DAY 2: I ❤️ AI Challenge: Build Your First AI Employee

Admit all, admit all. Oh gosh, I'm trying. All right, let's see here. All right, Jacinta, there we go. More. Confirm. There you are. If you could hit the— oh, you're froze.

But I'm here.

All right, well, thank you. Well, I gave it to Tara. No, I'm just kidding. Uh, yes, please just admit as people come in. Lots of people. Hello, welcome in. All right, let me see if I can do this again today.

Let's see.

All right, I'm gonna try it. Uh, I'm gonna wait because I did, uh, all sorts of stuff. So yesterday I realized I published our— this session is published on YouTube and Facebook and all the places. Um, and I realized Sky and I had a nice little conversation going before we turned on I went on YouTube, I'm like, I'm not sure I want everybody hearing that. So I edited it, uh, but it was pretty funny. It's not very engaging is really what it comes down to, is somebody sees that and they're like, uh, just some dark screens of random women talking. Uh, hello Frank, what's up?

Good to see you.

May is in the house, and Mary and Mark, Lexi, what's up? Natasha. All right, awesome. So I'm gonna ask, um, a few people who I saw Doing some good work. Great to be with all of you. Yes, I got some really exciting stuff. So we do have a little contingency of people from an organization called Mindvalley, the MV in the house. And I just got some notification about our AI Summit coming up here in March, and I'm super excited about it.

It's going—

I should— I'm not going to steal any of their thunder because they're going to be mad at me, but It's gonna be amazing and it's very much geared towards exactly, I mean, they asked me, they're like, I don't know, Noelle, can you talk about these things? I'm like, I'm doing a challenge right now on those things. Uh, so it'll be fun. Um, a little bit more advanced, I guess, not advanced, but quicker to develop cuz I only get 1 hour, right? In that whole summit I get an hour. So it will be like this 5 days condensed into an hour. Um, so, uh, stay tuned for that because it's free to join. Uh, about 150,000 people did it last year. Uh, and that's why so many of us know each other. So I'm really grateful for the opportunity, really excited for that, uh, coming up right after my birthday. So really, really cool. Um, and so March— February is like my favorite month, and March is kind of just as busy. So as we welcome people in, um, March, I end up uh, doing a few things. One, I'll be in New York City for my birthday. I'll be speaking to a company. I love it when companies, um, pay for me to go to my favorite city. I was born in New York City, uh, so pay for me to go to my favorite city right before my birthday because then they get to pay for my whole trip. Um, so that's really awesome. So thank you in advance to that client that's going to be doing that. And of course I'm going to see Hamilton. Y'all know how I feel about Hamilton, so I will be seeing Hamilton for the— it's really cool, it's my my birthday, 3/16, March 16th, and I will be going to watch Hamilton for my 16th time.

Why?

Like, I don't know if y'all believe in like power. Anyway, it's a good thing. It's cool. So I will be watching that again. You know, I don't have it yet. This is my Lessons Learned from Alexander Hamilton, but I actually have a signed movie poster or like playbill from with Lin-Manuel's signature. So I'm just such a big fan, so I'm excited for that. My next hope, and if any of y'all have pull, I'm saying this to the universe, is that someone, for the love of heaven, create Epic: The Musical as an actual musical. If y'all don't know it, like, go check it out. It's a Greek mythology based on The Odyssey. You remember Iliad? Anyway, I'm bringing back some 9th grade vibes. But anyway, that is my new dream, is I would love to see that on Broadway. But until then, Hamilton has has my, my number and I'm going. Uh, so that is exciting. But right after that is our AI Summit with Mindvalley. Then right after that, I will be headed to Microsoft, uh, for the very coveted experience of the MVP Summit. Never before have they actually limited the MVPs who could go. Like before, if you're an MVP, meaning a Most Valuable Professional, an award given by Microsoft to kind of top community contributors and developers, you never had to like, apply to go to the summit. Like that was your reward cuz they don't pay you to be an MVP, right? Like they, I get paid through them, like letting me go. And the benefit of the MVP summit is that you get to talk to all the product managers, you get to talk to the owners, uh, the, the marketing team. Like you get to see all the things that they're thinking about doing '26, '27. And a lot of the things I do, I get to seed and build right before it comes out. So the second it's available, I'm ready to go. It's like hugely valuable to my business. But yeah, about 6 months ago they're like, get on the waiting list. And I was like, waiting list? I don't do waiting lists. Like, it was very a humbling moment where I was like, what do you mean? Like, you have to— like, do you know who I am? Needless to say, I did get accepted like a few weeks ago during Christmas, and I was like, fuck. So I'm super excited. So that is the third week of March. So needless to say, as soon as I leave that event, as long as I'm not under NDA, meaning there's No secrets. I will be sharing with you all the cool new things. But what I love about that event is that I attend all the labs, so I get all of their live learning. Last year we did the same thing. April, May is just full of live learning opportunities where we just get on a call together. Granted, this is for our AI Leadership Institute members, but we get on a call together and we just build live. Like, because what I found out is when you're doing this work, which is what we're going to be doing today, when you're doing this work, it's all fun and games until you get stuck. And if you are not in a room where people can unstuck you— is that a thing? Anyway, if you can't get unstuck, you end up quitting. And that's like 90% of people, is that they start doing a thing and then they get stuck and then that's it. Um, so my hope is that you realize how important, number one, this room is, uh, full of people on the same journey. Um, and just know, you know, like On either side of you, uh, you will have people that are going to excel, uh, in what we're doing and move. And like, that's me. I'm ambitious. I will just push through by myself. But there'll be others that will desperately need people like us who are kind of pushing through things to go back and be like, oh my gosh, I just ran, I ran into that problem and here's how I fixed it. You can already see that inside the community with homework last night, the amount of feedback we gave each other. Um, I was watching a video, uh, by one one of my mentors. I've, you know, been in a bunch of his workshops and classes and, and met with his team. His name's Alex Hormozi. Some of you probably know him. He's really big in marketing. Uh, he owns a company called acquisition.com. Um, and one of his companies is basically is an AI company called turing.com, I think, based on Alan Turing. Anyway, a lot of people went into the business of AI about 4 years ago, so good for him. But I was listening to him earlier today and he was, just reiterating something that I always, um, am worried about, which is if you don't surround yourself with people that are just a little bit ahead of you, you will get stuck and you won't be able to get yourself out of it. And you'll ultimately not stay where you are, but you'll ultimately go backwards to a place of comfort. Um, and he read, uh, Alex Hormozi read one of my favorite things, which is the good old days of Jeff Bezos. This is all public domain, so you can go out and find it. But Jeff Bezos wrote a bunch of, uh, shareholder letters. So in— I don't know if any of you have stock with a company that sends you a prospectus, and in the front of the prospectus is a shareholder letter. But Jeff Bezos is like a poet. Like, he— I told— talked to you about this yesterday— he handcrafts every word. Um, and so his last shareholder letter is actually very, very good. And the reason I want to talk to you a little bit about it today, number one, because I just heard it and reminded me of it, But number 2, it actually is in the ethos of how I do things. And sometimes people don't know what I'm doing. Like they don't realize how I do, why I do the things I do, why I make the decisions I do. But he was talking about this tendency that we have as a species, and this is true for all mammal species, is that when we go into an environment, our temperature, uh, like literally body temperature, it, it, it is typically different than the environment we go into when we go into a new environment. Like me, in Minneapolis. I'm a Miami girl. I was born in New York, but I've lived in Miami my formative years and onward. I, my body is a different temperature, I think, than people that live in Chicago and Minneapolis and Iowa. Like, it's just, ah. However, there's two things that happen when you put yourself in a different environment, right? One is that the average, the typical, the thing your body wants to do the most is become the same temperature as you're outside. It wants to regulate itself. And what Jeff Bezos— his hypothesis was that the more you can stay in a dysregulated state, the more likely you are to think differently, act differently, be different. And that's why a lot of people, when they meet me, they, they kind of categorize me different. Like, they're like, you're eccentric, you're— someone said this to me last night, they're like, your approach is different. And that's how they said, your approach is very different. And I was like, I think that's a good thing. Um, and then I got— was reminded of this, uh, this article, this shareholder letter that Jeff Bezos wrote, and he basically was like, different is literally one of the main ingredients of success. And you could go back and look at all of them, you know, Michael Jordan, right? Uh, Jeff Bezos. I mean, you can judge success, right? Elon Musk. I'm— anyway, I'm not going to go any further down that dark whole of leaders that have, you know, defined success. But anyway, you know what I mean? Like there are people that have made all the money they wanted, achieved all the goals that they wanted, but what made them different is when you look at them, they're not the average. And what, what, uh, what I realized they were talking about was this term that many of us have come from corporate, what many of us have come to know as industry standards. I don't know if you've ever heard that term. Um, it is a place of comfort for many consultants where we're like, this is how the industry does it. This is the best practice. This is, you know, these industry standards are a way for us to normalize, right? To become the same temperature as everyone else. But what I love about Jeff Bezos, and now Alex having resurfaced it, was that he said, your job is not to meet the standard in the industry. Your job is to set the standard in your company. And that is what we're doing today. We are setting the standard. And your tendency, your desire, is going to be to do the bare minimum so that you could get it out the door so you can move on to the next stage. The only one who will set the standard of excellence that will distinguish you from everyone else is gonna be you. And you will need to set that standard incredibly high, meaning, I don't— if you wanna use the analogy, at a much higher temperature. So everyone else is gonna be cool and like, hey, this is kind of fun. You're gonna be like, this is crazy. What is happening? I feel like everything, like this is way harder than I thought it was gonna be. So what I'm presenting to you today is a technically, I guess you could perceive it as a harder way to do things. However, it is a distinguishing way to do things so that you are in the 5% of people that are successful when they deploy AI. Can you do it without this? Absolutely. There are, you go to school and type in AI, there are hundreds, maybe thousands of communities with AI enthusiasts in it. However, those that are actually being successful at a rate that I feel all of us can be, that is dependent on what we're covering today. So while I know many of you are like, let's just do it, I wanna build something, give me a tool, I will give you some tools in VIP. But today is where we take the ideas we had yesterday and we start to really think about them. Now notice the timeframe, right? This is, 2 hours, 1 hour to ideate on the thing that you wanna build, and an hour maybe to ideate on the details of that thing. Then we start building. You'd be surprised how many people just don't spend these 2 hours, and it's all a bit relative. So on a project, I would, let's say it's a 6-week project, I would spend the first 2 weeks ensuring that I have these 2 things locked in. Now, I was going to whiteboard something that some of you have seen before. But instead of doing that, I'm going to send you the whiteboard so that we could kind of talk through this and of course give you all some time to ask some questions. Now today, just like yesterday, I do have a— oh, I think I threw— I think I copied the URL. Let me go see. So I'm going to throw the URL into the chat. Yes, awesome. I love it. Okay. Awesome. Awesome. Okay. Control V. There we go. This is day 2. Those, that's the slides. It'll also be in the notes. I hope you saw, like, we worked a little bit hard to make sure that you got all the content in school. I personally emailed everyone who was not, I went through the school membership and if your name was not in there, I personally sent you an email and said, please join school. Um, so hopefully all y'all listening to my voice are in school. If not, it's, yeah. All right.

