# 164 How I Built an AI Voice Agent With Claude + Voiceflow (Zero Code)

# 164 How I Built an AI Voice Agent With Claude + Voiceflow (Zero Code) Speaker 2: All right, voice agents. Everybody is talking about them, and if you run a business, you probably had that thought of, I should have a voice agent answering my phones. But the reality is most people have no idea even where to start when it comes to building them and the tutorials that are out there either skip the hard parts or just assume that you know how to code. So on this episode, I brought on Susan Westwater, who is the co-founder of a company called Pragmatic Digital. Susan has been in the conversational AI space since 2017, way before ChatGPT. She's co-written two books on voice and AI and she's even built voice personas for major CPG brands. So today she's going to be sharing her screen, walking us through how to build a fully functional voice agent using a tool called Voiceflow from A to Z with zero code. So you're going to get the exact prompt structure, the knowledge base set up, the conversational flow that she uses with real clients. And if you stick around till the end, she's even giving away the entire system prompt template. So you can literally drop it into Claude, swap in your business details and have a working voice agent in under an hour. You're going to love this one. So let's jump right in. Susan, what are we building today? Speaker 1: Hi. Today we're going to build a voice agent. But I think what's more exciting is we're going to basically operationalize a repetitive business conversation. That sounds a lot more I guess higher level, but today we're definitely going to be taking something where we're going to operationalize a workflow that's been used for phone and appointment handling. We're keeping it simple and we're keeping everything pretty set, but we're going to be using Voiceflow as a tool and we're going to figure out how we can find ways to do more than just customer service calls and kind of breaking out of there. I wanted to look at a way that we could find something that was a little more than dealing with angry people. Speaker 2: Yeah, and voice is a frontier that I think a lot of people are afraid to touch because it is customer-facing for most people. So I think this is going to be an awesome use case. I'm excited to run through it. And before we jump into your screen and you kind of take us through the build start to finish, why don't you just give us your background? Like, why should folks in the audience listen to you? Like, what's your credibility? Speaker 1: Sure, absolutely. So, I have been in the conversational AI and AI space since about 2017 when I founded, co-founded our company, Pragmatic Digital, to start focusing on actually voice marketing and the whole notion of how voice and AI can impact customer experience. We've looked at it from a whole bunch of other ways, but it is near and dear to my heart for a variety of reasons. I've also co-written two books on the topic in addition to a bunch of papers and lead research programs. But we've been working really deeply in the space of everything from Amazon Alexa skills to looking at ways of adding an entire voice persona that would work with systems, not just Alexa, But across them for some major CPG brands and things along those lines. So we've been combining a lot of our experience and knowledge from traditional CX into now what is more of like applied AI for CX. Speaker 2: And, you know, you've been doing this long before it was cool, which I love. I've had some people on recently who have been in the AI space since, you know, I think you might even have been in the longest, but I had somebody recently who's been in it since 2020, actually a guy recently who's been in it for about 10 years. So I love talking to people like you who are, I feel like there's a pre-ChatGPT AI crowd, and then there's a post-ChatGPT, and of course I belong to the post, but... Speaker 1: Yeah, and there's nothing wrong with that. That's awesome. I feel like ChatGPT has helped change conversations in an interesting way. Obviously, there were a lot of different calls and like when people would talk about voice and go, oh, Alexa, oh, Google Assistant. When ChatGPT happened, it allowed us to allow people to kind of change that conversation and say, go play with this thing. And so that was really exciting. It's dangerous and exciting at the same time, which is the best place to be of pushing innovation. And that, I think, is the most exciting thing about that is that the whole notion of how generative AI has shifted. What we have to think about is an exciting time. I think it also helps leave room then for some other things as well and allows people room at the table per se of like, hey, I don't know how to code all this stuff. That's okay. Here's now tools and things out there to make it possible for you to do a lot more than you could previously. Speaker 2: Exactly. Yeah. And that's what I love about what we're going to get into today too. So you're going to share your screen and we're going to look into Voiceflow, which is the tool that you built this voice agent on top of, and we're kind of going to break it down A to Z. So whenever you're ready, let's dive right in. Speaker 1: Yeah, absolutely. And one thing we can even talk about as well as I'm getting ready to share, and I just realized. I have many and many of an account and I'm getting ready to share this and I don't know why it's it's not allowing me to here we go. Let's go. So we're going to go into Voiceflow. Now, I prefer to use Voiceflow because Voiceflow has been around for a very long time. After mentioned, that team has been working on all of these, has been helping with conversation design for quite some time. They're one of the OGs. What they have just recently in the past, I would say, 10 months have done is looked at how do we make this agentic? How do we work that way? And I think one of the other things that we definitely want to think about as we're thinking about those things is this notion of what is the difference between a chatbot and a voice agent because you're going to see in the interface that I'm about to show you, it actually has both. And part of that's just because I'll be honest. Incorporating all of the Twilio on this part for me today wasn't necessarily going to be a primary thing. We'll talk about it, and I can show you how easy it is for those phone numbers. But with some of those things, we're going to also look at it a way of making it a voice-enabled chatbot. But the reality is that What a voice agent does is so different is that it listens, it decides, and then it acts using those rules. And today, we're going to be looking at something where we're going to necessarily, there's three ways to think about the control level of where we talk about being in the loop. So that's all human approved. And I think that's what a lot of people are used to experiencing when they're dealing with agentic of, you know, I'm always going to be involved of, AI is going to gather, but the human is the one who decides and acts. But what we're going to be looking at today is actually something that's a little more on the loop. So the human supervises. So AI is going to decide within those