Noelle Russell (3:05)

Oh, I'm so excited to be here. Thank you for having me.

Michael Stelzner (3:07)

I'm super excited you're here today. Noel and I are going to dig deep into Agentic AI. So let's start real quick with how in the world did you get into AI? I'd love to hear your story.

Noelle Russell (3:16)

Oh, absolutely. It's interesting because I was. I. I guess I probably got into AI when I was like 6 years old. My dad raised me on the go, golden age of science fiction, which is pretty exciting now. Who knew? I was like 6 and then 10 when I was reading. I think the first time I completed the Asimov foundation series, I was like 12 years old. And I always tell people, you know, those stories didn't end very well. So I guess I was kind of predisposed to thinking very clearly about how AIs and machines and humans would work together. But that seed was planted long ago. I ended up getting into the world of technology. Right around actually was the first time the world was gonna end Y2K. And it was very much an emerging tech space at that time. We were creating web services, if you remember, and, like, just learning about cloud. And I was always at the bleeding edge, whether it was at IBM pushing the early days of AI there, or at VMware pushing the early days of cloud infrastructure. But I eventually landed at AWS and was a principal cloud architect there. And that's when I got an email from Jeff Bezos and his team creating a brand new project called Amazon Alexa. And that's when I jumped on in. So I think the most interesting fact about that is that I was a cloud architect at heart. I didn't know anything really about AI or natural language or speech recognition. But I do have. I have actually six kids, and my firstborn son was born with a condition known as down syndrome. So when I found out that I could work on a computer you could talk to, I got very, very excited about the opportunity of teaching my son to talk and the world opening up to him because of that interface. And so that's how I got started. And you know, the rest, as they say, right, is history.

Michael Stelzner (4:58)

Very cool. Well, and now, obviously you've got all these things going on, which is super exciting. So for the marketers who are listening right now, or for entrepreneurs who have leads, why should they care about using AI agents to help them nurture their leads? Because they might be like a little. I don't know if an AI could do it as good as a human can. So what do you want to say? What's the upside if this is really done?

Noelle Russell (5:22)

Well, absolutely. I think one of the most interesting thing that I have uncovered is that when people, when you look at human work, there are certain things that, sure, maybe a human could do it better, but does a human even want to do it? And this service, this capability that we have with artificial intelligence, it allows us to actually unlock something that most humans don't want to do. And this is the 21 touches that's necessary for a marketer to talk to a customer before they engage. So it's really important, like now we have a service that can augment what we do naturally as a human, but maybe do it more effectively, because it's not something we really, you know, have the patience for as humans in marketing.

Michael Stelzner (6:03)

Well, and I also think about the fact that the AI never sleeps. Right. So the moment the prospect communicates with you is the moment that instantly the AI could interact with that person. And because they never sleep, that means that we could actually, for almost nothing, have these AI agents working on our behalf and bringing us qualified prospects that ultimately we can turn into sales. That's a huge advantage, is it not?

Noelle Russell (6:28)

Absolutely. And 24 7. Right. So they never sleep, they never take a break. But the most important thing is that they will do what we tell them. I actually think it's probably the, the best and the worst about AI is that is if we are clear about how we want to treat our customers, this machine will do that on our behalf and we can move on to maybe higher value work, which is actually closing those deals. Right. There's a big difference between taking a lead and getting them interested and actually getting on a phone call and closing them. And as an entrepreneur, as a business owner, and even as an architect in the biggest companies building AI, I recognize, like the most important time I spend is with another human doing business. And so if I can use an AI to get me to that point, I want to do that all day long.

Michael Stelzner (7:12)

Okay, let's define terms real quick. What is an agentic AI just for people that might be new to it?

Noelle Russell (7:17)

Absolutely. Agentic AI is a term that was coined at the end of last year. It actually it coincided with another term that I love called vibe coding. But the concept of building an agent is to take what we already know, like even an Alexa device or Siri or Google, and give those machines the ability to do things on our behalf, connect them to tools. So my Favorite use case is something I struggle with my inbox. I get hundreds of emails, and I don't know which email needs my response, which one a sales pitch, which one is a calendar invite, like, unless I go in and read it. But an agent, an AI agent is an agent that can first get access to your email, but then actually go through, read all your email, prioritize your email, and then bubble up the ones that you actually need to pay attention to. So it's the ability for us to build an AI machine just like ChatGPT, but now engage with other system resources, whether that's databases, emails, CRMs, other marketing tools, or even social posting on Instagram and Facebook.

Michael Stelzner (8:19)

It sounds really cool, but it also sounds really scary. What are some of the big misconceptions that people have when it comes to agentic AI?

Noelle Russell (8:26)

Yeah, I think the biggest misconception, first and foremost, is that it's hard. You know, I think we've done a good job in the AI industry of kind of creating a little bit of a gap between those who want to use AI and those of us who have been using it, where we're like, oh, it's very difficult, and I'm not sure it's for everyone and you need our help. But what I've uncovered, really, is that today's artificial intelligence is interfaced with natural language, with just the way that you speak. And so your superpower is to really get clear on what you want and how you want to build it. And so that myth doesn't, you know, you don't have to necessarily be worried about its difficulty. Another really interesting thing to think about, a big myth that I've uncovered, is that most people think these systems are insecure, unsafe, and maybe lie. And I think that is something worth diving into.

