Name: Prompt like a pro: 6 ways RevOps teams are getting more from AI
Uploaded: 2025-12-11T20:30:42.350Z
Duration: 58 min 36 s
Description: Prompt like a pro: 6 ways RevOps teams are getting more from AI

Transcript for "Prompt like a pro: 6 ways RevOps teams are getting more from AI": Hi, everyone. I'm Katie. I'm here on the rev ops team at Zapier and excited to kick us off and talk about six ways rev ops teams are getting more from AI. Go. Just a quick run of show. We will run off some introductions, talk about when AI fits and when it doesn't, how to prompt and automate using AI reporting, and then talk about using AI co pilots to unlock your team followed by q and a. We are recording this session, so don't worry if you miss anything, and we'll break this into chunks. So excited to highlight some people here on our team. Everyone is on our rev ops team here at Zapier. Myself, I'm Katie. I'm a rev ops manager. I do a lot of things, cross sales and success for process systems and quota and commission planning, all sorts of really glamorous work. Sarah is here. She's our senior revenue operations specialist. She is an AI and automation queen. Do not let her fool you. She works really heavily with our success team, but has started working a ton with our sales as well. She's great hands on, working with the team, runs really cool build along sessions, and is so fast to automate anything. Sean is on our MOP side of the house, our marketing operations manager. He works really closely with marketing, with our sales assist team, and is a tools wizard. He can pretty much make anything happen. He understands the in and outs of all of our systems. And I think not really recently, but, like, really recently, it's really started to shine how much he's able to implement AI and how quickly he's able to stand up really well thought solutions. So excited to get into all of this with everyone. And I will start with my own section. I like to think of myself as likely the least technical person on the team. So if you consider yourself less technical, here we go. You're in the right place. Something I've been thinking about a lot with AI and use cases here is when to use AI and when you need things to be deterministic, when you need them to be exactly right. And, and that's become really important for me and my day to day work. So what are the things I think about? I think about, flexible versus flawless. When do things need to be exactly right and when can I leave a little caution to the wind? I start by thinking about what's really time intensive in my day to day and what do I hate doing. The things that I dread are often the things that are really great to be rolled into AI tasks and to use agents for those processes. I think about, like, meeting follow ups and recaps, reports of sprint planning and things like that that have to go to leadership. They're not hard tasks, but they're tasks that I find myself pushing till the very end of the day because I don't want to do them. And those are great use cases for AI. Something else I think about is, can the outcome vary or does it have to be exactly right and perfect every time? A big part of my work is commission planning, and that has to be exactly right every single time. That's very manual. That's very human driven. I can't hope that an agent gets somebody's check right or gets it to payroll on time. I have to make sure that happens. But when I'm thinking about meeting prep or drafting emails, things that I can review myself but can get a jump from a robot to help me out, that's a really great use case for AI. It doesn't have to be exactly the same every time. If it's a message to myself or to my manager, it doesn't have to be exactly perfect every time. So in, like, overall thinking, perfection requires human brain, but you can still siphon off these tasks which are better left to robots. So a couple of examples of things that we do really well here with, with AI and with agents, one being sales call coaching. Our enablement manager, Chris Ondara, put together a really cool process that tracks our reps' Gong calls and reviews them and looks for specific behaviors that we want to see and then help score those calls so that it can go off to the managers and off to those reps to get really quick feedback, and actionable insights on their calls. And it's a great use of their time. A great plug here is that Chris is also hiring. So if you've got a great enablement person, send them our way. So what this looks like, this is just a piece of a larger puzzle, but this is in a Zap, and we use an agent in each section to score. So this section here is on introductions. When I think about prompting, an agent, I think about what data am I giving it and then how am I telling it how to do its job. So in this instance, every section that uses an agent is going to be pulling in the call transcript. And then in the prompt, you'll see here it that it's telling it exactly what to do. It's saying evaluate the sales reps introduction at the start of the call and use this rubric. So we're giving a lot of guidance and telling it how to score, but letting it use its, brain, if you will, to make some calls on its own. And then giving it a little bit more structure saying, read the transcript, score the call, and then give it the one, three, or five. And here's the output that we want. I find in my work that giving it a suggested output is very helpful. It helps standardize what you get every time. Another example, that I've been working on recently is meeting prep. I did a test recently for the last month to take some SDR calls from inbound leads, which was really fun. But I also have a day job, so I don't have time to research every person who books on my calendar ahead of time. So this is another great use case of pushing off to an agent work that doesn't have to be exactly perfect, but is very helpful to get done in bulk. So what this does is it takes my customer data, creates a customer summary, and then sends that prep ahead of my calls so that I can be prepared. I