rules, and then we will be supervising those outcomes. We will be seeing that. And so we're more of an operator than we are the driver. So when we think about the other part of it, and when we think fully autonomous, and I think that's where people's brains go of how. And or Jarvis or any of those things, it is this idea of that AI can decide freely and that a human only is going to react if damage occurs. I like to think of it that way, of putting in the idea of if damage occurs, because that actually is typically going to be what is your first question. We're going to talk about how autonomous you're going to want your agent to be. What is the impact of a bad decision? What is the impact of those things? And that helps guide when we think about that in-the-loop type of thing. That's really obviously a lot of regulated things have to just play in that space. And by the way, we can do lots of really exciting things. It's just going to be way more deterministic. We're going to get into more of that traditional scripting of if-thens all over the place as opposed to when we start thinking about on the loop and things like that. So, let's dive in. So, the use case, too, that I've also been looking at today is, like I said, was an idea of coming up with a very simple use case because it's very helpful for us to think about that. And when we think about it from that perspective, when we're getting ready to jump within something, there's a lot of pre-thinking that we're going to want to do. Before we even jump into, let's just jump in and build. And you can see on the bottom here, I've got my Brightline appointment and then full experiences and a couple examples to show how complex you can get. What's really also exciting is that this is all things that someone can do on their own. So let's dive right into Build. And what you can see here is we're looking at a perspective of what we had started to build and put into place was an appointment. And we said, hey, I want to do an appointment set. And if you look at all of this, this looks really amazing, right? It's a lot of stuff. And it's kind of almost intimidating. I wouldn't know how to do all that. Well, here's the exciting thing about when we go into this. And I'm going to start a whole new... I'll show you how easy it is to kind of go from there. So we can do a new project. And it can say, we're just going to call this Live Build. Here's our prompt. This is how easy this can be. I have a prompt and I'm going to actually see if I can do this slickly. I'm going to shift us over to show you where some of that pre-work comes into it. The strength of what we're doing is all in this prompt. That'll be an exciting thing to see of what we can do. We've got live build and now I'm going to move us into where some of my prep work is. Speaker 2: And so I assume you probably have some prompts that you've used in the past to create voice agents. And it's, I mean, I assume it probably like a Claude prompt, right? So use a similar example of like, hey, we've got our prompt bank that's worked well for us in the past. And that's kind of our starting point for creating some of these agents. Like, am I thinking about that correctly? But it's obviously not going to like one-shot it. Speaker 1: Yes, no, absolutely. I believe some of this actually is going to be a little more one-shot than you expect, only because this is us prepping it to be able to get that set up. As we think about our agent instructions, here's how this agent is going to behave. We're saying that this is a customer service support and agent prompt. We're going to start with, and here we're going to say with identity and purpose. First of all, what are you doing? What's my goal? Why am I here? And then, also making it a point to say, here's what your role is. So, we kind of talked about it in the sense of what are roles and goals. We want to make sure that when we're covering that often, an identity and purpose. And I will admit, I've done a bunch of this a little bit ahead of time because nobody wants to watch me type and misspell. So, as we work through this, we kind of say, you know, we're setting some very strict parameters. For this particular instance, I've kept it pretty tight just so that way then we can see how what success looks like and keep it pretty simple. There's a ton of dependencies that can come up into place as we think through this and as we troubleshoot. But so we have said, you know what, you're a virtual automated scheduling assistant. Here's your roles to be able to help us determine not just I have someone who is calling and they want an appointment. I'm actually going to make sure that they should be scheduled. I'm going to make sure that a couple of things about that. And then after I do that, I'm going to book this appointment. I will send them confirmations and I'll be able to get that set up. The goal here is to not troubleshoot electrical problems. And it's really important when we're working with some of these. And we're setting these up to say, here's what you do do, and sometimes it's really important to say, and here's what you don't do. So, don't get excitable like I'm giving you really important boundaries. So, we then move into voice and persona. And with voice and persona, we're looking at an idea of what that personality should be, you know, friendly but brief. But the more specifics you give, the more you can shape this. Because this is just a generic type of person who is going to be called, there's not too much personality that we're getting too deep into, right? If I had an established brand that had like a spokesperson or if I had other things along that line, I might get a little more creative. I might add a lot more information to be like, I need to make sure that you sound like, say, an example would be like, if you think of Tony the Tiger, you need to sound more like Tony the Tiger. We have other places where we're putting in customized voice, but language is important as well. So we kind of cover within there of just kind of what that personality is. And then we move into speech characteristics. And this is part of the speech sculpting to just, do you sound like a human? How much of a human do we want them to sound like? And so we do want something conversational. Most of the time, we try to make that. Now, you'll notice that within there, I had said, you are automated. You're a virtual agent. We will keep that transparency throughout, but that doesn't mean that it needs to be an awkward robot. When we're recovering through it. So we kind of cover through that as well. And again, this is using to your point, this is using some stuff that will start as a base point as a template over time. And then we start to create that. The other nice thing about being able to set this into what we're looking for to happen is that depending on how I'm building this in Voiceflow, I can go back and start to adjust how this bot will or how this agent will behave as opposed to what it knows. There's two really important documents. Like we have all this agent instructions and then you'll see underneath we have another one. Love Google document tabs because it really does make it easier for us to keep things. It's kind of talking through the knowledge base and the