Loading summary
Transcript61 lines
- [00:00]
A
So voice agents are moving from novelty to revenue infrastructure. That is, if you stop treating them like talking FAQs and start treating them like a role, maybe. Qualifier, scheduler, concierge, onboarding guide, retention rep, upsell assistant. That's what we're going to talk about today. Hello, and welcome to another episode of the Duct Tape marketing podcast. This is Jon Jantz, and my guest today is Ryan Murha. He is the founder of yodelfi. Yodel Fi helps creators and brands stay personal at scale by letting followers call and text an AI that speaks in the creator's own voice, grounded in their content library. So, Ryan, welcome to the show.
- [00:50]
B
Hey, thanks for having me.
- [00:51]
A
Did I, you know, I asked you how to pronounce your last name, but then did I pronounce Yodelphi, right?
- [00:56]
B
Yes, it is Yodel Fi. Yeah.
- [00:58]
A
Okay, awesome. So we're talking about voice AI, so let's, let's kind of set the table. There's a lot, you know, there's IVRs, there's LLMs, are, you know, participating chatbots. I mean, so, so what's a voice agent?
- [01:14]
B
Yeah, so a voice agent or. I mean, most agents are just interacting with an LLM. A voice agent is essentially just an LLM that knows it's supposed to respond in a way that's like, naturally speaking. And then you use another tool to have it actually read that text out loud as it's coming.
- [01:37]
A
So, so typically, like, if I had a library, if I wanted somebody to be able to answer questions about my business or my service, I would just give it everything I could. That, that. And then hope, you know, when somebody asked a question, it would access the right thing in giving a response. I mean, is that. Yes. As simple as it comes.
- [01:55]
B
I mean, essentially, that's what it is. So you, you want to build a knowledge base, but there's kind of two componen. So one is, let's say all of the episodes that you've ever done, and we could take all that text and we could feed that to the LLM that it could use for, you know, context. But the other piece is that we also have to make the agent feel like you and act like you in different, in different points that you interact.
- [02:22]
A
So you mean literally, you. Like, it would sound like they were talking to John James?
- [02:25]
B
Oh, well, yeah. So we do also clone the voice. So we could take a lot of your audio and use that to clone your voice. But the thing that we've been finding is that a lot of people will say, here's a Prompt, hey, you're an LLM, be Jon Jantz. And here's all of his episodes, and they're typically getting pretty poor results with that because you as, say, a podcast host, you have a lot of different states. Sometimes you may be, I don't know, explaining something, and sometimes you may be asking a question or pushing back on something. And so what we try to do is we try to have a few different LLMs that an agent can call on and can be different versions of you and have different access to pieces of knowledge that you may need at a certain time. So that way it sounds like you, it feels like you, it responds like you.
- [03:22]
A
And would it be as simple to say, you know, when I hear you describe that, I'm like, oh, this is when John's feeling kind of sad. This is when John's having a good, really good day and happy. Or is it really more. This is John in his sales hat and this is John in his customer service hat.
- [03:38]
B
Yeah, exactly. It's going to be the latter. And that's what's going to make it feel like you're actually speaking to a person compared to, you know, just LLM. Because what a lot of people are used to is speaking with LLM like over a chat window.
- [03:55]
A
Right?
- [03:55]
B
You know, chat GPT or something like that. And that hides a lot of the sort of mistakes. But when you start talking with it, you realize, you know, very quickly, yeah,
- [04:07]
A
yeah, it butchers my name, you know, for example. But yeah, and I think. So where do you think we are in the world today? Are, you know, at one point, you know, people were like, I hate those things, or, gosh, I'm talking to a robot. And you know that. But I, I get the sense that now as more and more. Well, first off, as the technology's gotten better, but as more and more people have, have had good experiences, do you feel like the acceptance to where it's like, I know I'm talking to AI and I don't care.
- [04:39]
B
Yeah, I think people, I mean, I really think like 2026, 2027 are going to be the years of like real voice agents. I think people have been interacting in these chat functions for a while now and they're going to want to start having a more real experience. And kind of like I was describing how we build these agents, it's going to have to be a little bit more tailored to the experience of the user's looking for, I guess where we're at in it. I think we're still actually quite early a Lot of people are not even using any voice agent, for example. Yeah.