Good. Yes.

You were one of them. Noelle Russell. So I emailed every— there was only like 9 of you that hadn't joined at as of the time I emailed you. But then we got like a bunch of people that joined this morning, which is fine, and I tried to email them too. But it's in school, so please don't forget that is where we're storing it. As soon as I get the recording and the transcript, I throw those resources into the school. If you're part of our AI Leadership Institute, you also have a copy of this in the institute. Um, and it has a little bit more information, some of the links that were captured in the transcript. Sherilyn is doing that for you all. So thank you, Sherilyn. Um, uh, but everything that we're doing in the context of this—

what?

Lauren is in the house. Good to see you, Lauren. Um, every time I see your name, I'm like, old friends. Uh, so good to have you here. Okay, awesome. Um, yay. Uh, okay, good, good, good, good. Oh, good. Um, awesome. I'm so glad, Renee. Okay, good. So that means that hopefully we all have done day one work. That means, and if you haven't, you're gonna do it right now. You do have to pick problem. So I wanna just talk you through the successful process, and I'm not sharing my screen yet. But the successful process you wanna go through anytime you start an AI project, and the first thing you do before you start anything that involves AI, the first thing you do is you define the problem, right? The problem. You don't say, ooh, what cool model can I use? You don't say, oh my gosh. I mean, someone just yesterday, Jason Spaulding from our Leadership Institute, sent me a message and was like, Look at this cool new thing. And I couldn't help myself, like shiny button. It was like a moth to a flame. I couldn't help myself. I'll get time to go do that. But I did and it was cool and we're gonna do a live learning lab on it actually. Um, but it was a Claude, what do they call it now? Anyways, like, oh, a Claude plugin. And, uh, it was a Claude plugin to help you stay up to date. So it basically refreshes your AI knowledge every 30 days automatically. Like I was like, this is amazing, but not right now. So I have to co— like, wait, let's stay focused. But the first thing and the best way to stay focused is to clearly understand the problem you're solving. So that is why yesterday our goal was not only to of course define an agent, but actually understand what problem was that agent gonna solve. And it needs to be pretty narrowly focused. Now I say narrow, but like if you're talking about an executive assistant, which many of you in some way, shape, or form decided to go. Towards, um, like having to check email and manage calendar. Like those are all part of the problem it's solving, which is kind of managing like your, uh, inbox and your calendaring. Like that's all part of an executive assistant role. So I don't want you to get so focused that you're like, this agent can only do one thing and that's label my emails, right? Like there should be a collection of subtasks, if you will, associated with the thing you're asking, but understanding what its job is is critical. So that's why we spent all day focused on it. Many of you did this work. Um, I don't— uh, did anyone— is anyone feeling comfortable enough to share either in chat or, um, with me, uh, your, your experience identifying? Did you have any trouble? Was it, was it fun? Did you enjoy it? I'll look at your faces. Um, yeah, feel free to throw it in the chat if you don't want to share, or you can unmute yourself if you feel so, so obliged. I will, I will give you a couple seconds.

Oh, yay.

Oh, all right. Approve. I have some AI agents coming in. Welcome. Welcome, agents. Thank you for being here. We love robots.

All right.

No, no brave souls. All right. So, oh, yeah, there we go. All right, Alonzo, let's go. Thank you. I just need a couple.

Thank you.

Okay. So, yeah, I'd like to do the executive assistant item that you mentioned, and I figured that I'm just going to use it to take an email that I normally use for research just to keep up to date with AI stuff. And the idea why I'm keeping it simple is because I think you said start simple and then it could eventually do recommendations and other stuff. I just want to get that part right. Like it could do, you know, and your high, medium, and low, I like that scale. So I said, you know, tell me if this is a high value email, low value email, or medium. And making it simple for the metrics. Like, does it talk about revenue, income, some kind of way for it to figure out what is high value and things? So simple. And then later on we'll grow into it with iterations.