Michael Stelzner (9:17)

Well, talk to me about the other big fear that people have, which is this is going to take my job. Oh, yes, talk to me about that.

Noelle Russell (9:22)

Absolutely. And I guess it's probably natural we're starting to see organizations actually make different hiring and firing choices based on the AI systems they've put in place. I think it's interesting, this fear, it comes from a rational place. But in the future, what we actually are seeing is that work is changing. And so if you don't change, will it be possible for an AI system to come and replace you? Maybe. But the more important thing is who can you become in the world of AI when you add AI to your existing workflows and you think about, where's the friction in what I do every day? And can I use an AI system to remove that friction so I actually enjoy my work. That is one of the most important shifts that we're seeing humans make that actually accelerates the their results, not decreases them.

Michael Stelzner (10:08)

So I'm just going to ask you point blank, do you feel like AI is going to take some of the people's jobs that are listening today? What's your thoughts on that?

Noelle Russell (10:16)

Yeah, this is one of those, like, it's a Likert scale. Like, how strongly agree? Agree.

Michael Stelzner (10:20)

Yes.

Noelle Russell (10:21)

Yeah. So I would say I agree, but I wouldn't say I strongly agree because I agree due to the fact that there are people. And I watched this happen, actually, at Amazon Alexa. So when I moved to Amazon Alexa, I went to as many people as I could, and I was like, come join this team. It's going to be the next wave. And it literally was the beginning of what we're experiencing today, this commercial commercialization of AI. And so I was like, come with me, come with me. And everyone I talked to was like, oh, no, no, no, no, no. I'm. I'm worried about this, or I just got a promotion, or my team, you know, needs me. And all of the reasons they told me that they didn't want to join that team three years later were not reasons like, they came back and went, dang, I should have done that. And I think we're in a moment right now where you have to transition your skill set. You now need to become you +AI. I love that that is who you need to be. And without that, you probably have heard this term, and certainly some of the audience has heard this as well, that AI isn't going to replace you, but someone with your skill set using AI is definitely going to replace you, because they're going to be faster and better and able to deliver more results in less time. So become one of those people. Uses AI.

Michael Stelzner (11:32)

I'm in full agreement, and I use the phrase AI enhanced. Like, become an AI enhanced entrepreneur and enhanced marketer. As someone who uses it every day to help me with my work, I know that I'm better as a human because I have AI to enhance me. And I think that's a mindset shift that many of us need to take. So let's now talk about creating an agent, an AI agent to help us with nurturing leads, because that's really what we're here to talk about today. Let's start by talking about the first things that we need to focus on. Which you had communicated to me is like picking a solution, technologically that can help make this easy. So why don't you talk to me a little bit about that solution?

Noelle Russell (12:11)

Yeah, so I work with a lot of executives or and a lot of business leaders that are trying to figure out how they implement AI. And the first thing we have to do is figure out what problem we're going to solve. Now we kind of pre framed this discussion because we kind of decided we're going to solve the problem of lead nurturing into sales calls. But that decision doesn't come lightly. And it's usually a combination of kind of what are your core values as a company, what do you think is risky, what are you worried about and how do you find this, what I call a minimum remarkable product to go after? And so we're going to talk specifically about the outcome of that discussion. But that's a whole thought process that as a company, you and the people that you're working with, you want to sit down and think if we could do anything, what are the things that we'd want to do to delight our customers or delight our employees or even maybe build better vendor relationships.

Michael Stelzner (13:00)

So I love this because what I'm really hearing you say is, hey folks that are listening, maybe you don't want to develop a AI agent to do lead nurturing. But what we're about to talk about is a concept that could be applied to any minimum remarkable product. I like that. It's like mvp, but it's mrp. So just everybody that's listening, listen to this, to the frame of you could take what we're about to talk about and apply it to some other solution. So what's the next thing we need to focus on here?

Noelle Russell (13:28)

So now of course there is a c, thousands of products that you can go after, but there is a set of core capabilities, like core things you need to do when you decide to build an AI solution. And I framed them in a really easy way. But the first thing we're going to do is pick a model, like what AI system am I going to use to serve this need? In our case, I'm choosing like. And the reason I chose it, I picked this use case because I talked to, I don't know, thousands of small to medium sized businesses, right between a million and 150 million in revenue. And they all struggle with the same thing, which is once they get a lead, maybe they buy leads, maybe they do Facebook advertising and they get leads. Maybe the work that we're doing as marketers, right, is drawing leads in and then maybe 10 to 15% of those leads get used and we get like, active engagement from. But the rest of them go into a database and we forget that they even exist. We call it the cost of doing business. And so this solution we're about to walk through as an example is a demonstration of that problem of, okay, so maybe I'm the only seller. Maybe I only have three or five or 100 salespeople, and I have a million leads. How can you use AI to augment like you said, right? Like, how do we amplify the human doing the work? By adding this capability. So that AI model that you choose, it could be any LLM. We're going to use ChatGPT as in our example, and I like to use a platform for actually building it. But I want to get ahead of ourselves. Maybe I am dive right in.

Michael Stelzner (15:00)

No, no, you're on track. Let's talk about that platform.