knowledge base is the what you know. As opposed, and say what the company does, as opposed to here's who you are, agent. So you will see, then we get into, this is where we start to get these core rules, right? We're gonna confirm this information. You have to collect it. You can't skip fields. Is, you know, escalating emergencies. If someone's calling about an electrician appointment, we need a little, hey, wait a minute. Is this an emergency or not? So we kind of go through all of these thoughts. The more we think about this ahead of time, The easier and the more successful our first shout-out is going to be. So we will have, you know, we also are saying, you know what, we don't want to put an exact pricing because we don't want that promised. By the way, this is just booking an appointment. I'm not giving you an estimate. We don't speculate about safety. And then also, if I've asked someone something twice and they still haven't been able to give it to me and I'm not understanding, I don't want to continue to try to limp along in this agent demo. So that's also one of my fail-safes of when I exit or when I start to transfer to another, to a human. There's nothing wrong if your agent can't actually handle these things. What is wrong or whatever it is, is when it's a janky type of handoff or when you have to make someone repeat all that information. Speaker 2: I mean, I feel like that's been, and sorry to cut you off, but just to kind of add my two cents, like I've had, for example, if I've called Amex in the past, right? And they'll say like, well, what's the last five digits of your number? And I'll say it. And it's like, I'm saying it clearly. There should be no reason that it's not understanding me, but for whatever reason, it's not picking it up. And it's like you say it once, it asks again. You say it a second time, it asks again. And then if it fails the second time, it just, all right, I'm going to transfer you to an agent, right? So that way you're not sitting there kind of playing verbal tug of war with the voice agent. They're just like, all right, you know, we just got to get this guy to a human so that we can maintain this positive customer experience more or less. Speaker 1: Yeah, absolutely. And it's important to make sure that we need to address the root cause and not make our customer feel like they're a part of an experiment, right? Right, yeah. The job is to solve their problem. And some days, you know, if it's not working or they're in a noisy environment, that happens as well. We actually have the opportunity, by the way, to say that, where we then start to go through there from that. But what we also then figure out is what that conversation flow should be. And I add this conversation flow in here of just sort of, hey, this is how I always want you to kind of talk through this. There's ways that we can get variants, but this is a place that we can put that. If they sound upset, that's also something we can mess with as well of making sure that we acknowledge that because that's what good customer service is. Now, as I said, we have that emergency check right in there. Ask right away, hey, before, are you seeing something? And if so, that's where we just end the call and we move on. We've started to already build that immediate. I don't want you to troubleshoot. I don't want you to. And if you don't put some of that explicitly, It will sometimes try to problem solve because the way that I like to think of all of these types of models and what we're working with is they have, they're trying to win the game, right? It's sort of like the movie, you know, not to date myself too much, but that movie War Games where they want to win. They want to get a right answer. And sometimes in their mind, like in that the way that logic flow is, transferring is giving up. So therefore, they still want to stay within there. So we have to put some of those guardrails in there as we think about that. Speaker 2: Yeah, I like how you put that multiple times in the prompt. It's like you do not troubleshoot electricity issues because obviously the example you're using here is for a residential electrician. So yeah, you've got to make that abundantly clear or otherwise it will try to troubleshoot or it will try to kind of step outside what it's intended to do. Speaker 1: Exactly. And again, this is a very narrow case. If I had a customer service bot and I have something to show a different way of how that could be handled, I have another example that's a lot more robust and not built in five minutes. I can show you how that can get deeper and more complex, and you can start to kind of chunk them together. So that they can do that, but this can be that main, this is just that main piece of like, hey, this is an 800 number you can call and you can put that within there. So the next important thing is, can we even help you? So this is where then that qualification. So when we start talking about qualification criteria, this is where then we can also start to put that within that particular piece. And then, you know, these are the pieces of information that you have to collect. This is also if we also have to kind of plan out, hey, if we don't serve them, what do we want to do? What do we want to offer? We could leave this up and not be explicit. And it would probably just say, we don't help you. Have a nice day. We don't know. But if we want to make sure that we're giving someone a good experience, that's also then when we explicitly send them down that path of how do we handle that. And then we have intake questions after that, if we can help them. Okay, these are the important pieces of information that we need to have. And again, you'll say, do not diagnose. It's not your job to solve the problem. Your job is to schedule this appointment. And then we have confirming the details. If it's not emergency, we'll schedule same day. We can offer a human if it's, hey, I need to see someone today. There's also moments too where we can say, hey, if something needs to be escalated. So, you'll see where we have schedule, human, or escalate, which escalation typically is transferring as well. Again, we also have a situation, again, of where we're trying to not just solve the behavior or solve the question. So, it's, hey, I can't give you this information, but I can give you someone. I can transfer you to a human. There's a lot more as I was going through and rabbit-holing this. I'm keeping this very simple, but there are ways that you can also include rates and things like that. Speaker 2: You can obviously get super intricate here, but I like that this is more of a simplified version because this is a great way for anybody listening to start. If you're listening to this or you're watching this and you're like, I've always wanted to create a voice agent for my business or at least like an MVP of a voice agent. I mean, this is the playbook. This is awesome. Speaker 1: Yeah. Oh, great. Yeah. Yeah. And I'm happy to share this with your audience or if you want to, of a way to make this accessible from an idea of this is kind of the general jump on those things. This is a pretty solid We've been using it over time. The other thing, too, that we're looking at is because we all know LLMs can get a little