- [05:18]
A
So one of the things that I think I picked this up from off of your website. You talk about a voice agent that critically thinks, how is that happening? I mean, again, when I hear that, I hear like, you know, they're, they're actually making decisions. You know, they're not just, they're not just accessing stuff and predicting what you want to hear.
- [05:41]
B
Yeah. So. So without giving away too much of the secret sauce, we, we use like multiple levels of LLMs, right. And within those there's different instructions. Like one may just be orchestrating and another one may be doing an action. Another one may be, you know, calling a certain. A different LLM to give it a response. So we break up all of those tasks to be. So. So that way each LLM call is like, very targeted. And that's kind of the, that's kind of the mistake that we're seeing a lot of businesses, like, fall into right now is they buy a cool AI tool, it looks great in the demo, and then they, they get their hands on it and they're like, this isn't working for me. It's because they're using like a very general package. And the way the LLMs work is like, if you're very specific on what you want, you're going to get much better results. But it can't do too much at once.
- [06:38]
A
Yeah, yeah. You can't just brain dump the entire organization's knowledge base in there, unfortunately, and hope it finds what you're looking for. So I'm curious about this because I have, you know, the way people are buying today is really changing. Right. I mean, they're doing a lot more research. They don't want to do a sales call. I mean, they want to get all the way to. Almost to the point of deciding and then have like a consultation, you know. And so I have a theory that AI agents are going to play a role in that because where people will actually offer them. Not ready to talk to a human, you know, talk to the AI voice agent, they can answer all your questions and they're not going to hard sell you. I mean, they're not going to. Do you feel like there's a point in the buyer's journey where we are. Where that's actually going to be seen as a value add as opposed to a convenience.
- [07:25]
B
I love that you brought this up because we are actually planning on doing this. Yeah. You know, just like when you go to a website now and, you know, a little Chat thing comes up and it's like, hey, maybe I can answer a few questions. Yeah, yeah. The technology is there to be like, you know, take it that much further. And the reality, especially like in software and technology, a lot of the sales and procurement process is just about making sure that you get the legal documents passed back and forth. I think that we're going to see a lot more of the, like those roles focus on that piece and then the answering questions and explaining the product. People don't want to be sold, you know. Yeah, they just want to ask the questions. They want to get to experience it. So in some ways like AI is kind of perfect for that.
- [08:12]
A
Yeah. And, and they can hang up. Right? I mean it's like I'm not getting the answer I want, I'm just going to hang up. You know, it's like I'm not going to be rude to a person maybe, but you know, this, this, I can just hang up on this.
- [08:24]
B
And on top of that you can do that at three in the morning as well. Right. Like you don't have to be waiting for that, that call next week and they're busy or you got to go to this conference and you know, it's instant.
- [08:34]
A
So let's talk that through. Let's, I think you also use the term purpose built. Let's, let's walk through the framework of giving a voice agent a job description and then, and then maybe let's explore what the limitations are. So let's, let's go with a typical kind of business development agent. Somebody buys a, a low cost product on your website and you want to upsell them to the higher cost. You know, can a voice agent reach out or is that really more of a, we're going to train that person to be able to answer anybody's questions that they might have about what's next.
- [09:07]
B
So there's, there's full tools available already that have this like full, we've experimented a lot with, with one of them for building some of our agents just because the functionality that they come with, where they can, they can already call, they can lead the conversation, they'll have sort of what you can like if you can imagine like a timeline and then along that timeline you have different prompts and when the agent gets to a certain like criteria it meets that it goes to the next prompt. And so these tools are very cool. They, you can have a conversation with it and feel like you're speaking with a person and you can get very advanced with it. It can remember your names or your ticket number things and reuse them later and go update the database when it's done. So, and on top of that, you can use it a thousand times at the same second instead of just like an individual.
- [10:02]
A
So do, do, do you. Are we at a point where people, some people are feeling duped? Like, you know, where it's like, I thought I was talking to a per human, and even if they got the result they wanted, it still felt, you know, they, they still felt sort of deceived.
- [10:20]
B
I was on a call the other day and I was trying to ask the person, like, are you a AI agent? And I think they felt offended if they. Because maybe they weren't, but I'm still not convinced they weren't, you know, because. But there's, there's certain tells that, you know, if you speak with these all the time, you're like, okay, there's a delay here and the accent is changing a little bit and things like that. So, yeah, I think people, I think people don't want to feel that they're talking to an agent yet, but I do think that's going to change.