I love it. That is great. Yes, I— it was my first thing also. I was like, how do I just get my— the day— the way I start my day typically is to go to my inbox. And now I am a big fan. I have a book here, uh, David Allen, Getting Things Done, right? GTD, right? Yeah, I see some friends who share this. Um, and it's the philosophy of getting things done is don't go to your inbox. Like, it's the last place you should start because it's like a hole you fall into and 2 hours later you haven't actually done anything valuable for your business. Um, and so this is why I was focused on— I, I actually can't focus on something else if I know that there might be a contract pending or someone's asking me for help. So I now feel better knowing that my AI system is go— is going to bubble up things to your point on like, This needs your attention. This is urgent. Um, there I, I, am I getting ahead of myself? I don't think so. But I, when I build this agent, one of the first things I asked it to do once I trained it was to then give me a daily briefing, right? Like I want every day at 8 o'clock for you to give me what's on my calendar. Am I triple booked for something? What are the top 3 emails that need my attention, right? That, that takes me 30, 40 seconds to read as opposed to what I was doing before. So Thank you so much, Alonzo, for that. Christy, would you mind sharing with us your experience? And then we'll go to Art.

Sure.

So my, I was like, well, what do I want? I want AI to help me with my email, right? And then you have to say, well, what does that mean?

Help me.

Yes. So that was like getting to the help me, help me, I'm drowning. Yeah, yeah. But what does that mean? Like, well, tell me what's important.

Well, what, how, what criteria would I use?

Yes. And all of these, uh, fun was fun. And I'm not to a tight definition yet, but I'm, but I'm on my way.

Yes.

And the fact that you just recognize that it's not as simple as like, uh, I need us, you know, I want an assistant, right? Like now you're like, okay, well, and it's just like a human. I've tried to hire humans to support me in this and they fail, not because they failed, but because I can't clearly articulate how I want them to help me. And as a founder, I think this is especially, you know, founders, business owners. I think it's a huge challenge, this delegation gap. Like, how do I describe to you what in my mind I can do, like, without thinking? Yeah, yeah. So yeah, that articulation is really, really interesting. So thank you for sharing. Um, hey, Arti, how are you? Good morning, afternoon.

So, Christy, I'm with you. I was like, I did this asynchronously kind of jotting down what I wanted. And then I actually went back to ChatGPT and said, listen, you are a special— you are a specialized job description creator. Look at what I've put together and tell me where are the gaps, like what is not clear and what needs a little bit more detail. Ask me questions. I felt like just kind of changing the role helped me think better. So rather than asking GPT for questions, uh, for answers, I asked it to ask me questions. So I think that only helped me kind further crystallized your point, Christy, to really kind of get clear because I was like, oh, help me with my social media post. Like, what does that even mean?

Yes, I love that. Awesome, awesome. Wonderful. Okay, 2 more. We got Latasha and then you, Dave. All right, now, Latasha, how was your experience? Did you enjoy it? Yes, I absolutely enjoyed it. And here, here's what I realized. I'm trying to grow a business and I need help onboarding. Like when you talked about that yesterday, that really resonated with me and I was like, yes. This is what I need to focus on. Yes, that is. And I mean, me too. So all y'all, like, I'm doing the best I can. I'm just kidding. I do not have an onboarding agent. It is just me and my team. Like, you know, it's funny, I, I built this and I like to do it. I build my plane while flying it. I'm going to encourage you to do the same thing. You will attract people that like just appreciate the content in your brain you're willing to share with them., and teach them and help them with. Do I have a perfect onboarding process or any onboarding? Maybe I, I definitely have a mural with a couple boxes connected on my journey. So, uh, but I'm, I'm hopefully here among many to show you like the most important thing is to start and to continue. Um, because if you wait till it's perfect, if you wait until you've figured out everything, you'll, you'll have missed the opportunity during the time when you could have actually been getting really valuable feedback from your audience. So I am super excited that you were thinking about onboarding earlier. So that is amazing. And I can't wait to learn how you do that. All right, Dave, what about you?

Hi. Hi, Noelle. I have all these Python scripts to actually test our APIs. And when I was listening to you yesterday, I was like, oh, wait a second. If I had an agent that could actually just tell me what's going on, test it in the background, and then send me a summary going, oh hey, there's an issue here, or everything's good. That would be awesome. And so that's what I was thinking. I was like, oh, okay, I got this Python script and it's like, okay, how do I automate this so I have an agent that's actually gonna let me know as soon as something happens and can also run tests so that I'm not going, oh, okay.

Yes, exactly. And give you like a conversational interface where you could just be like, just tell me the time, like, Or even alerting, right? Like, tell me when something goes wrong. I don't need to see it all if it's working. Yeah, huge amount of time saved. Huge amount. Okay, so that is amazing. Kristi, did you want to add something else? And then Krishna, and then that's it. And then we're going to go on to our content. Oh, you just raised your hand. Oh, you had to put it down. It was still up from before. Sorry, I just want to make sure. Krishna, good to see you, my friend. Congratulations on your certification. All right.

Yes.

How was the experience for you?

So I'm, I don't have one picked out yet, I'll be honest, but it's not because I haven't thought. I'm trying to make it very meaningful to my current role. And nothing that I do here is, in my current role, is repetitive or rule-based. We're trying to go in simple for everybody's understanding. We're trying to take a company that has been entrenched in project-based approach for every project order we receive to becoming a product-based. So a lot has to do with how I'm, along with a couple others in the, in the entire enterprise, distilling that information and how we're steering the next steps. So my first thought went to, could I have another person like me who could be helping talk to the various people that I will be working in the next day, week, or next few weeks, where they could interact with this agent and kind of get what I would say in terms of understanding what does it mean for my role when you say product-based Krishna.

Yeah, yeah, exactly.

Stuff like that. And I don't know how to define it. So maybe if you can summarize that problem for me, I can dive in.