sound like I know what I'm talking about, and I am able to have a more, productive conversation. And that, I actually have a quick demo here that I can kick us over to to walk through how I built a couple different sections for this. Hey. So in this process, what I've done, are a couple things. So I ran this SDR pilot here at Zapier, taking customer calls. And a couple things that I stood up were some call prep and then also some automation to send notes to AEs after the call if I passed off that deal. Like I've been talking about, I think it's really important to weigh what can be exactly right or what needs to be exactly right and what can be left to AI to make judgment calls are a little bit more free form. So the way I build my processes, often are I build Zaps and then I run AI steps within my Zaps. So you'll see in here as an example, this is my call prep. What I do here is I pull my calendar events. I do some filtering, pull out who's in it. I look up my folks in HubSpot as well as the company. And then this point is where I layer AI in. So for me, there are some specific things that I wanna make sure it gets exactly right. So I pump that information into the agent from HubSpot directly. So, this is me, Katie Hewson. I've got my contact email, company domain, if there are free accounts, what their ARR is, industry ICP. And then this is where I let AI run with the information I give it. So I say you're an SDR at Zapier preparing for a first call. So I tell it what it is, what its job is, so it knows what it's doing. And then, I give it this little fun fact that there's Zapier user ID. It's an existing user, and then I tell it to provide clear, actionable insights. And I tell it what those are so it knows where to go when it's researching and what I want back. I want a contact summary, who they are, what their role is, a company summary, what they do, what their industry is, any context, and then automation signals. Do they already use Zapier? Are they using any LLM tools or AI today? Use cases, these are things that we know will help us in our business. And then call strategy. So what I want is to get an output that is tight and structured and immediately usable for my first conversation. SDR is not my day job, so I need to make sure that this is always ready for me to go so I can skim it really quickly, and then be able to run calls effectively, right off the job. So we'll go ahead and run this as a quick little test. And then you'll see here it runs me through a quick output here so I can figure out, what's going on. For me, it's easy to break down and see read it in this way, but you can prompt your, your agent to output in any way that makes sense for you. And then from here, it goes it'll actually pause. It'll drop that into my prep document, pauses this app, and then runs through updating my deal after that. The other side of this process is after I have my call, I can send notes to an AE. And this one, I find to be better to be a little more freeform. It's got my notes collected. It's got the its own prep collected. So what I do here is I just trigger it off of a table here in Zapier when I'm ready to send my notes, runs through HubSpot again, finds out who's gonna own that deal, finds their Slack, and then this is where AI comes in. So for this one, I'm a little bit less prescriptive. I just want it to summarize what I have, what it already came up with for me. So I've got my notes in here. I've got the AI prep summary pulled in, the deal ID, and when the next call is. And this one, I just tell it you're a sales assistant. You're summarizing SDR call prep notes to brief an AE before their first meeting, with a lead. So I want it to just take all my crazy notes, all of the prep that I already generated, And then say, I want you to give me an output that's a fun intro that notes the booked call date, a link to the HubSpot record, why it's qualified, clear overview, insight from the research, and then any other critical takeaways. And then I want it to be concise, scannable, and fun. Use emojis. That makes it better, I think. And then I give it an example format that I want it to come through. I do when I'm sending repeat messages, I want them to be, similar consistency so that they look the same. It's easy to recognize like, oh, this is an SDR call that's booked on my calendar. And then, I just run that through here and, gives me a similar output here. So let's do that. Let me pause here. Okay. So then you get an output here, and then almost live demo. I actually went ahead too and reprompted it, just so you can see my thought process here. I have it formatting now as a Slack message. Sometimes that helps with the output to really specifically tell it what you're going to do with this information. The other Zap is just a dump for me of information that I can read really fast. This one, I want it to look a little bit better. So this runs through and sends a Slack message to the rep, qualified lead, who they are, insights, critical takeaways, so that they can prep for their call and know what they need to do in the handoff. About, work that she's working on, and how she's using AI for automating and reporting. Am I unmuted now? Yeah. Okay. I'm good. Okay. Great. I just had to refresh my screen. Okay. I was ranting about how much I hate reporting. Anything weekly reporting, I always forget to do. Like, every Monday, I don't wanna come into work and have to dig through a bunch of dashboards and do weekly reporting. And so I am going to show a little bit how I use AI to do that weekly reporting. And now I basically don't have to do any weekly reporting at all. It's all AI, and then I'll often go into the Slack threads and tag people and do follow-up. But the process of layering in AI to this has made things so much easier for me, so I'm hoping y'all will be able to benefit from that as well. Okay. So this is kind of like a three part process for me. The first part is making sure you have a really strong prompt to start with. So I'm taking, what data I want, like, what information