate and use Zapier or use other tools.

Yeah. 100%.

Cool. All right, Wade. Well, thank, thanks so much for your time, man. This is awesome.

Thank you, Peter.