verbose, and I myself included, and I'm not an LLM, is to then say how many words. Keep it as short as possible. I want you to keep that there. Only ask one question at a time. So, when I'm doing that information collection, I don't want you to say, name, address, phone number. Unknown Speaker: Right, one at a time. Speaker 1: Yeah, humans also do that as well. And then using explicit confirmation as well. In this particular case, it's really important because if any one of those pieces of information is wrong, we'll probably have a breakdown in the best experience possible. Speaker 2: Right. Speaker 1: Same with technical jargon. I want you to keep it layman's terms. And then be empathy, but don't apologize. That's a really important piece. I don't know about you, but I cannot stand it when I have an agent that keeps going, I'm sorry, I'm sorry, I'm sorry. Speaker 2: Yeah, I'm like, I don't care. I just want you to fix the problem or transfer me to a human or, you know, get out of this conversation. Speaker 1: Yeah, absolutely. And sometimes when we say empathy, it's amazing how quickly it'll glom onto, it becomes incredibly submissive or just kind of apologetic. And it's like, no, I just need you to not be cold, but I also don't need you to take on the world's troubles. And then we then have also then escalation conditions. Now this is if there's a safety concern, if this person is irate, if it's someone we can't help because it's a commercial property, they want to negotiate prices or we have this, you know, like we, as we mentioned earlier, the not being able to get to the answer we need within two attempts, we let them then, we will move and we will transfer to a human. For frustrated customers, now we're calling out an edge case of like if there's folks that are just upset as they're trying to make this. We've also covered off on that of just trying to take some ownership and just trying to take a moment to pause and help with that. This is, again, an attempt at trying to do a little bit of containment for someone. But there's ways that you can also, if there are certain things that you know are coming through or if there are certain things that let's say not angry, but there's this one issue. There's ways that you can also start to branch things off. And that's kind of what this is a demonstration of. If there are certain scenarios where you're like, you know what, we'll have to treat them a little bit different and I want to be a little more control, have more control, then you can put that within there. And then we have call management. So that's background noise, any of these particular things of letting it help. That way then also there is an opportunity to say, here are the things we do if we have, we can't hear, we need a moment to pause because we have to look something up. Or if let's say the call drops, since we have their number, we can call them back. But let's look at how we're going to handle that as opposed to letting it go into the ether. That becomes the entire document that we have there that goes within here and we have this. What I'm just going to do is I'm going to copy that. Speaker 2: Now, Susan, you did mention potentially making this available to the folks who are in the audience. I mean, is that something we can take you up on? Because for one, I want it, so hopefully I can get it and hopefully we can give it to them as well. Speaker 1: Yeah, no, absolutely. One other thing, too, now that I've copied this that I wanted to call out, too, is I always like to think through the conversation flow as well. So, as I'm thinking through this, is there a greeting? Is there a disclosure? What's kind of that triage? What are the chunks of this to help me understand? Because that's also going to be what helped me fill out my agent instructions. When we think about the process of how we do these things, I got to figure out what the problem is I'm solving first. Figure out what that flow would look like. And then we think about, okay, what do we want the agent to do about it or that? It helps me. It also helps me figure out what information I need or what other pieces. This is a simple, and it's not too, you'll see why it's nice to have like, okay, this is five steps. Because once we move into the creator, so this is that live build where I've said, I hit create new project. We called it live build. Here's the prompt. I just pasted everything in there. Speaker 2: And then we're still seeing Google Docs. Speaker 1: Oh shoot. All right. Hold on. Hold on. Let me let me get back to the right document. That's not fair. It was all magical. No. Unknown Speaker: Let's see here. Speaker 1: There we go. Here we go. So we go from the Google Doc. So before, you'll see before we clicked on New Project, we got Create New Project. I'm just calling it Live Build so I can see. And then I have my prompt. This is where I've gone and I've taken my agent instructions and I paste them in here. Speaker 2: So that whole Google doc that you just showed us, you're basically copying and pasting that entire doc, like that is the prompt. Speaker 1: Yes, right into here. And you'll see what that does. So then what happens, I click generate and you'll see requirements. It's designing the flow. It's optimizing the interactions and it's finalizing it. So you'll see how simple this is without me doing any coding. Speaker 2: And that's what we love here is how do we kind of build these automations or build these tools or build these workflows without having to code. Yeah, I mean Voiceflow is something I've actually used Voiceflow. It was one of the first AI tools I ever experimented with for the purpose of kind of having my own internal knowledge base and having almost like a, basically before custom GPTs were a thing. That's kind of what I was using Voiceflow for. Oh, interesting. Yeah, which is funny because it's like not the tool for the job, but I was just kind of messing around with it back, it's probably a year or more ago at this point. But yeah, I can see it being a very powerful platform for obviously what we're using it for today. Are there other, I mean, I'm sure there are, but off the top of your head, what are the competitors of Voiceflow? Like the other kind of off-the-shelf voice agent builders that don't require coding expertise? Speaker 1: So, there are also ones that are also out there. There's like... Speaker 2: Hello. Unknown Speaker: I'm Laura, your virtual scheduling assistant for Brightline Electric. I'm here to help you schedule a service visit or connect you with our team for any electrical needs you may have. Speaker 1: Boy, she really got excited. That's awesome. So, you'll see here. So, VAPI is another one or VAPI, V-A-P-I. There's also a couple others, I think. ElevenLabs just launched something, I want to say last week, this week. The days run together of how to build your own agents as well. But I like voice, like I said, most of my familiarity is with voice flow, just because of being able to build that. There's also another platform called Voiceify, which is very much a centralized library that you can use for chatbot, voice agent, telephone. I'm going to be covering off on all of those different aspects that you can put where you can put even put that on. You can put it in