I'm trying to report on weekly. And I start in an LLM like ChatGPT or Claude, and I have it interview me one question at a time until it has enough information to write a strong agents prompt. So I'll just go in, and I'll show you what that looks like on the next slide actually in a little more detail. Next step after I have a strong agents prompt, I need to consolidate my data. And so I've used things like we use HubSpot as Zapier. So, I'll use HubSpot workflows to trigger and send all the data into an agent at once so that it can consolidate all the information. You can also use Looker or you you might just have to get creative on how you're consolidating that data, but, I'll show you how I do that with HubSpot in a in my little live demo. And then the third step is obviously posting those AI insights. We are a big Slack shop at Zapier, and so we have a lot of different program channels, and I'm posting my weekly reporting in whichever program I'm running. I'm running reporting on that week. Okay. So here's a little bit about how I go into ChatGPT and start to get a strong prompt for a weekly agent. I first give it context on what my job is or the job that the AI is going to be performing. I explain the data that I have, like, what I'm going to be providing the AI on each customer, each data point, whatever you're reporting on. And then I'm describing a little bit of the desired output, so I want a weekly status report for onboarded customers in this prompt. And then I just have the AI. Again, ask me questions one at a time. The one at a time thing is important for me because I hate when it sends me, like, 45 questions in a list, and I'm like, I don't you can't answer all those questions. Just ask me one at a time. And so it helps break it down a little bit and just makes it feel a little more digestible. What I love about this process is that it often asks questions that you might not think about or, like, you might assume AI has this context or might just know something, but it starts to pick at your assumptions about what the AI knows and make sure that you're on the same page so you have a really strong prompt to put into an agent on which whatever reporting you're doing. Okay. So I have a demo on how I've done this fully, and I'm gonna share it. Yes. Okay. So here's an an AI reporting process fully end to end setup. For this particular use case, I am starting with HubSpot deals, and so I know I can start in a HubSpot workflow. HubSpot has, scheduled frequencies, so I'm just scheduling this to go out weekly, end of day Friday, when I know all the renewals that have happened that week have been closed out. You can see I've set the deal filter criteria here. So close date is this week. Deal stage is close one, and deal source is enterprise renewal. The you can see I'm sending all of this to a web hook. The reason I'm doing this is because this is going to enroll each of these deals individually, and I need it all in one place so that I can send to the agent. There might be other ways to do this, and I mentioned earlier, I also use this process in Looker. When you use Looker, you can send all the data as one, like, output from Looker, but with HubSpot, this is sending the deals individually. And so I'm sending this to a Zap. And then in the Zap, I'm going to use a digest step here. This digest step is going to allow me to pull each deal into the same output. So what's this doing? It's looping through every deal that comes through this, adding it to one input. And then once all of those are compiled, so I'm just delaying it for ten minutes to give time for every single deal to flow through here. Then I'm going to release that digest that was just compiled here and send it into this agent. Now this agent is triggered by a Zap, so that is that previous step we saw. I'm going to input the prompt that AI and I wrote together. So back to the, like, interviewing AI to get your ideal agent's prompt. I'll input that in here, and then I'll also give it instructions to start a Slack thread. I think this one it doesn't matter. Anyways, I'm starting a Slack thread, and then I'm put putting in the threaded Slack message, more details. And so AI handles this really well. It has all of the deals that were closed this week. And what that looks like is this message in Slack. So this week's renewal report, how many deals were closed, the ARR impact of those close, the close closed deals, like, the outcomes of each one, and then it'll also give a brief explanation on what happened with those. So, like, you can see significant losses. One of these deals was turned. One of them downgraded because of a competitor, and then you get more details in the thread. So it's tagging you can link out to every deal. I'm linking. You can just click on the deal there. The sales owner is here, and then there's a little bit more information about what exactly happened with this deal. So I can see they had a use case surge. This customer had, some security concerns that we resolved, which is nice. And you can just see some brief explanations on what happened with all these renewal deals. So before, if I were to do this manually, it might take an hour for me to go through all of the deals and, like, read what happened with every deal. But now it's just looking at the information from the deal that I'm sending into this agent and coming up with this really structured weekly report that leadership or sales reps or whoever can come in and engage with really easily. Okay. Cool. Can can y'all hear me? I have backstage people. Thumbs up? Okay. Great. We made it through that demo. Okay. Cool. So, basically, I've done that reporting process for a bunch of different programs. I've done it for renewals. I've done it for onboarding. I have basically, like, five weekly reports at this point going out through this AI process. It means I don't have to do five hours of reporting every week. I can turn my focus to other initiatives, but still deliver on reporting and outcomes on the initiatives I've agreed to in the past. It has