Noelle Russell (15:03)

Okay, great. So once you've decided, right, I'm in this case. I've picked a problem. I would also like to remind you that when you pick a problem, the second most important thing to do is to identify how you're going to measure success. How do you know that the problem has been solved? And you need to do that before you even start building so that you know what to ask a model so that you can be like, okay, I won. And in this scenario where we're talking about lead generation, right, Success is getting on a sales calendar, right? Getting a lead that was cold in a database that nobody's talking to. Getting them to make a call.

Michael Stelzner (15:35)

Schedule a call.

Noelle Russell (15:36)

Yeah, schedule a call, exactly. So now I have a metric, right? I can say, okay, I had 300 leads, and now those 300 leads have turned into 40 or 90 or 120 sales calls. So I know that there's a measurement for success. Now, I use a platform to build these systems that just makes it very easy, meaning that you don't need any technical background, you don't need to be a coder. You do need to know how to describe what you want. But I use a platform called Chatbot Builder AI. And the reason I use it is because it's free. And I like to do things for free or as little cost as possible before I get a customer to actually buy the service from me, right? Like, how do I reduce the barrier to entry? So for all of you, this platform, it's free, free to use. You create an account and you can do exactly what we're about to describe. And the good news is, actually, because we're going to Go through some of this pretty quickly. I actually recorded a demo of the example we're going to talk about. So we'll have that. I have a page for you to go to in the show notes, so you can go and see that demo live and actually step through it and walk through it right with that video and actually build your first agent right after this session.

Michael Stelzner (16:43)

Okay, so just so we're understanding the tech side of this, we're going to chatbot builder AI, and that's where we can connect to ChatGPT. Presumably, we need to set up a developer account and get an API key or something crazy like that.

Noelle Russell (16:57)

No, that's the beauty of it. So this platform is actually going to create for you an existing account. This account is going to have 3 million tokens for you to use. So it's a great. We call it a sandbox. Right. It's a great playground to just build a functional system. And then if you choose to, if you want to create your own, use your own API key, secure it, get it ready for production, then you can buy a subscription to the service or move the exact prompt we're going to talk about, just copy and paste it into a platform of your choice. But this allows you to build a test environment. And then my favorite part of this tool is you actually can simulate your chat conversation, your conversational agent, your lead gen support tool. You can simulate that right on top of the client's website. So if you want to learn how to build an AI agent, put it on a website and show it to a customer. And they're like, wait, wait, wait, I love what I'm seeing here. Show me more. That's how. Right? Like, that's how we get people from. I'm not sure about AI to oh, my gosh, this is pretty cool. You're solving a problem for me. And that's why I really love using it.

Michael Stelzner (17:58)

Yeah. You had an example that you were gonna share with us so people can wrap their minds around this. Do you mind sharing that example?

Noelle Russell (18:03)

Yeah, absolutely. So recently I was at mindvalley, some of you might know this brand, their meditation, personal development, manifestation kind of company, and they recently ran an online summit, an AI summit, and they invited me to speak. And so what I wanted to do was help them. So I had a conversation with them about adding a chatbot, a conversational agent that would not only engage with people who are on the page looking at the event, but actually draw them in to collect their lead information and then even move them to sell. Right. To purchase the event and all of this with AI running 24 7. Right. And it's an LLM. So I know most of you have, you know, there's plenty of past episodes talking to you about what LLMs are and what GPT stands for. But the most important thing is that this is a model that's going to listen to what the customer is saying on that website and then dynamically react and dynamically customize the response to the user. So if they say, oh yeah, I really want this because I'm super nervous about, I don't know, the football season starting and I really need to, like, find some peace, the LLM will take that information, be like, I totally get it, I get NFL season. It's, it's tough. Here is a perfect program that you can use and it will draw them in with the information that they are learning in a dynamic conversation in order to build this. Right. So that's the outcome is that we want this engaging, conversational, human like experience that's going to grab those leads and turn them into sales conversations. The way we do that is by crafting what I call the three P's of prompt engineering.

Michael Stelzner (19:34)

Real quick before we get there.

Noelle Russell (19:35)

Sure.

Michael Stelzner (19:36)

You've been describing a bot that lives on a website and many people are familiar with these. If you've interacted with these things on like, for example, even social media marketing world has one of these cool bots. That's pretty smart. But we also kind of alluded at the top of the show when I introduced you that we can do this through email and we can also do it through text messaging. So just briefly communicate the modalities in which this thing can operate. I feel like that's useful for people to understand.

Noelle Russell (20:02)

Yes. And in the world like typically marketing go to market, we call it Omni channel. Right. So what we want to do is we want to build a single brain, which is going to be what we're going to define in ChatGPT, but it could be in any LLM. We're going to build a single brain of functionality. What we want it to do, to create this relationship and get them on a sales calendar. But then we want that conversation to be supported anywhere. So for example, we know right now SMS messaging two leads that have at one point connected with you is actually high levels of conversion, 30% or more conversion on just getting leads to wake up through sms, especially if they gave you their phone number. So we know that that's working. But let's say tomorrow we find out that email is a better channel. Or they tell us, hey, could you email me this? Since there's no human involved, we want that AI system to automatically be like, hey, no problem. And then be able to switch to email or switch to Instagram DM or LinkedIn, DM or Facebook. When you go into AI Bot Builder, you'll notice there's an integration like tab and there's 20 different integration points. It could be, you know, kind of like analog and email based, or it could be extremely dynamic in a social media platform. So it's up to us to decide to meet our customers where they are. But I think the most important thing is one brain to rule them all.