your apps or you can even put it into your mobile app of having an assistant within there. So there's different places to manage it. And you'll see here, because this is so simple from a perspective of where we started, there isn't going to be that much iteration. And apologies, my mouse is not cooperating today, which is typical. So, based off of what we did and the conversation flow we did, it actually went ahead and mapped out all of this for us. So again, I didn't have to go as far into the into mapping all of this stuff out. Speaker 2: And is that because of the strength of your prompt? I'm assuming if your prompt was much more rudimentary, there probably would be a lot more tweaking under the hood to be done right now. But because you had a really solid, all-encompassing prompt, it kind of did just one-shot this use case, right? Speaker 1: Yeah, and it'll feel like it's one-shotted, but there are spots where we might want to tweak something or we can add something. And there's a couple other things that we have to do as well. You'll see that we've got this whole thing put together, and that's question clarified, failed after two events. This is covering off on all of those particular things. It's covering off on everything to getting us to where we need to be. Now, where is it? Appointment confirmation and wrap-up, or actually where we have customer intake details. One of the other things that we need to put into here is our knowledge base. We have information, but we don't have as much as we normally should. So we click on this and we hit. So I've clicked in here. This is introduction and greeting and I have my agent. So you'll see that I have a little button here called Knowledge Base. You also see here. So here's instructions of what to happen at that particular case. So this is the greeting. Here's what I want you to say, state your capabilities, making sure that this is covered off. Again, The system built this all off of that prompt, like you said, because of having that template. You can also see all of the tools that we have involved right now that you can be able to pull in. So you can work in to where it'll shoot an email. You can make it so that the tool is in there for the 800 number of adding that in, and I can even show you where you're able to put those things. It's not connected. Because of the key, but you just have to enter these keys and it will then connect it for you. Once you've got that on your file, in your account, you've got the 800 number, it will also put that within there. Now, knowledge base, and I'm going to switch to another screen again for you. Speaker 2: And just so I understand correctly, so all the tool integrations there, right? So like make.com was one of them. You know, the Google suite of applications. There's a lot of tools that can be connected. So I guess you can connect a tool to that specific conversation point. So I'm trying to think of an example. So let's say maybe the customer asks the agent to send them an email. If that Gmail tool was connected, the agent could just automatically do that. Is that kind of the purpose of having those tools integrated at different steps? Speaker 1: Yes, and we're going to, when we get to one of the other steps, I'm going to show you how easy it is to do that. That's the nice thing. So this was, again, with the way that this is set up, it's pretty simple. And it's just a matter of a couple of clicks as opposed to embed code. You have all of those options if you have a really complex or you have someone who is more dev and technically minded. But what's, again, what I really like about Voiceflow is it's not a either or. It's a here's a way that you can work within this because you don't have, you know, an IS degree. Here's a way that you can be able to put this through. But we'll start with, let's talk about the knowledge base before we jump into that. Because here, how are we going to tell you about my services? How am I going to tell you what you can do? Well, I'm going to need to enable my knowledge base. Now, if I click on that, you'll see that then that opens it up. And this is saying the LLM description, but we're actually going to Close that because we've turned that on. And we're going to come over here. So over here, we've got our dashboard. We've got our workflow, which is all these one little arrows and tools. Then we have knowledge base. So I've clicked on it. Right now, we don't have a data source. So I was talking earlier about how there's two really important documents. One is your prompt that's in your agent instructions. And then we also then have the prompt that is very important from a perspective of tell me what I can say about this. Knowledge base is also important too if you're saying I don't want you to go and just pull something and make something up. I want you to have definitive facts. So these are the facts around that. It's important with your knowledge base to keep that just the facts because A, I've got a centralized place, so if something does change, we change our hours. I have an easy place to change it. Speaker 2: Right. Speaker 1: And that information. If I were to put that all in that prompt, I would have to go back, dig through that prompt and fix all that out. So you can see that there is some information, even though that was a really comprehensive prompt in those agent instructions, there's actually stuff I've held back from a just-the-facts perspective. Speaker 2: So tell me if I'm understanding this correctly. So the big prompt that we copied from the Google Doc into basically to build this flow, like that's our system prompt basically, right? So that's our system prompt. And then what you were showing us a second ago is we also have almost like many prompts at each step of the conversation. Is that correct? Speaker 1: Yes, we do. Speaker 2: Okay. So we've got, again, when it kind of built the branching logic for us of the voice agent, it's like we've got almost like a little mini individual prompt for each step there. And then the knowledge base is essentially the foundation of when it's responding to a customer or when it's going to answer a question. It's consulting the knowledge base to get the content of that answer. Am I understanding all that correctly? Speaker 1: That's correct. That's correct. And each of those steps, we can tell it, hey, I'm fine with you just using general knowledge or, hey, you know what? No, I want you to look and reference the knowledge base. And that's why we've got that. I can show you there's an even more rapid, simple way to create everything off of what we've got. That version is incredibly Open the version we have here, the more you map those things out, the more control you will have, the more deterministic it will be. So when we're spinning some of these up, you know, faster isn't always better. I'm not going to lie, sometimes when, and even we were talking and I think we had said, but I was like, oh, I'm not sure 15 minutes, like spinning one of these up in 15 minutes, I'm not sure that's the best thing to advertise because you break a lot of eggs making that omelet. But the reality is, is that you can get more control. And I actually have a version of this where we're just using, we're one-shotting it. And you can see