made my life a lot easier, so I've been super pleased with that. I am going to pass off to Sean now who will hopefully join me on stage. I. I'm up here now. I'll be great. Sounds good. Awesome. Thanks, Sean. Thanks, Sarah. So, hey, everyone. I'm Sean. So I work on our marketing ops team as, Katie was saying. And what I'm gonna do is a little bit of a different approach, where Katie and Sarah have talked about using AI within the workflows. I'm gonna take a bit of a step back and talk about a topic of, copilots in how you can use AI to actually help you build those tools. So let's just dig right into it because we got a lot to cover here. Now as far as, preparing for this talk, I thought it'd be interesting on our own side to take a look at our own data of how many people are using Zapier's Copilot. In the last three months, I know that we've got about 600,000 people who have tried it. But what I found was really interesting is that about eighty percent of people who tried have given up or only used it once or twice to kinda kick the tires, but never really found the value in it. And in fact, only 2% of people, have been able to really unlock value in copilots to leverage their day to day workflows and really unlock some hidden potential that I think all of you maybe have. So I think the key line here is that barriers are not really the tools. It's more about knowing how to use those tools effectively or knowing how to unlock it to leverage your day to day. So I wanted to talk about that a little bit, today, and I'm just gonna start. Obviously, I work at Zapier, so all my demos are gonna be very Zapier focused. But I also use a lot of AI tools such as Claude, ChatJBT, Gemini, Cursor in my own workflows, But everything I'm gonna talk to you about can be can be, applied to any other tool. It's not so much about using any specific tool, but it's the thinking methodology of how to leverage AI and copilots into any tool within your tech stack. So there's this rev ops reality that I think many of us face where we are all really good at coming up with ideas. We know how to design systems. We know how to orchestrate and put those, ideas on paper to explain to other people. But when we start getting into really custom ideas or complex situations or systems, custom code, APIs, data parsing are often needed and can become a bit of a barrier for us to connect our visions to the end execution. And while we have lots of engineering resources and data, to no fault of their own, we just don't have enough of them. And so that can then lead to us having to put in request tickets, wait long queue times for our requests to get prioritized, and really can put a damper on taking our ideas and shipping something that is going to impact the business. So with AI advancing every day, we're right now at a point where it kind of has an interface that rev ops can really, really handle to start unlocking some of the things that were maybe blocking us in the past. Just to set kind of a timeline stage, I mean, ChatGPT only unlocked AI for the mass public three or four years ago, but it's really been in the last year that the technology has allowed us to really leverage it for agentic AI, which maybe is a bud buzz term you've heard or not. But basically means we've got AI that we can now think for us, create advanced systems that automate automatically run for us, connecting different tools. There's now large context models where we can put tons of data into it and ultimately use it for all of the tools in your tech stack to run things for you. The bottom line being, you don't need to know how to build this stuff all all by yourself or with engineering. You just need to know how to describe what you're looking for. And I think that is a really exciting time for us to be in, especially for those that are less technical. So I wanted to try something. My moment, if you will, for what brought a lot of this to life, because I've been playing with AI for a while. Maybe some of you have too. But I really struggled to find that big use case. So I wanted to, have a big system of all of our data. I needed to connect it. I needed to cross reference different systems, format outputs. And, historically, the old way that I would have to do this is put together a document brief, share it with my engineering team, submit tickets, wait for things to get queued. Maybe some of this sounds familiar to you all. And that could take weeks or months from everything of writing that brief to getting to a point where we've actually had a shipped system process or product. Now with the new way that we're that we're doing this, I wanted to show a couple demos of how I'm leveraging AI to now, write code steps and actually become a partner for me along my journey. So instead of having to go to engineering teams and, request resourcing, I can now have my own dedicated engineer that is able to take my natural language and apply it to any process that I want. And what I'll show you is in two demos here of how I do this specifically around coding. That's been one of the biggest unlocks for myself. I am not a coder at all. I'll put that prefix out there. I do not know how to write Python or JavaScript or or anything, that may rely on a back end engineer historically. But what I do have the ability to do is describe in simple language what I'm looking for, my process to do and then have that AI create those steps for me that then unlock the potential for, all of my my workflows. So I'm gonna start with one example, which is very simple, just to show, a a mental model of how you can think about using this to unlock more technical, functions for you. And then I'm gonna go into a really complex example that is, kind of to show the extremes of where you can take it, and all of these do not require any engineering support, data support. They're all written from, end to end by myself, who has no coding background whatsoever. So let me just queue up the first video here. Alright. So I'm gonna show two demos. One the first one is gonna be very simple just to