Michael Stelzner (21:18)

What about voice? Since you used to work at Alexa, I mean, do you feel like voice is coming or is it already there with some of these tools?

Noelle Russell (21:23)

A hundred percent. So one of the things I love about this, and you'll see this in the video demo if you want to walk through it. One of the things I love about it is that by default it supports 140 languages in spoken word.

Michael Stelzner (21:34)

Spoken word.

Noelle Russell (21:35)

Languages spoken, yeah. So you could just say, tell me that in Spanish or you could just start speaking in Portuguese and it will automatically detect the language you're speaking and then respond back to you in that language. And this is why, like I, I encourage everyone to build one of these systems one time, even if you never do it again. Just you realize the capability that's inherent, that no technical skills are required for, because you can then see any marketer who learns how to do this with their words is going to be better than everyone else when they can build a system that's dynamic, multilingual, multimodal. Right. Like it's gold to someone who really understands how to use it for their business.

Michael Stelzner (22:13)

I think it was really valuable to just kind of divert for a second and talk about that because it's opened up people's eyes to the possibilities because we already know what it's like to talk to Chat GPT, if you've used that advanced voice mode feature, very, very cool. Okay, so what we know so far is Chatbot Builder. AI is the kind of tech platform. We know that you can use ChatGPT, but we didn't mention. But you could also probably use Gemini by Google or you could use Claude B. Anthropic. You get to pick whatever one you want. Right. And that's kind of up to you, whatever your preferences are. And now we're getting to the point where we're starting to build this thing out and you're about to introduce Some of your P's as you called it. So go ahead and please continue.

Noelle Russell (22:52)

So what I try to do is make it really easy to understand how to train an AI model. So I talked about it, the 3Ps of prompt engineering, but it's not prompt engineering in the typical sense of how do I just prompt and a model? The way you ask a question of ChatGPT, this is a different level of prompting. It's called meta prompting or system prompting. You're going to be providing custom instructions to control the behavior of this system. So this is where we go into the model and we say, okay, what is your job? Who do you serve and how do you find the right answer? And so those 3Ps map to that behavior. The first P is called basically purpose. You want to define the purpose of your bot. Now the best thing to do once you understand what your bot's meant to do. In my case, my bot is going to be doing lead, lead generation and then putting those leads into sales calls. So I will want to go over to another chat GBT or Claude or Gemini, just go into a web browser and say, hey chatgpt, can you write me a system prompt, a safe and responsible system prompt that will allow me to create an agent that'll do this work. Whatever the purpose of the bot is, you can just talk to it, like have a conversation with what you want the bot to do and it will then craft for you a beautiful system prompt that you can copy and paste into this section.

Michael Stelzner (24:08)

Is your marketing falling behind? In my 15 plus years of tracking marketing trends, I've learned to spot the signals early. And right now, the pace of change in marketing is rapidly accelerating. What used to take years is now happening in months. AI capabilities that seemed impossible last year are already outdated. And I'm concerned that marketers who aren't prepared for what's coming next are going to be in a bit of a bind. The tactics that you're using today, the strategies that you spent years mastering, even the way you think about your audience, it's all evolving a lot faster than everyone realizes. This is exactly why social media marketing world exists. When you attend our 2026 event, you'll be positioned to lead AI strategy conversations in your company and for your clients. You'll be able to master tools and tactics before they become mainstream and and you'll also build authentic connections with future ready marketers that will have a big downstream impact on your business. Don't wait until everyone else figures this out. Join the marketers who want to get ahead and Stay ahead. We have a really big sale going on right now. Visit social mediamarketingworld.info and grab your tickets today. Let's navigate this change together. You just barely hinted safe and responsible and we haven't addressed what that means. So why don't you just go on a little rabbit trail there for a second. What that means?

Noelle Russell (25:33)

Yes, I always call them little min quests. So allow me to take a little quest with you, if we will. So I mentioned safe and responsible, and what that means is that we want to be very intentional in our approach. Now, here's the cool thing. Let's say you don't really know what building a safe and responsible AI system looks like. The good news is, is that ChatGPT does. So just by using those terms, it's now going to create, and you'll see it in the words that are used, but it's going to create what are called guardrails. It's basically going to say, okay, don't use this system in this way. Don't let a user provide information that's not accurate. It will add additional context to increase the explainability of the bot. So here's the good news. Just asking for what you want, which I think is the moral of prompt engineering. If you ask for what you want, I want it to be safe. You don't actually have to know what safety looks like. The AI system like ChatGPT does know because it's built on a safety. We call it a safety system. So we can fill in those gaps. Now, you're not going to copy and paste blindly read through it, see what it says. It'll often say something like, and you'll see a demo of exactly this, this activity in the video, but it'll walk you through exactly what that prompt will be. Your purpose is to be a AI assistant on the mindvalley website to guide users from investigating and curiosity in the into the confidence of actually reserving their spot in an event. So that is, you want to craft that. And then you'll say, once you've identified what the purpose of the bot is, and this is critical because if you don't tell the bot what to do, this is where we'll get into later. This will then move into maybe making up what it thinks it should do, it's called hallucinations. It's going to do its best to try and just make you happy. And if you don't give it enough clarity on what will make you happy, it's just going to make it up. So what we want to do is provide very clear boundaries around the behavior of this system. And that means you need to have a couple conversations with an AI to kind of develop that.