how, wow, that's really amazing, but also you can see how it's fraught with, I have given up so much control of what this agent gets to say. It depends on your comfort zone. So faster isn't always better. It's just a matter of understanding the limitations of what you need. Speaker 2: Right. Speaker 1: So, now we're going to go back to our knowledge base. And I actually have created that. Here's where I have our knowledge base. So, this is everything. So, this is our source of truth. So, we have determined with our instructions, essentially, here's how I want you to behave. Here's how I want you to handle things. This is what your source of truth is. So, again, this is incredibly simplistic, but like, you know, this is a, it's open to, operated to residential electrician. Here's what you have. Here's our service area. We can also include zip codes. Again, this could be a lot more robust, but for simplicity's sake, keeping it simple of just Just pieces of here's when we have appointments, if we have an urgent situation, here's how we handle that. These are all things, by the way, is the questions that you'll probably keep coming up with, which is why it's also helpful to keep that in one spot where you're like, oh, you know what? When I first did this, I forgot, you know, like business hours. I'm like, wait a minute. If I don't tell this agent what we do with an emergency, then it's not going to know and it's going to say, sorry, can't help you. Let's make sure that we've got that covered for. It feels like a one-shot, but it's over time. We also keep the appointment duration in there of like typically how much just so that way, one, it helps with scheduling on our end, but also so that it can help with when we send over that confirmation. And then we have common services just in general of what those things could be. Again, there's a lot more to it. This can become more robust or less robust depending on what you want to put. Here's what is urgent, what's an emergency situation, making sure that we've got that defined, same-day service, just covering off on all of the things and the questions that they will have. But again, this is just the facts and keeping those two things separate. That way, then if I do need to change, oh, you know what, we're actually going to do We always do a consultation feed. I can just come in and change this and update this instead of having to rebuild anything. So I'm only maintaining one document and that's what's really helpful. So I exported this. I've exported this as a Word doc. So that way then we'll go back to Voiceflow as it's doing that. And what it'll let me do, sometimes it lets me switch packs really quickly and other times not so much. We go back into right here. All right. So I'm going to want to add that data source. So I hit click on add data source. I'm gonna add a doc. You can see I can just say just go to my website. Speaker 2: Oh, you can add a webpage. That's interesting. Speaker 1: Yeah, you can just give it the URL. You can give it a sitemap, you can give it tables, you can give it plain text, or you can even connect it to if you have a Zendesk piece so that it allows that to work. Speaker 2: So could you give it like a Google Doc URL so that way all you have to do is update the Google Doc and it pulls that in or do you have to kind of export that as a PDF? Speaker 1: You know what? I will admit I have not done that. We could see. I'm wondering if with the tabs it will confuse it. Speaker 2: Oh, that makes sense. Speaker 1: Yeah. So that might confuse it. But if you have a solid single page, yeah, I would think that you could. But I'll show you where I've been able to pull it from. So I'm going to pull this. And then, oh my goodness, as you look at all of my downloads. All right. Let's see. Oh, that's the plain text text. Of course, I should be able to find this faster, but of course, that's how it is. This is very interesting to me that it did not. Hold on. Let's redownload it. That's what happens. I was so prepared and then proceeded to... Speaker 2: That's usually how it goes, right? Speaker 1: Yeah. I'm like, I have this sorted, but yeah, yeah, yeah. Speaker 2: No, but it is cool that it can pull, so it can pull from a URL. And I guess it probably makes sense to whatever Knowledge-based documents you give it is like, well, what are you going to keep the most up-to-date? Because when I think about giving it a URL, like if I just gave it my website URL, for example, a lot of business owners don't really update their website that often. Maybe a couple times a year, if that. So it's like, if you're giving the voice agent your website as the knowledge base, but it hasn't been updated in two years, then that's probably not the best idea, where it might be easier, kind of like you said, And we're excited to have a Google Doc or some external knowledge base that you do keep more up-to-date so that way the agent always has the most up-to-date information, right? Speaker 1: Yeah. No, absolutely. And yeah, let's be honest. Most of us don't, do we? We don't keep shoemakers' children most of the time when it comes to websites. Or if you write an article though or something along those lines, that's something you can also do as well. But yeah, let me see why this is... Oh, I see what I did. In my organization, I overly organized and hid it in another file. That's embarrassing, but we'll pull it up real quick. There it is. So I downloaded a Word doc and now you will see, I'm just going to pull this here, and it is just as much as just pulling that in. And you'll see I just dragged and dropped it in. I don't even have to really search for it. So yeah. So then I just, you can add certain things in there. So chunking, so one thing you can say is if you, the more copious, like this is a really, again, a simple doc. But if I needed to, and I had like certain things within there of, hey, you know what? I want you to keep this logic. I want you to organize this information. I want you to keep it set like this. You know, only use summary. Don't get into the details. I'm giving you a ton more information. Like I'm giving you the math, but I only need you to share the answer. Those are other things you can do as well, depending on what it is. That can be helpful if it's a general customer service bot or something along those lines. There are other times where it's like, you know what? No, don't summarize because they're going to need that step-by-step. We're going to import that. Now we have our voice agent knowledge base. Now we have that set and that will then be available to the rest of the project. Unknown Speaker: So we'll go back to our workflow. Speaker 1: And it's really easy to navigate through here. You also can have your analytics that'll share. And that's the other thing I like about Voiceflow. It's pretty easy interface for me at this point with the familiarity. But going back to this, so showing where we did and we edited agency, now we've got our knowledge base. Now, the other thing that we'll want to do is, as we are taking our customer intake, we're going to want them to collect that information. So, what we do is, this is also going to require an edit. Because I'm going to want to make it, you can do Calendly, you can do a couple other