Absolutely. And this is actually a really great opportunity to talk about a prompting technique that I personally use. And again, it's one of those things that I did so inherently that I didn't realize everyone didn't do it. And you'll find this with the way you do things too. But I mentioned it briefly yesterday called Socratic prompting. I don't know how many of you do this, but even in our own personal lives, right? Like taking a Socratic approach. I have lots of books on Socratic journaling and it's how do you switch from asking an AI system to do something for you to asking it how it would solve the problem you're asking it for. And it sounds a little confusing the way I just said it because I'm not a Socrates. But this concept of asking questions versus, um, and I mean asking questions to derive an answer as opposed to asking for it to give you something like a vending machine. Most people use AI like a vending machine. Give me, uh, an email. I mean, I use it sometimes this way. Write me a, uh, thank you letter. Give me, um, a schedule. Uh, take a picture of my pantry and give me a menu. Like that's the vending machine method of using AI. What we wanna do from now on is approach it from a more Socratic position, which means— and I'll just give you some examples. I'm going to throw them into the chat of ways that I have adjusted my prompting when I'm deciding on how to grow my business or come up with different business values or OKRs or how I'm going to measure success. So I'm going to copy this one just to give you an example. Of what it looks like, but I'll read it to you. This is my prompt, right? What makes a value proposition compelling to a B2B buyer, right? Like, what is it about? And I, I'm not asking it for a specific value. I'm like, in general, what are the characteristics of a value proposition? Because I said that and I thought I knew what it meant, but I'm not sure my customers know what I mean when I say that. And what they say oft— what, what is, um, a typical marketing challenge for many of us is we speak from the side of we've got the solution, and most of the people that are trying to find the solution, they don't know the words we're using. And so, oh, let me hit enter. So this is an example of, what some Socratic prompts that I've used. I just pulled these out of my chat library so you can kind of see what they look like. Like, what types of LinkedIn content generate the most engagement when it comes to B2B SaaS? Now, I'm not a B2B SaaS person. That's a prompt I created for a customer. What posting Frequenting avoids audience fatigue. Like the model knows this because they watch it happen, right? How should, like, instead of, most of us don't ask questions like this, right? We're not asking it to look at all the data it has and build us an answer that's contextualized to the data that it has at its disposal. Most of us are like, give me this, and we make assumptions about that answer. And I just say like, switch your questions to challenge every assumption that you're making because now everything you do can be data-driven. So this means that when you're thinking about this, Krishna, and this is what I do, is I literally will, and sometimes I might even need help, like, how could I pose, right, a Socratic question about this topic? But I just have a conversation with, now granted, I have 5 years, I have 1,583 days on OpenAI right now, right? So I have like 5 years of prompting experience with this model. So it knows so much more about me. Than if you just started by prompting with it this year or even last year. Like I've been using it and I've been using it to run the same business. So it is very good. So when I ask it a Socratic question, it has a lot of business context. So I, I would encourage you to kind of think through what kind of questions you might ask it that remove the assumption that you know the direction the answer should go in. We call it in, in legal terms, leading the witness. Many times in our questions to ChatGPT, we're leading ChatGPT to an outcome. Like, we're like, build me a, you know, social media that does this. When instead I should be like, here's my customer. What social media would I create if I wanted that customer to get engaged? Right? Like, why make the assumption when you can ask the model to, to look at its data and give you that benefit? Especially now when these can be deep research, deep thought, depending on what model you're using. So this is called— the other thing Excuse me. I'll just put in the terms. You can Google it yourself, or ChatGPT it, or Perplexity it. It's Socratic prompting. It's a whole prompting philosophy. It was written about at MIT. And yeah, exactly, exactly, exactly. It is very much a legalistic approach, but the benefit is you get so much better data extraction when— in, like, content resolution. When you do it this way. Yes, exactly. And I think AI red teaming is very much based on this method as well. So I think it'll change the way that you approach asking questions because so many of us have gotten in the habit of— I mean, Google trained us all really, really well, right? Just ask for what you want. That is the opposite of a Socratic request, is like, give me a link for this, or tell me about, you know, like, The Muppet Show. Yay, thank goodness it's back. I'm so excited. So I'll be like, tell— give me an update on The Muppet Show, or who created The Muppet Show. Like, I'm asking it a question and it's vending machine style giving me that answer back. This method that I'm sharing with you is exactly what we're going to leverage over the next— the rest of this session, because we don't want to make assumptions. We don't want to— we want to leverage the full capability of these systems, and they're going to get better and better and better over time. Okay, so thank you, Krishna, for that question. Hopefully I answered a little bit. If not, let's Talk again in VIP. Okay. Gabriella, do you— would you like to share this last, last one? It's the last, last one.

Yeah.

Just for anyone here in the group today, has anyone kind of thought about what kind of AI agents that will be helping with client context in particular, like private sensitive or cost sensitive firms where they're can't share their kind of protected data. So you can't exactly have much maneuverability to, you know, go test your stuff or as well as the norm, just using stuff for emails or whatever.

Yeah.

So that's just—

yeah. Number one, if anyone in the chat wants to provide a chat-based answer, you're welcome to do it. But I will give you the distinguishing piece of data that is often missing from most approaches, which is the difference between a public model and a private model, right? When you own the model, you own the moat, you own the walled garden around your data. So what you consider PII, it's only— PII is only concerning when you are crossing the boundaries of a private system into public domain. When it's inside of your private walled garden, what we call a cloud tenant, then you don't have to worry. It's your data. Like, it's— that's what that, that grounding is for. Ways to resolve that is— I don't wanna get too technical, but like creating a DNS, right? Uh, creating a privatized network and closing it off. This is what people do even with LLMs. You can close an LLM off from even talking to the internet. So a lot of, uh, financial organizations known, uh, I mean, hacked I guess if you consider breaking, um, you know, 64-bit encryption hacking, maybe. Usually when we say hacked, um, hacking is often exposing vulnerabilities that you could have closed, right? So oftentimes, like cybersecurity attacks, a lot of the things that are really dangerous are, are often because the person either didn't lock the shed at all or they put a little lock on it. Have you ever seen that with like a locker? You're too lazy to like remember the combo, so you just turn it so it looks locked. Like from far you're like, yeah, yeah, it looks locked. But like if you actually turned it, you're like, wait, nobody actually locked this thing. So many of the, the challenges that we have in cybersecurity, especially for you all small businesses, is usually because you're not just doing the needful things to protect your own system. But the question you're asking, Gabriella, specifically around PII is often Worry— you're worried about it because you're using a model that you don't own with data that you own, and you're worried about how vulnerable that makes you. And just to be clear, it makes you very vulnerable. So the first thing I tell companies is that you need to create your own privatized model. Oh my gosh, I do teach it, Jill, just not in here, not for our one, you know, hour. But yes, and we will do it actually more importantly in, in the Golden AI Academy, we will definitely be doing that. And that it, there's a lot of different ways to do it. Like I'm kind of speaking of the more, most complicated, which is create a cloud tenant. You own it, it's expensive, but you get to encrypt all your data. You get to encrypt it at rest and in transit. You get a lot of control, but you can also just indemnify, right? You can get a business relationship with a, with a, uh, like an OpenAI, not necessarily ChatGPT, but you can create a relationship with companies and they will protect your data for you. So there's lots of different, uh, variances to that. Lots of cool— oh yeah, Socratic. Okay, cool. Um, oh yes, Nazar, I know it's a— so the Golden— I shouldn't have mentioned the name. Golden AI Academy is a special academy where as many of the members here, kind of like Mindvalley was his own little community. Um, Golden AI is for event professionals, so it's narrowly focused on people who are running the events, by the way, that all of us attend. Um, so If you, you know, it's a really cool opportunity for us to help business owners that are growing our event space. And I don't know if you know this, but post-COVID, we're actually seeing more interest than pre-COVID on getting together. Humans, I mean, look at us, like humans wanna get together, they wanna talk to each other, they wanna learn from each other. And so we actually desperately need to help event professionals build good events, use technology to scale their events to more people. And so that's what that community is for. So thank you for asking. Yeah, and it just goes to show like every little niche, right? Trucking. Where was I the other day? Oh, Mississippi Chamber of Commerce. Like all of these very niche areas geographically, domain-wise, they all need help with AI. So if you are in that world where you're like, what should I do? Learn what we're doing here and then go help people. So many people need help. There is no scarcity of clients or customers. You, we, the, you just have to have the courage to get out there and help someone. Um, okay. Uh, awesome. Okay, so let's, let me go into the deck. Uh, I gave you, I'm gonna see if it's still in my, oh no, I, I copied the Socratic prompting, so hopefully someone grabbed the link and can throw it back in the chat. Um, if you can, that would be awesome. If not, it will be in the notes and I will grab it at the end.

Share.