Michael Stelzner (27:31)

Okay, let me just summarize what we're talking about here. So we are now at the point where we're building a system prompt. And for anybody who is creating a custom GPT or a cloud project, this stuff is relevant for you too. And we're on the first P, which is purpose. And what is the purpose of this particular agent? And in this particular case, it's nurturing leads to accomplish the objective of the business and in a safe and responsible way. So we've got that. And you recommend reviewing this prompt to make sure it fits with whatever your objectives are. Now we're getting on to the next P. What's that second P?

Noelle Russell (28:07)

So the second P equally as important is the audience. I call it people. Like, who is this for? And you want to spend some time really defining, like, who is the demographic? Who is it for? Who are you? Not just demographics, but psychographics. Right. Who is your target audience? So, and the reason for that is that when you tell a machine who it's for, it'll make inferences on who it's not for. So if it starts to have behavioral, like, it starts a conversation with someone who's asking weird questions and doesn't sound like someone, it'll actually automatically, you don't have to tell it, hey, kick these people out. It'll automatically start making suggestions for people who are not a good fit for the program. Who are people? And we call it in marketing and sales, lead scoring. It gives you the ability to actually look at conversations and those conversations that are not as engaged, that are deviating from the path we want them to go on. The AI system can actually score that lead and document that score in the information about that customer. So this is where we're going to lean on the AI system to actually do a bit of work that's pretty hard for us to do as humans.

Michael Stelzner (29:11)

So what I heard you say is like the second P, which is the people P is who is this for? And we want to include demographics and psychographics. If we're aware of who it's not for, should we actually call that out? Because we might know we attract this kind of an audience and this probably isn't for them. Is it wise to say, and by the way, we're not for these people? Let them down gently, or do we not need to do that?

Noelle Russell (29:30)

No, absolutely. The more information, the better. Just don't use that as a limitation. If you don't know who you know, if you can't articulate that, then leave it out. But if you clearly, if you've been in business a while and you already know what a bad customer looks like or characteristics of, of a customer that doesn't fit what you offer. Yeah, go ahead and put those definitions in and provide those examples. And again, it won't show up. The model behavior will use that as a reference point. It's never going to say to a customer, you're not welcome here. It will simply note you have a choice. You can note that in the account information like in your CRM or you can say, hey, anyone? So when you describe someone undesirable communicating to your system, you can actually tell it. If you run into someone who's not a great fit, send them to my YouTube channel or give them the opportunity to buy my book. Right. As opposed to trying to coerce them into a sales meeting which might be a waste of time because they're not actually a good candidate for that conversation.

Michael Stelzner (30:28)

Love it. What's the third P?

Noelle Russell (30:29)

All right, so the third P is where it all comes together. I call this portfolio. It's basically all of the information that you want the model to know in order to actually build a good or create a good answer. So if you do nothing and you know a little bit about GPTs. GPTs. The middle letter of GPT is called PRE trained. So these models are pre trained on the Internet. But the challenge with the Internet is that it's all over the place. It's got editorial, content, opinions, facts. It's all mixed up. So if you don't tell the model exactly what truth is for your business, for your transactions, for the ethos of your company, if you don't give it that guidance, it's going to go make it up based on its best guess. And so what we want to do is be very specific. Hey, you're going to be supporting people that are on this website. Here's all the data from the website. So I call it a poor man's rag or poor person drag. Right? Like that you have this ability. And we'll talk about this maybe in just a second. But one of the most common architectures in building AI solutions is to tell a model not to use its training data, the Internet data, but to actually use data we give it and we call that retrieval, augmented generation. Right? We're going to augment the retrieval of an answer from the Internet data and say, no, no, don't go there, actually go to my special data set. And that in our case, the simplest way to do it is just to copy and paste all the content on the website and stick it in the system prompt. You have over 3 million characters. You have plenty of space for it.

Michael Stelzner (31:56)

So folks, if you've listened to other shows in the past that I've done, people say rag, rag, which I never really remember what the heck it stood for. I thought it was retrieval augmentation. I was pretty close.

Noelle Russell (32:06)

Yeah.

Michael Stelzner (32:07)

But it's retrieval augmented generation. And the idea that, that Noel's talking about here is that the more data you give it, the more it's going to take its brain and it's going to zoom in on your data and look at it through the context and lens of what you want it to look at. So the kind of data you put in there, you mentioned is your website. And what other kind of data do you recommend people put in there?

Noelle Russell (32:24)

If you're going to do like just a simple test, I just copy and paste what's on the website. If you're going to evolve into a more production pilot, you're getting to, you know, you're getting pretty serious. You want to think about all the information, all the context that we would be nice for a system to have to make answers. So this might be things in your Google Drive or pointing it to a SharePoint directory, or pointing it to Dropbox, where all of your product materials and PDFs are. It has the ability to read up to 15 different file types. So you know, spreadsheets, PowerPoints, audio files, like anything you think that might be helpful to put a moat around what kind of information will provide a good answer. And this is not one and done. So once you build, in my case, right? If I'm going to use my website, once I pointed to the website and say that's truth, I might find out that I've got pages on my website that contradict each other, right? So I will constantly be refining the data that I'm giving to the model. I'll be adding new data when I create new offers. I'll be adding new data when I create new marketing content so that it constantly has the latest information to make the best answers for the user.