different things, especially if you have your APIs. For simplicity's sake, today you'll see, I'm just going to have it put it into, I'm going to have it add to a Google Sheet. So all I'll need to do is connect to Google Sheet. Speaker 2: And the step that we're looking at here, right, is the information-gathering step. So what you're saying is that as the agent is asking those questions one at a time, like, hey, what's your name? What's the service address? What's the best email? You're saying it's just going to take that information and dump it into a Google Sheet. But we could dump it into our CRM, or we could really dump it wherever we want. Any tool that we can connect to Voiceflow, we can transfer the data. Speaker 1: Yeah. And these are, as you can see with Google, with some of them, there's some super simple ones. There's other ones where it's like, you know what, I'm going to have to just do an API call. And that's not that big of a deal. I believe they also have some MCP stuff in there as well, which can be helpful for some of those other pieces. So I've signed in, signing in, signing in. Now it's going to say, because I've been working, as you can see, I'm going to say, yeah, that's fine. And then it's going to say, all right, well, because it's pulled that sheet, I'm going to manage my sheet. You can see where I've got that. Hold on a second. Oh, because I know why. I was going to show you how we were able to just pick that sheet, but because I've picked it before, it's just going right back. And you'll see that I have this electrician appointment. And you'll see what's been set up here is this is all of the questions, right? And description and pieces like that. And then it's putting all that information in. Now, right now, let me just... So you can see from a previous one that I had before when I was just covering it. It was covering through saying name, et cetera. I do see one thing that's coming in a little funky because it's not putting the timestamp in. But that's something, again, so that's where you were talking about one shot. That's where we start to play with it and say, how do I get you to pull correctly and from there? Speaker 2: Right. Yeah. How do we optimize it further? Speaker 1: We've got this whole piece here, non-services, all of these covering up. Let's give it a run and let's give it a test. Speaker 2: Now, will it talk to us or is it only going to type? Speaker 1: I can actually get this if I wanted to. And because I didn't do the Twilio, but I can show an example of where if I hit call, so that's my cell. I'm just going to set that up because I do not disturb someone, of course. Speaker 2: Oh, yeah. Speaker 1: So I just told it to call me. It takes a moment unless it just, oh, hold on. Do not disturb that and try it again. There we go. Speaker 2: Yeah, I keep mine on do not disturb too, as you should, especially during a podcast. Speaker 1: Yes. Hold on a second. And of course, my do not disturb got in the way first, so that's hilarious. All right, let's try that again. This is the fun. Unknown Speaker: All right. Speaker 1: So I'm going to put this on the speaker. Unknown Speaker: That's perfectly fine. I'm here and ready to help whenever you're ready to proceed. Take your time getting set up with the speaker. Just let me know what electrical service or scheduling assistance you need, and I'll be happy to help you out. Speaker 1: Hello, I'm ready. Great. Unknown Speaker: I'm ready to help you. What can I do for you today? Are you looking to schedule a service visit or do you have questions about our electrical services? Before we get started, I need to ask an important safety question. Are you currently experiencing any sparks, smoke, burning smells, or total power loss at your location? Speaker 1: No. I'm good. Unknown Speaker: Thank you for confirming that. Now I can help you with scheduling a service visit. To get started, I'll need to check if we provide service in your area. Could you please provide me with your zip code or the city you're located in? Speaker 1: Sure. I'm in Clarendon Hills. Unknown Speaker: Great. We do provide service in Clarendon Hills. I'll need to collect some information to schedule your appointment. Let's start with your name. Could you please tell me your full name? Speaker 1: Susan. Unknown Speaker: Thank you, Susan. Could you please provide me with your last name as well so I have your complete name for the appointment? Speaker 1: Westwater. Unknown Speaker: Perfect. Thank you, Susan Westwater. I have your name confirmed. Now I'll need a good callback number where we can reach you to confirm the appointment or if our technician needs to contact you. What's the best phone number to use? Speaker 1: You can use this one. Unknown Speaker: I'd be happy to use that number, Susan. Could you please tell me what the actual phone number is? Speaker 2: I'll need to get this information and add your appointment information. Speaker 1: Yeah, all good. 312-933-6090. Thank you. Unknown Speaker: Let me confirm that callback number. 312-933-6090. Is that correct? Speaker 1: Yep. Unknown Speaker: Excellent. I have Susan Westwater in 312. 933. 6090 confirmed. Now, could you please give me a short description of the electrical problem or service you need? Just a brief overview of what's going on. Speaker 1: You can see how this goes through. I don't know if you want to go through the entire piece, but you can see how the whole thing goes all the way through and how you can work through it and talk through it. I'm going to actually cancel out of that, but it will go all the way through and we can get through that and you can start to see where information. Because we didn't get to the point where I give that info, it's not putting that info in here. You can see that that hasn't hit yet because we haven't hit that part of the flow. Speaker 2: But the fact that you just, I mean, you pretty much just spun this agent up over the last 45 minutes and we can call it and have a full conversation with it and it can save the data and push it to a Google Sheet. I mean, that's pretty impressive to spin that up that quickly. I mean, there's probably a lot of small businesses out there right now who could get this set up in under an hour, same way that you did, right? Speaker 1: Yeah, no, absolutely. And it's all about just making sure that you, the clarity, That's the important part of it is covering off on that clarity and figuring out how to work that. Now, again, this allows us to have different pieces of what information is collected and where it gets covered off on and we can put that in a different sheet or different things like that and get pretty sophisticated. So, yeah, it doesn't take much anymore. And that's what's kind of exciting about being able to dive into it. Plus, you can test and test and test, which is also really helpful. Especially when you're looking through some of these pieces here of like introduction and greeting, we have triggers. So some of those triggers could be that someone calls or that you hear something else that helps. There's one other thing I wanted to call out. Yeah, are these exit conditions? You can also spin that stuff up. So here I have it where it's just a very simple end the call. I could make that exit