So this, I'm going to remove my arc. I don't want too much distraction. But today we're going from problem determination to the next stage, which is at a highest level. I call it kind of the data stage. Like, how does the model know what good looks like? How does the model know what an accurate answer is? What's the desirable outcome we want? What's an undesirable outcome? So this is where we're training the model. So it will build on what you did yesterday. So if you haven't had a chance to do that work, you will need to in order to continue. And I will tell you, ChatGPT is your friend. If you feel like you can't figure out what to start with, it means your scope is too big. Shrink it down to something much smaller just to do the exercises, because what I'm teaching you this week, you're going to be able to repeat for everything like it is a completely reproducible process. Um, and my goal is of course that you would do that and add complexity over time. But if you start with a, usually if there's some, uh, delay of game and getting started, it's because you're picking something that's just big and you have to think bigger. So think smaller, think minimum remarkable products, tiny little thing that you could do where it would be like, wow, that thing Noelle Russell said, that's, that's working. Now let me go bigger. Let me add more to it. So today we're gonna be talking about how do we transfer Remember, we just started with this at the beginning of the session. What is in our brain? How do we transfer what's in our brain into the system that we're building? And this is the hard part. It's hard for humans talk, you know, training humans, and it's also hard for humans training machines. But the problem is pretty much the same. AI implementation, right? It's not about building an agent. Anyone can build an agent. An agent is just an AI system that does something. And it does something typically through the use of another technology by calling on an application. So imagine if you're building an executive assistant, that executive assistant is only as good as your connection to the things that that agent needs, right? Like it could be brilliant, but if it doesn't have access to your email, it's not gonna be super functional. If it doesn't have access to your calendar, it's not gonna be able to help you with calendaring challenges. So this is not just about the tech, it's about the connection to the resources, the data that it needs in order to do a good job, and the definition of good that we define for it. So what I like about this is, again, depending on where you are, if you are an ambitious professional, it's taking your ideas of, ooh, I could use ChatGPT for that, to how do I build a sustainable operating model? It's a completely different shift in your mindset from how, how you're using AI. Most people are like, hey Chat, help me with this email. Hey Chat, here's a transcript. Tell me details about this. Hey, you know, like they're asking these kind of one-sided vending machine type questions. When if we shift our mindset to an operating model, now we're gonna talk to it more like, your job is to make sure that I don't triple book myself. Your job is to make sure every contract that gets, comes into my inbox gets answered with a 8-hour SLA or 2-hour SLA. That means that it's gonna now be looking for that content and ensure that I get that information within the boundaries that I've defined. It's a totally different behavior, right? And the only thing that's different between the way you're doing it now and the way that I'm talking about it is just a choice, is the words you use and the amount of work you do to prepare the agent to do those tasks. That's what I love about it, is that the technology doesn't actually change. It's us as leaders who have to change. We have to evolve the way we're using this technology. Now, when I was at Alexa, one of my biggest frustrations is that we built a technology, um, that could do anything. If you talk to it, it could do whatever you asked it for. It could look up sales data, it could, uh, read you analyst reports, it could, um, read you an Audible book. Like it could do anything. When we went out of the mar— like into the market, our go-to-market strategy was to teach people to build little skills. For Alexa. And our hope was if we teach you to build a little skill, you'll go big, build big skills. Turns out I taught tens of thousands of people to build little skills. We called them fact skills. As a matter of fact, the article I wrote was so popular, it is referenced on Wiki I mentioned yesterday, right? So here's how to build a fact skill, which means can I train Alexa, uh, to build an application that literally will just regurgitate facts from the internet. Not very fancy, but it teaches you the mechanism, right? The mechanics of how speech recognition works, natural language understanding works, and calling code works. It was very, very simple. I don't know if you know this, but like 90% of the skills on Alexa are fact skills, which means no one ever went back and built anything else. They kind of were like, I built a banana facts Fact skill. Maybe you wanna know banana facts. I don't know, like cat facts, dog facts, meditation facts, um, mountain climbing facts. Like they named it a little different, but they were all just going into a database of data and regurgitating something back. Super simple, barely AI. The only AI was in hearing a human say the words and translating it into a machine call. That's it. That was not fancy. But our, our assumption was that you would build more. That you would have the ambition as a developer to go build better stuff. That didn't happen. So that's why I'm so focused today on how do we change our mindset to build, like, you're gonna build something simple today, but your mind should be set on the things 19 use cases from now. Like, what does this little thing allow you to do at scale? And you don't just think about the little thing, the little fact skill. You think about, if I enable this, that allows me to completely retransform my business in this way. Be thinking big, but building small. This is something I learned from Jeff Bezos, and it's how he does everything. He thinks big. We're gonna take over healthcare. He starts small. We're gonna start with prescriptions, right? Like, how do you get into the habit of reframing your mind in that way? Now, I do wanna give you a warning. There are a bunch of reasons why AI projects fail. I talked about this yesterday. I did feel bad that I didn't have the data to support it, so I wanted to show you what that looks like. This is that article. You can see the article reference at the bottom. We'll also include that in the notes today. But this is that article I was mentioning yesterday. It is a Forbes article that actually combines a couple of different academic, uh, papers and research studies that have happened in the last 3 years to quantify, which is what we, we all want, how is generative AI actually changing, like moving the needle, and to to some, some people's surprise, not mine, um, to some people's surprise, they're like, oh my gosh, almost everyone is failing. But I want you to look at the reasons why people are failing. Why are these AI projects missing the mark? And notice the one that has the biggest impact, an unwillingness to adopt new tools. Can any of you imagine you go into a company and that company's like, hey, we build this cool AI for you to use. Why might I be unwilling to use an AI system given to me by my company? I mean, you can ask the 16,000 people that got replaced in a certain company and the excuse being, we have AI for that, right? You could ask other, I mean, these are big tech companies that have, and I mean, there's a, there's a company, Salesforce, they actively say in the hiring of humans, you must prove that an AI system can't do this. Before you hire a human. Well, if you're already employee, right? You're an employee doing work and someone gives you an AI system.

I don't know.