give you the overall concept of how Copilot can help you, really go far in some of your processes. Obviously, this is a Zapier example, but it works in many other tools. And in this case, I wanna show you how I've used it for unlocking my ability to write custom code. So right now, what this app is trying to do is I get transcripts from Granola, which is, an AI tool that comes with me to all my Zoom calls, transcribes the meeting, takes notes for me. And what I wanna do is convert those notes into markdown that I can then feed into my other AI tool that is always learning and creating a, personal repository of knowledge for myself. So in order to do that, I could manually do a, a couple of steps here or use AI even to convert it. But AI can be expensive if you're using it in large scale, and isn't the best use of it. This is something that we can do programmatically, but I don't know how to write Python or JavaScript, so I'm gonna ask Copilot to do it for me. Can you create a step after the trigger that uses code by Zapier to write a code step that converts the transcript into a markdown format? Make sure to highlight the names of each person, whether it's me or them, and also include a field in the output that is the date stamp for the transcript in a year month day format. Alright. So this is my prompt. Looks good. And let's see what the Copilot does. So right now it's just thinking about how it's going to do it. It's saying it'll create a code by Zapier step to convert. It's decided that it's going to use JavaScript for this, which is great. When I actually ran this the first time, it used Python. I don't think it really matters, so we'll see where the results go for this. And it's just gathering all of the information, opening up the JavaScript script step, and it's saying there we go. It's automatically grabbed all of the inputs from my meeting. Let's minimize that, and let's maybe expand this. So you can see it's actually gone through and written all of this JavaScript code. Thankfully, it uses comments, so that you can learn as you go. Let me hide myself here for a sec. And you can see that it's running all the conversions. It's doing some pattern analysis to format, putting bold around the speakers. Sure. This looks great. Let's give it a run. See what happens with it. And then you can see that it has now converted it all into markdown. And if I scroll to the bottom, yep, there is a, date formatted field there. So I can now use this to map into all my other steps. And, what was that? That was 38 lines of code that I didn't have to touch. I just had to use natural language to explain to AI how to write it. So that's a really simple example of ways to unlock, maybe things that aren't natively able to be done in your processes or in your tools. But now let's go into a much more complex one. So I'm gonna go over to here, and I've got a prompt ready to go. And this is going to that larger system that I had explained to you guys before, where I wanted to collect all of the information about a lead in the previous sixty days. This could be event registrations, form submissions, emails between our our marketing team or the sales reps, website activity, basically everything under the sun about the lead from the last seventy days. And I want this Copilot to now go through, parse it, organize it, clean the data, and compile it into a single summary that we could then pass into ChatGPT. And so you'll see some key phrases that I do when I talk to this is I give it all of the criteria up above of, what I'm looking for it to gather. I give it a requirements over from HubSpot contacts over the last sixty days. And after each step, parse and format each engagement separately because AI really struggles, I find, when you try and get it to do too much in one go. So it's best to take a waterfall approach where each step is very specialized. I'm asking it to default to code steps because I have a hunch just like, in my other processes, it will try to default to OpenAI steps or other AI steps that I've got connected, which I don't want it to do. Those can work, but I think there's a lot that we can do programmatically here that'll be much more efficient. So I give it that requirement. Then once it gathers some information, compile it all into a well structured summary for ChatGPT so that the goal is that it will analyze it. Can you help me build this this workflow? And so I'm gonna give this a run. And hopefully, for the sake of this demo, it will come out I know it's not gonna be perfect because this is a very complex process, but just to show you some of the examples of the code steps that it's going to go and create. So I'm gonna hit start building. And this is gonna take a few minutes to build. So while it is doing that, let's jump over to my completed Zap, that I can walk you through what the end result was. Now right out of the gate, I'm gonna preface that this is version 35 of this workflow. You have to work with the copilots to help give them context and correct things when they don't get them right. But you can see that in this process, I'm collecting all the information when somebody is MQL ed in HubSpot. We pass a webhook. But what is really cool is that we get into this app, and I'll zoom out just to show you. It's it is a large waterfall of steps where we go down all these waterfall, areas and slowly piece together all the information. Within each of those waterfalls, you will find that there is a code step. So if I click on this and take a look at the code, you can see that this is 200 lines of code. And what this step is doing is I asked it first to go out and get the contacts accounts. So every account that they are associated with. And then I wanted it to find out which of those accounts was the highest priority account or on the highest tier plan, most valuable plan. And it wrote all of this Python for me that goes out. It calls API calls here by passing it information. And you can see that it then organizes all of this information, collects the data, grabs the accounts, and, yeah, does