condition, then you move on to asking something else because they wanted to know the status of their project or if they wanted to know something like that. That's where then you can start to branch. But like I said today, we're going to keep it pretty simple and then you can iterate off of that. Speaker 2: Yeah. Well, no, I mean, this is I think it's the playbook's pretty clear, right? And for folks that want to duplicate it. So I will take you up on on getting access to the Google Docs. So we will. We will put that in the YouTube description. And for anybody listening on podcast platforms, we'll put that in the show notes as well. But I mean, you can pretty much take that. So here's what I would do, right? I would take that document, which is pretty much the template for one of these system prompts for one of these voice agents. And I would give the template to Claude. And I would tell Claude to maintain the template, but then obviously swap out the content for whatever your business is. Save that as a Google Doc and then boom, there's your fully functioning voice agent in a matter of minutes, right? Speaker 1: Yeah, absolutely. And I'll be honest, sometimes when I am like, gosh, this is getting really, really long. Am I getting down too many rabbit holes? I'll actually run it through a Claude or I'll even run it through ChatGPT or any of my other favorite LLMs for the day and mess with it and say, hey, help me make this clear or help me make sure this or what is something else you'd want to know if you were building this agent just to help me make sure that I'm refining it. So using it to refine or even what are some standards? I'll admit that when we worked with some of these before, it's like, what are some standard T's and C's that I'm missing? Or am I missing an edge case that could be a major issue? But yeah, so it's very simple. And the other part of it too that I do like about Voiceflow is they have a lot of documentation. So if we go back to this, the amount of information that you can get of where you have documentation or even Discord, that's really, really helpful. So there's a couple areas too that I can just share here of like my usage right now of what I've got, plans or billing my members. But when I start talking about my phone numbers, There's my Twilio number that I've been able to add. And so you can keep those within there and then it can ask you, what number do you want to use? What piece do you want to use? So there's some information you can also basically store within here. So there's a ton of help though to help anyone figure out how to do this. I am a double English major. So any of these integrations that I was just showing you or things like that, I'm pretty confident anyone can do that. So yeah, I can code, but still, this is a lot of, this doesn't take that much. And there's enough things where you don't hit publish, so that way then you're keeping yourself safe. You know, not to worry about what's out there. Speaker 2: Right. You can kind of sandbox it until you've perfected it and then put it live. Speaker 1: Yeah, absolutely. Speaker 2: Awesome. Well, Susan, this has been fantastic. I've certainly learned a ton. And I mean, the whole idea of voice agents, kind of like I said in the beginning, I think it's intimidating to a lot of people. One, because the stakes are pretty high for most. They don't want to put a voice agent product out there that's going to paint their business in a poor light. So they want to make sure they get it right before they put that into production. And for most people, that means they never do anything with it at all. So if anything, this This interview here has kind of shed light on the fact that anybody can do this, right? You just need a pretty strong system prompt and a platform like Voiceflow and we're going to give that system prompt to folks and they can run with it. So I think this is great. Speaker 1: Yeah, I know. I'm so glad. And there are ways to, like, you heard that it just defaulted to that voice. There's ways that you can change those voices. There's an integration with ElevenLabs, too, if you wanted to use the recording of your voice. I'll be honest, I don't usually typically use my own voice. Just because of all of the things that are fraught with voice security in general, of those pieces, but obviously I'm recording right now, so what have you, but you can come up with some synthetic voices or you can use other model voices that are out there that you can either create through ElevenLabs or there are some that are just built right into the system. So if you wanted to use a more masculine or if you wanted to use a different accent or if you wanted to use a different language. Making sure that there are voices that can handle, say, if I would need something to speak, you know, Catalan, Spanish, I can. It's got the ability to do that. So that's another part, too, as well, of where you can kind of think about what is the voice of this brand or, hey, this is a, we have a spokesperson. We always want to use their voice. That's perfectly fine. There's ways of being able to integrate that within. But again, those are things that spin up pretty fast. So you can kind of go from there. It's all, they have lots of help questions too to help guide you through that. Speaker 2: That's awesome. Well, Susan, thank you so much for your time today. Like I said, I've certainly learned quite a bit. So before I let you go, where can people follow you or find out more about you or work with you directly? Speaker 1: Sure. So our website is www.pragmatic.digital. And that's where you can learn about all of the different areas and flavors of AI, of how we help and the services we provide. We're also in the process of spinning up a couple of programs to help We cater to small to medium businesses because we work with anyone. It's more about mindset for us, quite frankly, but we want to make sure that we're catering to that because there is, like you said, there's that intimidation factor and there's a lot of things that you can do on your own, especially with the knowledge that most folks have done. If you've set up a CRM system, trust me. Any more complicated. The other thing too is I'm pretty active on X just because that's where the AI community lives. And I'm at SJW 75. That's how old my handle is. Yeah, and there's a lot of great brethren who I also follow as well, like Robert Scoble, Brian Romilly. If you really are interested in understanding more about some of the edges of VR, ambient computing, as well as some of the innovation of local AI, since that's been a major hot topic with everything that's going on with Clawbot and all the, you know, every week there's something new. Speaker 2: Yeah, awesome. Guys, we'll put all of her information in the show notes as always and in the description if you're watching on YouTube. Go check out her website. Go follow her socials. And Susan, thank you for the time. We'll definitely have you back to do a follow-up in the future. I can see a lot more people wanting to get into voice agents here in the near future. Speaker 1: Oh, for sure. Yeah, happy to do so and thanks for having me on. Speaker 2: Of course. All right. And for everybody in the audience, we'll be back next week with the new build.