I personally might be like, well, what's that gonna do? Like, right? There might be this mass level of friction when people realize that some leaders are choosing AI over humanity and that that humanity might be them and that their job might be at risk. So there's fear, concern, job security, a lot of reasons why new human, like humans in a company might not even use the AI that you're building. But the one way to get around this and what the 5% know is that the way that you get, rather than building something and lobbying it over to a human staff and be like, here's some AI, I hope you use it well. Instead, we do something I learned from the accessibility world. We never build for people without them, right? If I'm building a solution for my chief of staff, I'm going to ask my chief of staff what hurts them first. If I'm building for my customers, I'm gonna go to Reddit and find out what they're complaining about, right? Or I'm gonna go to, you know, um, I don't know, Substack. I'm gonna go online, I'm gonna Google, I'm gonna go to Answer the Public. I'm gonna find out what the reason, what, what are their pain points, and I'm gonna build with them in mind. And even better, I'm gonna identify some of them to build with me so that we are gonna not be able to, of course, go channel who we're serving. I mean, that's why I said build an AI employee for ourselves, cuz we're, we're serving us. But the macro level of the solution we're building this week is actually not only how do I build solutions that help me individually as a business owner or as a founder or as a chief architect in a company, But also, how do I leverage this to equip my staff? And this is what Amazon does really well. They have a, uh, a leadership principle called invent and simplify, which means as an employee, your job is to constantly look at how you do your work and invent and simplify it. Meaning build macros in Excel, create, right? Create little agents, figure out a way to do your job well, so well in fact that there's actually a set of awards they give at Amazon. That one's called the Just Do It Award, which is given to someone who just built something to help themselves and then shared it with people and their scale, right? The scale of that assistance actually impacted the business in a positive way. There's the, um, what is it? And Just Do It, by the way, you get a sneaker. Like, I, I, I don't think it's used, but still, like, it's a sneaker just bought off the shelf and you get to have it on your desk and you're very proud of it, but it's just a sneaker. It's not even put on a plaque or anything, right? But these awards are given for people who innovate on behalf of their customer, who take action, and usually they're taking action from their personal lived experience. And other people in the company are like, oh my gosh, I have the same problem. I, that's amazing. That's not typically how we build stuff. If you are an engineer, oftentimes we build in engineering and we lob it over to the business and we're like, isn't that awesome? Don't you love that? I, I hope you enjoy it. Don't call me. It worked on my machine, right? Like we kind of like dis, you know, distance ourselves once it leaves our hands. And that is the opposite, uh, of what the 5% are doing successfully. Now, this is the, the mental shift. I've talked about it already, but this is the mental shift that we wanna make as we move into this stage and the homework that we're gonna do today. It's a little bit more writing, and again, do it in a notebook or in Notepad. These eventually will become the guardrails, the instructions that we're giving to the model. But I need you to spend, I'm going to say, a good 30 minutes drilling into this. We're going to do this together if you're in VIP. You want to shift from, hey, write me an email about this, right? Or answer this email, or create me a deck, or build me up, you know, an Excel spreadsheet, create a, you know, whatever, right? Go from asking for something to operational thinking, right? Go from prompting to operationalizing, which means now there's way more detail. This is the Socratic method, right? When pricing exceeds $10K, I want you to escalate to sales. If it's an FAQ level, do this, right? We're gonna create, and this is something that I got, uh, a lot of, I don't know, I guess challenge for in the early days of Alexa, people would be like, it's not AI. Because you have decision trees, you have if, if this, then that in your code. However, as you build a, a generative model, right? It's very helpful to create these decisions and map those decisions so your model at least knows the general direction to go based on those examples. So that's what I do is I shift to operational thinking of like, how do I protect the model from itself, but also how do I protect the user, me or my employee or my client, by asking and rounding out the instruction set to not just be, write me an email. Sometimes it gets fancy, right? Some of us have been like, write me a highly engaging email. Um, write me a highly engaging email for LinkedIn. Like, it's still a prompt where you're asking a vending machine question. What we wanna do is flip the script on that and start asking questions that actually demonstrate operational knowledge of what this thing is meant to do, right? What is its job? If its job is to support sales, then we wanna equip it to know what those— remember yesterday we talked about thresholds. I didn't have you, I didn't have you define those thresholds, but you will today, right? So as we lead, as we begin leading these AI systems, we need to shift from consumer mode into producer mode. And this is an analogy that I use because I use it with my kids. Like, we have a rule in our house in order— before they're allowed to consume social media, they have to produce before they consume. It's a great rule. Um, they don't have to produce social media. They can do canvas art, they can build a bracelet, they can do anything. Um, they can sell Girl Scout cookies. We're selling, by the way. Um, uh, and any of those things count. But they— it's so interesting what has happened over the years is that as they become producers before they consume, Guess what happens when they start consuming technology and AI and social media? They're actually looking through the lens of their production. So they're now looking for better ways to build bracelets. They're looking for, you know, like it's very, very cool. And that's the same thing that happens operationally in your business. When you shift your AI systems to be operational in nature, they actually start improving on themselves. Self-learning is what we call it in machine learning. Oopsie, someone wants me to approve, so There you go. Welcome, Firefly agent. All right, so that's the first, uh, kind of mindset shift. Now these are the 4 categories of explanation I provide to a model before I even ask it to solve a problem. The first thing is context. Some of you know, I mean, it was very hard for me. I'm a ChatGPT girl. Like, I grew up on ChatGPT since COVID So a lot of my habits go right to ChatGPT. Judge me if you will, it's like an old friend, but no, remember zero trust contractor, but it's like a habit, right? So I have a habit of using that tool, but then I found out that, you know, Anthropic developed a model that has a million token context window and I was like, that changes everything, right? So now of course I'm like, zoop, for certain things that I need more context, in my prompt, I will go now and take advantage of that. Now, context includes all of the data sources. So this again is a thought process. What I often ask, I've now got a problem before I'm even thinking about AI. In order to solve this problem, not all problems, for this one problem, what are the data sources? Some of you picked a nice tightly scoped problem, email and calendar.

Beautiful.

So where does it need to go? Your inclination might be like, I have 6 Gmail accounts and I have 12 calendar accounts and I've got Calendly and this. I'm going to encourage you to pick one email account and one calendar to start. You can add more later, but to start. But that means I now understand that. And now I need to say to the model, based on this data that I'm choosing for this AI system, what are the constraints on this data? Are there any? And yes, there should be. What is its operating environment? Are there, for example, with my calendar, if it can't extract from parameters I've, or, or, details, data that I've provided, right? If I have not provided information inherently in the calendar around like the days of— like right now in my calendar, I've actually got time allotted for most of the hours of my day in like free boxes so that my assistant can see, oh, this is your podcast block, this is your sales block. So it knows where to stick things. But that is me applying those constraints. That's me educating the model on here's where stuff goes when you actually have to use it. And I can say, look for the free time on my calendar to get an understanding of where to put sales calls versus where to put learning and focus time. These are all things you need to think about. How are you going to equip your model to make good choices, right? To make good decisions. It is all about the context, all about the data you have, the constraints you give it, and the operating environment in which it sits. And the more simple your initial use case the better, because we want to exercise the muscle memory of building the entire agent end to end. If you build a problem that's too big, you're gonna get stuck here. And you're gonna end up with a lot of context that you're gonna be managing. I would try to keep it small in order to just build the entire thing out and then go back and add more. You can always add more. I almost add more instantaneously. As soon as it works, I'm like back in the loop, and I try— I start over. I add more context. I then do the next piece. So right after I get the data I need to make a good decision, you'll notice on the slide it says standards. This is my description of— for those of you Scrum people or Agile people, right? What is your definition of done? What is your definition of success? These are words that tell the model, this is how you know you do— did a good job. This is how you know you did a bad job. Now, many times the model doesn't know. Now, you cannot rely on reinforcement learning inside the model itself because you don't own that model, right? So we need to do this declaratively, which means with our words. I need to provide it a signal so that it knows what that looks like. What I typically do in my prompts is I will create a section called, this is what desirable conversations look like, and I will give 5 to 7 examples of good interactions based on the constraints. Then I will also give 5 to 7 examples of bad conversations, right? So if my model is helping with email, I don't want, like, if somebody, for example, I had to create this recently, it was creating draft responses to newsletters and I was like, yeah, so I appreciate it and that's really nice and maybe the person who wrote that newsletter might even write me back, but I don't want it for newsletters. And it's because newsletters have gotten really crafty so they don't look like newsletters anymore, right? The, the headlines look like emails intentionally so that I open it. So I get it, I get what they're doing. So I had to give it a little bit more education on what a newsletter looks like, and that's what I did. I gave it 4 or 5 examples from my inbox of newsletters that look like emails but actually are newsletters. Now it has to do a little bit more work to look inside the content to see, is it a newsletter? Is it an actual email? Does it require a response? So these— it didn't do a good job at first, but I over time started adding context to the standards of what looks good. This standards piece is often a part that most people either forget, grow tired of doing, um, in an effort to just get it out the door. And this is what will come back to haunt you because the model won't know that it's doing something wrong and you won't have any mechanism of providing feedback to that model. And so it will not just do it wrong, but actually accelerate the inaccuracies that it's generating. So this is where we'll put our limits, our thresholds. I usually like to also separate out decision logic. So this is like once— let's say, for example, I'm asking a model to triage my inbox. This is my decision logic that helps it identify. Now, a lot of— some models do this automatically. Meaning they already know how to manage a typical inbox. So don't overthink this. But if I have any special logic around managing my inbox, for example, my contracts, my contracts are called partnership agreements. So if you're, you know, if I'm asking it to manage my inbox and look for contracts and it sees partnership agreements in it, in the initial stages, it did not pick those. And I was like, oh, well, this is how I describe things. These are the definitions. These are the thresholds in which I want you to operate to ensure that it grabs the right information. So this is what I go through step by step. First, I define the data sources, which is my context. Then I define the standards, which means what is my definition of done? What is my definition of success? What does good look like? Then I provide decision rules, typically if-then statements, right? If it's this, then do this. You don't have to do 100,000, which is what we used to have to do in the olden days of like unsupervised learning or supervised learning. Like, we don't have to describe or provide tens of thousands of examples. We just need— for me, I typically like to use 3 to 5. Microsoft feels like 21 is a good number. There is no— like, a model doesn't need 21.