- [10:52]
A
Well, do you think we're at a point where, and I'm not saying disclose it because it's an ethical thing, but just disclose it because people want to. It's a transparent thing. It's like, hey, talk to our AI advisor. They're, you know, they have all the answers for you. So, I mean, it's like right up front, even though it feels like a conversation, I know it's not. I mean, you think that's, that's kind of the crossroads right now?
- [11:18]
B
I don't know. I, I'm, I'm one of those people that, you know, they're like, do you want to share your data? And I'm like, yes, take all my data and customize my experience and things like that. But I could imagine there's a lot of people who want to be very private. Yeah, I think that's going to be a hurdle that we have to, we have to face. And it is going to be a deciding factor, like how people decide to do business with certain companies. You know, it should at least be on the website.
- [11:44]
A
I forgot to tell you when we booked this interview, I do need your Social Security number.
- [11:49]
B
No problem.
- [11:51]
A
Okay. Now, so, so to walk me through, if I came to you and said, ryan, I need this business development agent, like, what's. How. How's the process go? What do you need from me? What, you know, how do we put guardrails on it. I mean, what, what's the pro. How's the process work?
- [12:06]
B
Yeah, so we're always going to start with like a single small use case and try to like nail that down and then kind of build things on top of it. We're also going to just try to like, like, for me, it's very big about matching to a brand in brand voice and making sure that it's consistent with the experience you want your users to have. We build a lot more agents that are in the, like, we have a big one for facilitation. So maybe it's not trying to sell you something, but you still want it to experience like a full facilitator. So what that looks like is breaking down what makes a good facilitator and then building all those different pieces, putting them together, matching it to your brand and letting you use it in your company.
- [12:57]
A
Let's just go with a really, really basic receptionist. I mean, is that a use for this or is that too. Almost too basic?
- [13:05]
B
No, I think, I think basic is good. Yeah. You could definitely, you can have an agent receive a call quickly book an appointment with you. Kind of like what you talked about or asked about. Are people going to feel kind of duped by it? I think there's a lot of scenarios where people are actually going to appreciate it more. Sure. And maybe it takes some time to get there, but I mean, if you can offer me a product at a lower cost and because I speak to an AI agent and. Great, you know, well, and I think
- [13:35]
A
for a lot of routine things that people want to do, I, I know personally things like, you know, once a year I go get contacts, you know, and I just want to be able to go on there and schedule an appointment. I don't want to call somebody, you know, to do that. And so I think there are a lot of things like that that are going to be AI enabled that, you know, that people are going to actually want and appreciate. Because as you said, it's 3 o' clock in the morning. I want to do that, right?
- [14:00]
B
Yeah, yeah, exactly. It changes the game. And it can also be a hybrid approach where, you know, yeah, hit zero if you want to stick to that person. But yeah, I know one of the
- [14:09]
A
fears that people sometimes have is that, you know, the agent, AI agent's going to hallucinate, it's going to be wrong. It's going to actually say something that is maybe counter to the brand. How do you, you know, are there, there's probably some instances where you should never Use this would be one thing. But, but, but how do you also put the guardrails on?
- [14:33]
B
Yeah, so we do put guardrails in the prompts. But I, I'm a big fan of the Gemini models because of that. Even though maybe they're a little bit less fun or something like, to talk to, they definitely hallucinate less. So that's probably the biggest step you can take. But it's also just about being specific. If you give the agent the right context of what it's trying to do, then it doesn't have to go fill in the blanks itself. So a lot of it comes out in testing. We'll find, okay, why did it come up with that? And then we'll go back, we'll revisit the prompts and find out, oh, we maybe overemphasize this or didn't give it clarity on what to do here. One thing you can also do is just. Is give it like a document in your knowledge base, kind of where it can find things. If it doesn't find something, here's some, you know, ways you can respond.
- [15:24]
A
So if you're using Gemini, then could you put a lot of these sources in, like a Notebook LLM or something, or. And then be able to tap it that. That make that be its library?
- [15:34]
B
I have not tried that. I do love Notebook. Do you use it a lot?
- [15:38]
A
Yeah. Gemini does connect directly to Notebook as a source.
- [15:42]
B
Okay.
- [15:42]
A
Now. Yeah, yeah, yeah. So it's. I've been shortcutting training because I'll build the notebook LMS with 300 documents in it, and then, you know, just be able to say source these three. So it kind of gives you. It's a. It's a good. Best of both worlds. Your model is voice and phone number, right? Voice and phone call.