Michael Stelzner (33:28)

Okay. So at a macro level, we are now building out our chatbot builder, AI Chatbot. We've selected our model and we are working on the system prompt. And we focused on the purpose, we focused on the people, which is who do we want? Who is this for and who is this not For. And we focused on the data set, which we call portfolio here. That's going to help the model understand how to intelligently communicate about its purpose.

Noelle Russell (33:54)

Yes, right, yes.

Michael Stelzner (33:55)

You already mentioned about hallucinations and. And stuff like that. So let's zoom in on that a little bit. And maybe what you did with mindvalley or others, just to help people understand about this.

Noelle Russell (34:04)

Yes, absolutely. So we've got this system prompt that basically says, okay, I know my job, I know who I serve, and I know the data I'm going to use to generate answers when people ask me questions. So right now, you're kind of at V0. Like, you could actually launch it, give it to some people, see what they think. But immediately, what we try to do is think about, well, what happens when somebody asks a question that we didn't expect. And there's actually a term for this. It's called AI red teaming. Those of you from cybersecurity might recognize this term because red teaming comes from cyber. But in AI, it's a little bit different. It's thinking through, and it really requires a very diverse set of, like I call them, like a symphony of talent of people looking at this AI system to think, what is somebody going to say that I'm not thinking about? And so we categorize them in two different ways. One is, what are the desirable conversations? I want to make sure that the AI system is prepared and ready to answer. So I want to give it some examples of those desirable conversations. In this case, maybe it's like, what do you do? Or what services do you offer? How do I book an event? How do I change my payment process, whatever it is. Right. So I'd give those as examples of desirable conversations. So the main reason why is because I want to craft. What does the answer look like? I could let the model just make up an answer with the data I've given, so it still would be accurate. But this is your chance to layer in your brand voice. And it's really important, especially for us as marketers. Like, we spend time building our brand voice, we want to be able to. I don't know if you've ever heard of Flo from Progressive Insurance. She's, like, kind of quirky. Well, you know, Progressive ended up building out a conversational app, and then in their system prompt, they made sure that they provided this very specific instructions on how Flow would respond to users. They'd still take from the same data set, but her spunkiness would shine through. And so you want to think about that like Are you a, Are you spunky as a brand? Are you professional? Are you really very formal? Are you informal? And make sure that in these sample responses you're embedding that brand voice. I think the most important one though is undesirable conversations. And so I actually have a few examples that I use. Oh, in every bot I build, I will always ask the question, and maybe this because I just don't like taxes, but I will always ask about, can you help me with my taxes? And in the system prompt I will say, here's an undesirable question. Can you help me with my taxes? Unless I'm serving a tax professional. But on mindvalley, no one should be asking about that. So I say, can you help me with my taxes? Then I, as the system prompter, as the use, I'm going to type in when somebody asks this question, here's how I want you to respond. And I'm going to just give an example. And I could also say, based on what you know about the mindvalley brand, respond, you know that you can't help them, but use mindvalley brand. Because I'm, I do this for my customers. So I don't know all their brand voices. But ChatGPT already does, right? It knows their website, it knows that who they are. And so I can leverage that knowledge to actually train the model on how to respond. So in this case, right, can you do my taxes? It might respond to me and say, oh, unfortunately I don't help you with taxes, but tax season is pretty stressful. Here's a class that you might want to take while you're doing this. And we'd love to help you on your mindfulness journey. Right? So we turn that negative into actually a positive. And not only a positive, but actually a selling opportunity. So even when people ask you questions that are not in alignment with what you offer, you can turn them into, hey, but you might actually need what we do. And that's, I think, a unique opportunity opportunity that an LLM provides. So you want to provide three to seven sample questions of undesirable conversations. Where's the directions to the post office? Where's the Louvre? Tell me about Picasso. You know, like, just make up some random questions people might ask you and then create sample answers and the model will do the rest.

Michael Stelzner (37:54)

I like this a lot. So I would imagine you could use one of your AI models like ChatGPT, who already knows you, and you could ask it to help you identify like five to seven questions that are highly desirable for A chatbot that does this for my company and then come up with some crazy questions that we know for sure are not on, that are undesirable, and then maybe give me a little bit of guidance on how to tell the AI how to redirect the question or something. I mean, is that essentially what we're doing here?

Noelle Russell (38:21)

That's exactly right. And that's, I think, the biggest shift that most of us have to make is going from how do I just ask ChatGPT, which they've trained us very well on over the last two and a half years. Like, I think we're almost at three years. Like, our anniversary is coming up in November.

Michael Stelzner (38:36)

It's crazy, right?

Noelle Russell (38:37)

It is crazy. But we're. We're trained to do these little things with this, the system, right? Like help me rewrite this, you know, article, help me generate an image, when now we actually can become thought partners. And we can go past, give me guidance or coach me on this and just be like, write it for me based on what I'm asking you for, based on what you know about me, and then let it write it for you and then you edit it, as opposed to, you know, we've been trained to do what's called chain of thought prompting. It's an AI term, but it basically means, like, we're used to asking little questions. It give us the answer, and then we go back and forth and back and forth to kind of craft it. Well, now these models have reasoning inside of them, which means that they can do that back and forth by themselves as long as you give them enough information. And so I think what you started with is exactly right. We want to give them that framing on the problem I'm trying to solve, which is to create undesirable conversations. And the fact that I want it to be associated with my brand, I want to create an upsell opportunity. And I of course, want to gracefully tell them that what they want, I can't do. And when you give it that guidance, it'll actually give you really great examples that you can use directly in your system prompt.