It—

I do think it probably needs at least 3, maybe. But somewhere between 3 and 5 and 21, create some examples that give it context so that it has some room to consider what good looks like and, and what good might look like, because the words might change but the essence is the same. So the more context we can give it on that, the better. I use if-then statements. If it's this, then do this. If it's this, then do this. More importantly, if it's this, don't do this. If it's this, don't do that. So I do both desirable and undesirable. And then of course, we mentioned this yesterday. I made you identify this yesterday, but today we're going to actually write out the logic for escalation. Who's it gonna go to? How does it do that? For me, again, with my inbox manager, like its job is to, it's gonna label it in a different way and it's going to send me a message in my con— my conversational app that hosts my agents. So it's gonna ping me and be like, by the way, you have, you know, and I actually will tell it, uh, at 8:00 AM, at noon, and at 5:00 PM, I want you to surface any urgent emails to me.

Me.

So it's automatic, right? I'm, I'm asking it to reach out to me, not the other way around. Push instead of pull. But that's my decision. I chose to do that. You could instead be like, keep it where it is. I'll look at it twice a day if you, if you want. Okay? So that is the way that you're going to transfer what you are, the assumptions you're already making about what a good job looks like into roles. Now, this typically would go into what we call a standard operating procedure if you were going to give this to a human. Right? However, if you're not going to give it to a human, we're going to build what's called a system prompt. Many of you know what that is, hopefully. But we are going to define that system prompt, and we're going to use way more characters, hopefully, than you've ever used in the past. But more importantly, they're going to be structured. And that structure can now be injected into any agentic platform. So you're going to build this, and you can use it to build a Claude agent. You can use it to build a, an agent, a custom GPT. You can use it to build a— what we're gonna do, we're gonna use a pla— an agent platform. You could do it in GoHighLevel, which is a CRM. You could build an agent there. This standard becomes copy and pastable as injected system prompting for across multiple platforms. And that's my goal, is that I don't want you to get stuck vendor locked in to these systems. Now, once we've got that, that is our rule set. The last thing I want to mention to you, and I'm a little bit over, so, um, because I wanted to hear what y'all were doing, um, but the last thing is to actually codify judgment. In other words, how do you make sure you know that the model is working well? And these are the criterias that I have set. Now, this is the homework: what 5 major decisions do you want this agent to do, right? What is the decision criteria, similar to what we just mentioned? What are the escalation triggers? Not what does— what happens once it's triggered. What is the thing that will be experienced in the model that will trigger an escalation? You want to describe that so that the model knows what to look for. We are leveraging the benefit of a generative model to give it instructions that then it's going to look for things that are similar and slightly different than what we're providing, but allows us to accommodate a wide range of things that fall into that category, right? Oh, I think I was going to put— I think I stopped writing. I was going to give you some don'ts, but I guess I'll wait.

Duh.