quite a lot that some of it I don't understand. Thankfully, there again, there's it leaves comments for me so I can learn as I go. But it's all code that I would never be able to write on my own, and I would need to rely on data or engineering teams to help me. And then we go and do, the same thing. We then run a step to collect all of the email engagements from HubSpot, but I don't know that I need to pass in all of the, data and HTML and CSS and meta stuff that comes within all that data because it is very bloated. So I asked it to run a code stuff that would clean it up. So it takes all that raw data from HubSpot and then passes it through this code that detects CSS, HTML bloat, and email signatures, and ultimately just tries to get to the essentials for the emails. And then it so it strips it all down, processes the email data, and then recompiles it all into an array that I can then pass to our AI staff. And so you start getting the sense of where this goes, where I then run additional steps and, then summarize the email threads for AI, then I go and get all the form submissions, I clean them up, I get all of the products steps. I check if they're on a trial. So for example, again, we come in here and we look at all of the accounts, get the contact. I check for the is trialing properties. If they are, do this. If they're not, do that, and piece together everything that I need through, a waterfall approach that then eventually gets me to a point where I can pass all of this information to an AI step. So I create an agent, collect all of this this information, and then I'm able to pass our AI, in in this case, an agent, all of the information about them and each piece that we separately parsed out, cleaned up, and optimized for an AI agent, all using CodeSteps. So that means that all of these pieces that I used CodeSteps for were not things that I had to, use our AI credits or AI usage tokens for, which can be quite expensive depending on how much you're passing through. And in our case, in MQL and leads, that can be quite a bit. So I'm gonna take a look now at our sample, which trigger approach let's oh, it stopped here. HubSpot. Let's do that. It was asking what trigger I needed to do. So I should have taken a look at that. So let's see if this can quickly whip up, an example. When should this workflow run on MQL event? I don't think it's gonna know what MQL event means, so I'm going to assume that the trigger is just going to be pretty generic and something that I would need to clean up. But for the sake of this demo, let's take a look at what it builds. And so it's going to go forth now, and it's searching through all the actions. I probably am not gonna sit around here and, bore you all with what it builds, but you'll get the idea that it now understands what it's looking to do and would come back with a bunch of steps much like this. And then you would test, iterate, and go through the process that way. So I'll give this one second to see where it gets to, and then we will return back to my slides. Alright. So it looks like it finally finished, and I had to do a little bit of prompting here just to help clean it up as you can imagine. So, some things had froze, so I had to nudge it again. Sometimes you gotta especially for the complex processes, have to nudge the AI. And I went through all these processes, and then I finished up after it had built all this by asking it to rename each step and add notes, which it does great. I love using Copilot for helping my team with documentation. And so you can see all the steps are made and, like, parse event, registration, and attendance, it's gone out. It's going to it's written a code step four. Once we collect all this information from step two, when we go connect HubSpot to it, it would take all that data and pass it through. And all the other steps are very much the same. Now, this was a very blank run. So there's a lot of stuff that we would need to do, such as actually giving it a real trigger to see what the data it receives here, and then it would be even more intelligent at parsing. But you get the idea where one simple prompt, it was able to write, what is that? One, two, three, four, five different code steps of different sets of Python, that it is then able to go through and clean up everything, provide metrics, analyze data for us, and just really create nice systems that, for me personally, would not have been possible prior to having a Copilot assistant that I can now ask, in regular language what I want it to do. And then it just goes out and finds the right approach to do so. So that's that's the demos that I wanted to show you guys just to give you a high level, idea of ways that you can use Copilot. The opportunities are endless. But let's go back to my slides, and I'll show you some, frameworks that I would use as you get started. Sorry about that all. I think the system was glitching a little bit there. Can you all see me now? Alright. So, yeah, as those examples, were showing, obviously, I was leaning heavily on code stuff. So I think that's been one of the biggest unlocks, for myself, from being able to do end to end systems, whether they're simple or very complex. But I think over my time, there have been a a bit of a formula for things that I would share with anybody looking to get started with Copilot. And, this has been alluded to from Katie and Sarah as well as when you're working with any AI, Copilot or not, you need to start with context. You need to give it examples and describe the constraints. Give it the output of what you're looking to do with that data so it has a full understanding of, what it should, do with the request that you're giving it, and that'll give it the best start at performing for you. Next is you're gonna need to iterate, often a lot. In my really complex example, there was 35 versions of that workflow that I had to go through. And then when I did that first blank run, you saw there's only, like, eight steps. So from that first run, I would then go into each step of the