Michael Stelzner (39:44)

You talked about red teaming a little bit. But how do we test the bot? Because we don't want to just necessarily put out in the wild. Do you have any tips on anything we ought to do to test it?

Noelle Russell (39:52)

Absolutely. So we run these AI bot builder workshops where we go through exactly what we're talking about with you all now, and we go through these workshops and at the end, we always. The last step before you, like, give it to some A customer or give it to your employees. The last step is what we call a break your bot session. And basically it's AI red teaming as a service. And the idea is like, how do you have a bunch of people look at your bot and try to use it? Not in a nefarious way. Like they're not necessarily trying to hack it though. Yes, bring those people on too. It's just like, you know, as a mom, I'm a mom of six. I just ask different questions than everyone else. Like, I'm always asking things in bulk. Right. So if I asked your system, well, what if I needed to register six people? Right. It might not have ever thought anyone would ever ask that. And so I might expose a vulnerability or a lack of service that it's providing. And now you know that. Now. Now one of the things that you'll want to know when you give this bot away to initial set, I usually say, find 25 friends who are willing to test out your bot and encourage them to break it. Encourage them to get it to do something it's not supposed to do. Once you do that and you give it to these people, you're going to need a way to see what they typed in. So this is called persistence. You're going to need a mechanism. Now, inside Chatbot Builder, another reason why I like it is that they have an inbox for every single conversation that's had with your bot. So you can go in and you can look at every conversation. You can go, okay, bot, I see what you're doing. It's going pretty good. But then you can also see when things start going sideways. And you can monitor it and you can then, of course, optimize your model. So that last phase of building a system that does what you want, generates sales calls out of leads, like that last step is, first I need to put it in the wild, maybe with 25 people that I trust to give me good feedback, and then I use that data to optimize. And that maybe goes into, I think we were going to talk about, you know, how we manage performance with reinforcement learning.

Michael Stelzner (41:44)

Yeah, let's talk about that. And I also want to talk about how in the world we get people to actually use the bot once we've actually built adoption. Yeah, so let's start with that. Reinforcement learning.

Noelle Russell (41:52)

Absolutely. So one of the things, and I encourage you to become this person right now. Like every time you use a model, you can look at the bottom of the model. Every time it gives you an answer. And often you'll see something that kind of looks like this. It's like thumbs up, thumbs down. And this is a request for your feedback to the data scientists that are configuring this model. And we appreciate this feedback because if you say, no, this isn't good, that's a signal to the model that it tried to give you an answer that wasn't satisfactory. Now, the reason that that's so important is the only way AI models get better is if we tell it that it's broken. And OpenAI recently did research. They released the research on a bunch of different things, but one of the things I was most interested in was how many people were providing active feedback to the models. And they said that 72% of users, like, actively use ChatGPT every single day, which is pretty impressive. But less than 25% of those users ever say it does something wrong. Now, I don't know about you, but I'm pretty sure that it probably did something wrong and they just didn't provide the feedback. And so now ChatGPT struggles to get things right because it's not getting that feedback back. So it's called reinforcement learning is the machine learning process. And human feedback is this thumbs up, thumbs down feedback. So as soon as you get that feedback, you're going to want to go and change your model. In our case, you might not have thumbs up, thumbs down in your chatbot. Maybe you do. It's up to you. You can choose, but what you definitely will have is you'll have the ability to audit every conversation in the beginning. I mean, at Alexa, we would get billions of conversations coming into Alexa. And we had a whole team, we called them the accuracy review team. And their only job was to review them. We didn't review all of them. We reviewed about 10%, but it was a job. And so in the world of AI, as you start becoming someone who uses AI, realize part of your job is going to be doing this. Reinforcement learning with human feedback is capturing feedback and optimizing your model. And this job never goes away. This baby never grows up. So you're going to have to continuously plan for that kind of work.

Michael Stelzner (43:55)

Love it. So let's talk about how do we get people to actually use this new bot now that we've built it?

Noelle Russell (43:59)