- [16:03]
B
Yes. So the Yodel Fi model is his phone. We can text it. We can also deploy it within the web app, just like the service we're using here.
- [16:14]
A
But there is no avatar, Right. There's no video component to it? No.
- [16:19]
B
The way we see it is that a lot of people are going to want to be able to have conversations with their creator, the creators that they follow. So, you know, maybe when you were a bit of a smaller creator, you could, you know, interact with all of the different, you know, fans and everything and respond to every comment. And then as you get bigger, it becomes more and more difficult. But that doesn't mean people still don't want to communicate. So we can do that with sort of them being able to just text you directly and have conversations and, hey, I'm going through this. What's your take on it? And, yeah, it's not the real thing, but it is, you know, still valuable for a lot of people.
- [16:56]
A
So where do you. Where do you feel like you fit in the category? You know, is eleven Labs a competitor or are they just tangentially related? I mean, where do you fit in the category?
- [17:09]
B
We use 11 labs. So.
- [17:11]
A
Yeah, okay.
- [17:11]
B
Yeah, yeah, they provide voices and they do. Yeah, a lot of great stuff. Yeah, no, we're. We're sort of a. We combine the different pieces, the different tools that these. These producers are making and try to bring them to market. I think. I think there's a lot of cool tools out there, but people haven't figured out really, like, great use cases that are going to enhance people's lives. So we're trying to, you know, meet them there.
- [17:41]
A
Yeah, I kind of laugh at. Some of the tools are like, well, okay, it's cool. It can do that. But like, why? Yeah, yeah. Where. How, you know, would you use that? So if somebody's listening and they're like, hey, I want to try this out, like, next month. What's the. Let me give you a concrete example. I have a marketing agency, so you can use that as the example. What would be the smallest kind of safest experiment that you think a marketing agency could do that would still provide roi, either in marketing or for their clients or just even in efficiencies in the business?
- [18:17]
B
You mean sort of to prototype themselves?
- [18:19]
A
Yeah, yeah. To kind of give it a test. Like a pilot?
- [18:23]
B
Yeah, I would say. I mean, ChatGPT has these. I think they're called GPTs. I think that's a nice way to test. Yeah, I think that's a nice way to sort of test. You can upload a few files and, like, talk with it and be like, oh, is this interesting for us? Definitely have a few customers try it, because there's no point in building something that your customers don't want. And then, yeah, if you're getting a lot of good reactions, then you can, you know, engage us, or we can point you in the right direction to somebody that would.
- [18:52]
A
Well, I guess I was asking specifically about Yodel Fi. Like, if I wanted. If somebody wanted to do a pilot, came to you and said, we heard this show and we want to do a pilot, but we want to start really small. Is there a place that you would say, hey, this is a small, safe experiment that I think you'll get some value from?
- [19:10]
B
Yeah, so what we would do is we would probably do like a single prompt, LLM. So very, very basic, which is basically what I told you we don't do. But it's kind of the starting thing that you can play around with. We'd have like a single prompt. We'd upload a few of your, your files and then we would let you call it and be like, you know, maybe we do like a very quick and dirty, like voice clone. We'll say like, okay, is this interesting for you? Maybe show it to a few customers, get some feedback and then yeah, we have different ways we can price it. We like to be an additional revenue, revenue stream for creators. So, but, but yeah, it could be an ad agency or, you know, we can, we can build all kinds of agents, but for our creators, we try to be an additional revenue stream. So maybe they already have a paid tier and they can kind of incorporate it in there and add 2 cents on or something like that.
- [19:59]
A
Gotcha. Gotcha. Okay. Well, again, appreciate you taking a few moments to stop by the Duct Tape marketing podcast. Is there some place you'd invite people to connect with you, learn more about Yodelfi?
- [20:08]
B
Yeah, so LinkedIn is my main social media, so you can find me on LinkedIn. Ryan Murha. Yeah, we have yodelfi. Com and then that's actually a brand that belongs to another big project, Methotic, which is actually going to be launching here, the beta version. So if you're interested in checking out, AI Facilitation would be awesome to get some beta users.
- [20:29]
A
Awesome. Again, appreciate you stopping by and hopefully maybe we'll run into you one of these days out there on the road.
- [20:35]
B
Sounds great. Thanks for having me.
- [20:36]
A
Thanks, Ryan.