process, work through it, go back and forth with the copilot saying I wanted to do this or I wanted to do that or change this method that you came up with and go, and and just iterate accordingly. You're never gonna get it done the right, right the first time. And then lastly, I think this goes for all things AI's question everything. The reason I like co, Copilot and CodeSteps a lot for processes is for those that are can be best done by formulas or determinant work that you don't need an AI to analyze and make intuitive calls or analysis, it can add reliability to your processes and is often more affordable because, you're you're doing formula work. However, do question everything that a copilot does. Sometimes they they try to do things that don't really match what you were expecting, so, you do need to keep an eye on them. And that's everything that I think, we were gonna cover today. So what I wanted to leave you all with, and then we'll open it up for q and a, is I started with the talk about 2% of, power users using Copilots and AI, but every single one of them started with a conversation that unlocked some really massive use case for them. So going into your the rest of your week, think about the processes that are blocking you, and I'd be curious to know what is that conversation that you're gonna start with that is gonna unlock your next level of rev ops value. Alright. So with that, I think we're gonna open it up to q a if anybody has any questions for us based on what we presented. Looks like this one, John, came in while you were chatting if you wanna cover it. Perfect. Yeah. So, Robert, yeah, this is a huge process. I did design it, to be a bit of a waterfall by design. Originally, I had tried to have the AI do everything upfront. And as I mentioned, if you try and have it do too much, it can often get too much noise and get too lost in the weeds, if you will. So it's best when you're working with this goes for copilots, agents, any AI tool, even if you're just chatting with ChatGPT. Give it small pieces and work your way down the line. So in my case, I found out that it was gonna be much more efficient to specialize find event data, find form data, then find the email data, and, like, do each part in, a piecemeal type of fashion to then, at the end, bring it all together for the final passing to an AI step to analyze. If you can do that, and keep that methodology in mind, that'll, often solve a lot of issues that people have with AI when they try and pass it too much. Yeah. One stopping it is that you definitely don't wanna do that with AI. I encourage multi steps for sure. Alright. Anybody else have any questions on whether it's using AI or how you're trying to build workflows, things that have been blocking you? We've got some of the top Zapier builders here for the revopsorg, so feel free to ask us anything. We nailed. it. Awesome. There might be enough uncomfortable silence. Yeah. This is recorded. I think the recording will be sent out to everyone who registered. If that's wrong, someone can correct me. But, thank you everyone for coming, and we're, excited think we got share all this. Oh, here we go. we got one more here, Katie, from Anthony. So have we run into any challenges connecting to bespoke programs or software? I don't know about bespoke software or programs. I mean, we've got our own in house tooling that we do use for some cases where we've got our own APIs. This is something I've actually used some AI's Copilot steps as well because I've got our own internal and private APIs that I've connected to. But I can give access to the Copilot or instructed how to use those. And, again, not knowing how to create API orchestrations or connecting different systems, but it can read those APIs, know learn what the functionalities of those are. And then that those custom type of code steps or custom API calls, can become connected to any of the other systems that we are using. And, again, I don't need to know how to do it. I just need to know how to describe it in regular language in a way that the AI is going to be able to understand to really leverage. Is there anything you would add to that, Sarah? I know you do a lot of custom stuff too. Yeah. I honestly didn't know how to read APIs before I worked at Zapier, like, use APIs, and so that's been a skill I've learned over the past couple of years. And I don't think I've run-in as long as there's an API, I don't think I've run into anything where I'm like, I I have trouble connecting this. But I have used AI to help me navigate API docs before. Like, HubSpot, for example, has crazy long API docs, and so helping, like, narrow down exactly what I want. I will sometimes use AI for that. Yeah. That's and, I mean, that's a great example because I know one thing I've done before too is going with Copilots, or any AIs. I'll often just go to a tool's API docs, grab the URL, and drop that into my message to a Copilot and be like, hey. Here's all of HubSpot's API docs on how to handle company objects. Can you help me build a process that blah blah blah blah blah? And it'll go research the documentation, make sense of it because it makes no sense to me sometimes, and then build what we need to do. So I think as a rev ops professionals, if you know kinda where to find even just documentation, you don't have to understand it, but you know where to find it, and then you can put in language your your idea and vision, it can go really, really far for you and really unlock a lot of options. I see we've got a couple more questions here. Which one? Any examples of how to iterate and improve 35? It's a long slog. How many I can I can answer from mine? I mean, 35 is a lot is it don't be intimidated by that number from mine. So, yes, there's 35 versions, but in all honesty, building that entire complex system took me probably less than a week. In real in in real time, it was probably three to, around three full working days of q and a back and forth with the copilot testing things out, which is nothing in comparison to what the historical