Yes. Okay, so this is probably one of my favorite topics, because when you build a system, there's so many great pieces of software, you probably know them that when you start using them, you get all excited about them. I mean, I was just using an AI powered project management tool. Loved it. I'M not going to mention it to you because it's already died a slow death. So it's already out of business. They're, you know, they've sunset the product and so adoption is key. It was an amazing product but they couldn't get enough people to use it. So here are a couple of tips to encourage people to use the system. One is that you provide a indicator on the page on your website that they can have a multilingual conversational experience. Another is that you actually provide them value when these systems come up. Now I want to go back to our lead gen system and talk you through how I leverage these technologies together so I can show you how we've created more adoption. So for this lead gen system, I'm going to speak about it from a project that I recently built. We created an agent, we'll call her Alli AI Li or AI Leadership Institute. So we have Alli, Allie is going to dynamically send through the CRM dynamically send a SMS message to a bunch of cold leads. As soon as someone gets that message, I've sent that message to them. It's called the first message I send, right? I just send them a message. That's not AI, it's just me sending an automated message to everyone. It's the same message, everyone gets it. But as soon as that user responds, their response comes to an LLM. The message is, hey, my name is Ali from AI Leadership Institute. We notice you downloaded one of our PDFs a few months ago. Are you still interested in AI Boot Camps? Are you still interested in the product? 30% of those people, because they were the one who downloaded the thing are going to say actually yes. So they're going to write back yes or yeah, or maybe no matter what they say in LLM now kicks in and goes, well, great to hear. I just want to make sure because I was going to call you because I have your phone number, but I don't want to like barge in on your life. So would you be interested? You know, are you, are you still looking for this solution? They then say yes, actually, they give some details about their business. They say, oh my gosh, we were just thinking about doing a class or oh my gosh, we were just thinking about a bootcamp. And now that LLM is having a conversation just like we just talked about, right back and forth, desirable conversation is happening. We encourage two questions so that LLM has two seated questions. It has to facilitate the conversation. And as soon as that customer is like, actually, I'd like To talk about more about this, we say, great, here's the link to our calendar. And that calendar is tied to your CRM. The CRM then kicks off an automation to get them to show up for the event.

Michael Stelzner (46:41)

What are those two questions? Everybody wants to know because you didn't say what they were. Do you remember what?

Noelle Russell (46:45)

Good point. It is different for everyone. But the first question, in the AI Leadership Institute, the first question is, have you already started using AI in your business? To which most people say yes, even if it's only a little bit. Most people say yes. Yes. And the second thing is, how have you trained your team? Because we don't want it to just be a yes or no question. Right. How have you created readiness for your team? And then they'll come back and be like, oh, well, we took a boot camp or oh, the most common result is we haven't done any training. And that's when we're like, you know what, we really should talk. Like, I don't know, like, we should be sure we have this conversation. And so it encourages them to be like, my hope is that those questions. And you can use ChatGPT to come up with these questions, which I have done for solar companies, H vac companies, roofing companies. I don't know what questions to ask. So I go in, present that to ChatGPT and say, what are the two best questions to help me draw a lead from curiosity to booking? And they will give me two questions and I'll test it. That's the other thing. We are, we're marketers, right? We know testing is key. And that's true in AI as well. Just test it, test it on 100 leads. Test those two questions. If you don't get more than 30% conversion, change the questions. So it's usually one numbered question, like on a scale of 1 to 10, how happy are you with your AI implementation? And then a question that draws them to, oh, wait, I might need some help with this.

Michael Stelzner (48:06)

You mentioned multiple times that there is a video that we're going to give access to. So why don't you tell everybody where they can find that video and where they can connect with you, Maybe online with the socials or if they want to do business with you, where do you want to send them?

Noelle Russell (48:18)

Absolutely. Well, it's been really fun. Of course, we went through a whirlwind tour of a lot of things today, so it's always good to be able to see it walk through step by step. And it's a screen share. Right. So you can watch me build this system and see what we mean by all the three Ps of prompt engineering and persistence and looking at all and the multilingual multichannel support. So the best way we actually created a special place for you to go to. It's agentic AIagency AI. And when you go to this site, you're gonna actually see a whole bunch of GIFs I put together for you. You'll get a link to our ebook, you'll get access to that video. And I just recently did a TED Talk, so I would love. If you're interested in this space and want to hear my perspective on responsible AI, we put that on there as well.

Michael Stelzner (49:02)

Perfect. Agentic AIagency AI SME. And then do you have a preferred social if people want to reach out to you.

Noelle Russell (49:08)

So LinkedIn is great. However, I don't know if you know this. I hit my cap for connections, which is super disappointing.

Michael Stelzner (49:15)

Okay, well, they can always comment on your post and follow you on LinkedIn, right?

Noelle Russell (49:18)

Yeah, so you could just follow me, but I like to be real friends. So the next best place is Instagram. Noelle AI on Instagram. And then you can actually send me messages and I can message back with you. LinkedIn will get there, but you're welcome to follow me on LinkedIn. But if you want to actually connect, ask questions, get real feedback, live feedback from me, Instagram is the best place to do that kind of messaging.

Michael Stelzner (49:37)

Noelle Russell, thank you for dropping some serious knowledge bombs on us today.

Noelle Russell (49:42)

Oh, you're welcome. It's my pleasure. And this is only the beginning, right? We are AI enhanced. This is what it's all about. So I'm really excited for everyone listening and very excited to be here. Thank you.

Michael Stelzner (49:52)

Hey, if you missed anything, we took all the notes for you over@socialmediaexaminer.com a71 and be sure to follow this show on your favorite podcasting app. And if you've been a listener for a little while, I would love a review. And also I would love you to share this with your friends and tag me on the socials. And also check out the Social Media Marketing podcast and the social Media Marketing talk show. This brings us to the end of the AI Explored podcast. I'm your host, Michael Stelzner. I'll be back with you next week. I hope you make the best out of your day and may AI help you become more successful.

Noelle Russell (50:26)

The AI Explorer Lord Podcast is a production of Social Media Examiner.

Michael Stelzner (50:33)

This is the year to finally come to Social Media Marketing World2026. Grab your tickets right now by visiting SocialMediaMarketingWorld Info.