method would have been, which it would have been probably spending a day writing out a really detailed brief for an engineering or data team, submitting it to them, and then waiting weeks or months for an open sprint where they can prioritize this work on top of other priorities. And I think that's really the key here is that AI has evolved to a point where it's so knowledgeable and so powerful with being able to take action on the tools that are in your stack that if you know how to describe that brief in prompts, you can actually have a dedicated engineer with you twenty four seven doing everything you want to do. Katie, sir, I don't know if you've got any answers around the number of examples or versions that you guys went through for your stuff, but Yeah. For mine, I I think I used to iterate a lot more. What I found works really well for me now is telling, chat g p t or something what I'm trying to do and then asking it to write a prompt for itself. And I found that that takes a lot of time away from having to iterate and try to fine tune. And then I can kind of fine tune from there and give it a little bit more guidance as I get results. I was gonna say the exact. same thing. I used to have to iterate a lot on the prompt for these reporting things, but ever since I discovered that, like, having Claude or Chatibiti interview you on your prompt, it really I mean, it does a really good job on its own because it's gathering all that context and taking away a lot of the AI guessing what you want out of it. And I I mean, it might have been two or three prompts until I got to that renewal report, and it might have taken me two hours maybe compared to an hour a week. It's already saved me a lot of time, and it's only been running for maybe two months. Cool. We got one more question here. Jessica, I think it was mentioned that you might cover some deal desk approvals and prompts. Any deal, details on those? Katie, I think that falls to you or Sarah. I know one of you. I can speak on this a little bit high level. We don't do a ton of AI within our deal desk. Some things that we've floated around using are, like, being able to, like, chatbot or ask questions, understand, I, what's been approved before, what can and can't work, in your deal negotiations. But it's not one that we use super heavily, Sarah, unless you've got any ideas. think that's our big one. Yeah. I I don't think I can think of anything off the top of my head. I know I'm using some for handoffs, but I don't think specifically deal desk. Sorry. think that kind of falls into one that you don't wanna make a mistake on. You don't wanna tell a customer you can do something that you can't do, give away too many things. You know? Alright. I guess on the prompting, so you you both were saying you talked to Chat GPT about creating prompts and then giving it to your AIs or your copilots. Yeah. I I think, generally, a copy paste can be pretty good. I I will caution sometimes. I don't know if you've seen this, Katie or Sarah. Like, I talked to Chad GPG. The prompt it gives me can be massive sometimes because it's used to a chat interface that is a little bit different. Like, every AI tool has its own kind of flavor and nuance to how to get the most out of it. So I will be usually, when I do that, I will be a little bit more specific or give a constraint of give me a series of prompts. So, again, not batching everything into, like, one prompt to go, but can you give me a series of prompt? Or if it gives me a really long prompt, I'll be like, can you break this up into a series of steps with individual prompts? And then I will give those to the the Copilot one by one to slowly progress through it. Yeah. And I think for Katie and I, we are having Chachipudi interview us for prompting more for AI steps within automations and within agents. So not necessarily the building part, but just, like, what do you want this AI to actually accomplish? That's when Katie and I are doing this copy paste situation. Yeah. And I see the follow-up question here is, like, do you just feed it the documentation? Pretty much. Copilots and AIs these days are great at web browsing. So, what I've done is I just grab the URLs of any API documentation or, any guides or resources that you have, and I I've given it, like, five links before in one prompt. I'm like, here's everything I know about this API. I specifically want to do this type of call to connect this data to this data and do this thing with it. And it will go out, and it'll visit all of those pages, ingest all of the content from the documentation so you don't have to. Again, the whole power of this is that you just need to know what the output is that you're looking for. And if you know where the resources to help guide that output, then you just provide the links to those resources, and it can go out and do that. Lately, one of the other things, taking that even a step further, is if you, explore MCPs to connect to other tools, like connecting it to your Google Docs environment or Google Drive or wherever you your internal documentation goes, you can also then start leveraging it to ask it to go out and visit, the internal docs on, say, system briefs or overall programs that have been written up and collaborated on by your team on top of the API docs that you wanted to go and use, and that can be really, really helpful context for it to go and, go and build with. Yeah. Not a problem. Glad that I answered your question. I think we got what do we got? Three minutes left here, two minutes left. Any other questions? Alright. I think that's enough awkward turtles. I hope all of this was helpful for everyone and just kinda showed how we leverage AI internally. None of us are coders. We are all learning as we go with AI as well just like you all are. So I would encourage you all to explore. It's it's definitely a very powerful tool if you can leverage it and, learn how to harness it for your day to day. You don't need to be intimidated by code anymore. Thanks, everyone. Alright. Thanks everyone for joining.