So I guess I saved myself 3 minutes of whatever I was going to talk about right there. But these are the 4 things that I ensure that— and this is my checkpoint. How do I make sure that I have embedded into the system prompt all the information necessary to make good judgments? This is what most people don't do. They do not spend time doing this. And the way I usually test it is, you know, I go and I ask questions that are outside the scope of what the model's supposed to do. And many, many times it's a parlor trick actually in, in a lot of executive briefings in companies where I'll go in and ask it to give me information in SharePoint that the model shouldn't have access to. And it will because they never built the boundaries. They never created the moat. They never identified the decisions the model should be making. And tell the model not to answer anything else. Right now, I mentioned this yesterday, but I do want you— you'll notice in the instruction set you do need to identify this as a lever. So you're going to choose— I'm going to encourage all of you to choose number 1, but these are the 3 different opportunities. It's not prevalent now, but there's a concept called AI assurance. This is a way that right now, um, regulated industries, healthcare, finance, they right now are looking for ways to document AI in their organizations, all the AI systems. And one of the things they're looking for is what is the level of human involvement for each AI system? And it'll be different for each one. So you want to at least have labeled, there is going to be AI drafts, human in the loop, right? Human has to approve everything. As it gets good at that, and you'll identify the determination of confidence. As it gets good at that, you then are going to shift it to, okay, human on the loop. I'm gonna now observe and monitor and look for errors and retrain as necessary. And then eventually, like autonomous vehicles, you will get to the point where you've identified the rules and you will embed them as boundaries for the model because you have done this so long now. I used to say 3 years back in the olden days when we were building traditional models, it took about 3 years to get a model to be actively capable of adhering. Remember I talked about prompt adherence, adhering to boundaries. Now it's happening faster because these models are more capable, but you still need to budget for time to climb this escalation ladder. Like you can't go from ideation to autonomous execution without some significant risk to your business. So, as we mentioned here, this is not a 1-week transformation. This is a maturity model. And the best way to test it is to put in front of— put it in front of your users, right? Whether it's you, whether it's your clients, whether it's your employees, get a couple sample ones, let them test it, and they'll let you know, like, wow, this is not ready for prime time. And the more users you feel comfortable sharing it with, the more likely you'll get feedback that's useful. So just really quick reality check when you're building. So one of the things I like to do is build in bulk. I'm not encouraging encouraging you to do this. I built this challenge to be Build One Agent for a reason, because as soon as you know how to do it, you can build as many as you want simultaneously. I often build in threes or sixes. Um, and so what you have to make decisions about though, if you're going to build more than one agent, is do they share the same standards? Are they going to be governed and documented the same way? Are you going to monitor them with one team, or are you going to separate them out by business function? How are you going to increase the, the maturity curve, right? How are you going to— where's the gate between human approval and human oversight? Where's the gate between oversight and autonomy? Like, those are gonna be definitions, thresholds that you define as an individual. It's a little bit out of scope for what we're gonna do cuz we're just building an agent. But these are definitely the very next questions you're gonna have, right? Is like, how do I manage these at scale in my business if I go from one to 5 or 1 to 15 agents. And if Satya Nadella has his way, 1 to 1 million agents in my business. How are you going to manage those? Document them, know who, which ones have a human in the loop, which ones have a human on the loop, what their escalation path is. How do you document, document those escalation paths? How to monitor, right? The ins and out, inflow and outflow of requests to the model. Like, it gets very, very tricky very quickly. So if you build these systems in place, When it's a baby tiger, you have a lot more opportunity to scale successfully over time. So today, this is the homework. We are going to apply decision rules, governance logic. We're going to set the context. So you might just write down like Google Drive or whatever you're gonna, you know, think through it. Some of you will have a simple answer, Gmail and Gmail Cal— or Google Calendar. Some of you might have Gmail and Outlook. Whatever for assistance. If you're building something a little bit more complex, maybe it's a Drive or SharePoint Drive with Python scripts in it, right? Maybe like whatever that context is, just document what it is because you're going to need it if you want to build it. And when you do this, just start thinking about building it as an SOP, right? These are all the requirements. Every time I build an agent, these are the things we're going to do. I've given you the framework, but it's not locked in. Like, this is what I've done. This is what's been very successful for the 5% of companies that are winning. But you can, you can adjust it as you work with it. Depends on your industry. If you're building software, it's gonna be different than if you're doing executive assistant type behavior. So just kind of think about that. We're no longer, you know, experimenting. You're not using this as a vending machine to get you a good answer. You're building an operational system. You're architecting a business, and this business is gonna have AI employees doing a lot of the grunt work where you as the owner or the leader or the manager are going to be supervising that work. And no matter what you think or feel about that, it is only going to become more of that, not less of that over time. So the better you get at removing yourself from doing the work to being this one who sets the boundaries, identifies the decision logic, validates outputs, like that skillset is going to become extremely valuable to you as you decide to either again scale your business or become an operational leader inside an organization. So next we are going to do tomorrow. So we're gonna jump over into VIP, uh, right after this. Um, but tomorrow we finally get to look at the tools that you can use to build out these agents. Now you should have a Word document I hope, or a piece of paper that outlines your system prompt, which I will give you a more structured format that you can use tomorrow. But we need to cover all the tools that you can use, and we're going to go from very technical tools because there's a few, a dozen engineers in the house, very technical tools all the way to no-code tools. And so I will give you a wide range of tools to choose from. But most importantly, When you're doing tool selection, what questions should you ask, right? What, what type of system do you wanna invest in? What do you wanna know about the ownership of the company? Everybody who is trying to sell you a tool that isn't Microsoft, Amazon, Google, or NVIDIA, right? Anyone who's not a big player in this space, what we call a frontier firm, is likely building their tech on a frontier firm, right? So you wanna ask that question, are you an AWS shop? Because that means different things for their security model than if they're on Azure or if they're on NVIDIA. So you wanna be able, be able to have that conversation with a vendor that you're talking to. And I'd encourage you to talk to them, get as senior in the company as you can. I always ask for a meeting with a CEO if I'm gonna invest. And again, some of these companies are small, some are big, but you'd be surprised how inter— now if all of us ask for CEO-level meetings, I don't know how well that'll go. But, but depending on your level of investment, you'd be surprised how high in the organization you can go. How do you test it? How do you monitor it? What are inherent tools? What are tools you can add? What's its integration? So I'm gonna give you that roadmap tomorrow of the tools. We're gonna demo a bunch of tools. Uh, but then we're also going to give you a checklist of questions that I'll cover in the presentation. It'll be available in the slides that you can use to vet these vendors that you're going to be using. Because I wouldn't recommend, unless you're like hardcore engineering and you just wanna build an agent from scratch, you can do that. But that's not where the, the scale is, right? Scale is in the commoditization that's happened in these platforms. And I wanna give you the right tools to use, the right questions to ask to make sure you pick the best platform for the type of agent that you're building. So that is it for today. I know it was a bit of a whirlwind tour, um, but these are hard questions. And I'll tell you, before AI, these were hard questions and businesses lived or died by answering these questions. If you answer them, you do better. If you don't answer them, you'll struggle. It is not a new thing, but the— I think the better scenario we're in now is that we control the narrative to these AI agents, and I can talk to them all day, at night, 2 in the morning, right? I don't have to wait for that employee to come in at 8 o'clock or 9 o'clock in the morning before I can train them. I can actually be working on these consistently over time until I get it right, and that That has been really valuable to me. All right, so is it Jahan? Jahan? Oh, good. Okay, awesome. So I have a question from Jahan, and then we will— VIPs, you have to wait until I leave for us to be able to start the next one. So you can just hang out if you want. Otherwise, you're welcome to go.

Thank you for being here.

We're a little over today. I'll do better tomorrow. And thank you for being here. I hope you enjoyed the session. You have this link. You'll get the video recording, and I think I put the homework in, but you'll see the homework in the— like I did yesterday. I'll just post the homework thread, and if you want to answer those questions, please do. You can either share it with us or keep it to yourself, but sharing is caring. So, I see you, Jan. You missed— I don't see you anymore.

I'm here.

Oh, good. There you are.

So, um, Can you help me out because I don't have any background on business or my AI skills. I think they're like very poor yet.

Yes.

Still. So the question is, I want to create a bookkeeper, but I don't know how to ask the question to ChatGPT or Perplexity because I need to provide context, but I don't want to lead the witness.

Yes.

The issue that I'm always having with AI is that I'm giving too much information that is sidetracking it from the main task. So what would be the keywords or the structure that I need to use for specifically asking, I want to create an, an, an identity workflow for a bookkeeper. What, what are the needs, the things I need to give it for context?

Yeah. So, well, the first thing I would break down, because bookkeeper is a big bucket of stuff. So I would break down bookkeeper into different tasks and then focus on a single task. So maybe it's— or maybe 3 tasks. But what is— so yesterday we asked the question, you know, like if you were going to walk away from your business and you had a bookkeeper agent, you know, what's the one thing you want that bookkeeper agent to do, right? If you're gone for a week or a month. So what is one thing that a bookkeeper does that you can focus on?

Yeah, tracking the expenses and income.

Perfect. Expense tracking. So that's great. So that's where I would start because now like bookkeeping is just such a broad term. That's why an AI model would not give you very good details, um, because it's gonna be very vague. Uh, so now you've got expense tracking. Um, and now I would ask, uh, the same question, but now through the lens of an expense tracking system, right? Because that's now one agentic process that your entire bookkeeper role might have. Um, minimum remarkable product. Now think about Now you can answer this question a little bit easier, right? Because now I can say, okay, well, for an expense tracker, what am I going to need? Oh, I'm definitely going to need receipts, right? And then now when you see about receipts, you need to make a decision on, do I care about— from an IRS perspective, do I care about receipts that are $75 and less, or do I care about receipts that are only $75 and over, right? Um, do I want to recategorize receipts that are at different price points that are in different categories. One of the things— oh goodness, I'm going to try not to— it's a slippery slope. But there's really— there's a lot of really interesting AI systems out there. So, for example, Microsoft has what's called Forms Recognizer, and it's a receipt— I even think it's called Receipt Recognizer or something like that. But, but it is built for receipts, right? So it's an AI model that its whole job is you give it a receipt and it analyzes the content and turns all of that data into metadata you can now mine. So now your job becomes, okay, what do I want the receipt tracker to do? Do you want it to create a report? Do you want it, you know what I mean? Like, that's the next question. So the context will be for receipt, uh, or for expense tracking, um, is what type of expenses do you want to track? Uh, how, what's the number? What's the constraint, right? On the type of maybe the financial amount of the each of those receipts. But I think that's the key is just breaking it down to be a much smaller component cuz bookkeeping is too big. And then from there, once you've got a receipt tracker working, you're like, ooh, okay, what's if, if they have this, what's the next thing they want? Maybe they wanna do P&Ls or based on that, or you know what I mean? So now you'll, maybe you do a P&L and then you do a balance sheet. Each one of those could be different Agentek systems that you train and then you put each agent in a workflow, right? So one hands off to the next. Does that make sense?

So yeah, I'm basing this simplified workflow. I am the one that is providing the receipts manually.

Yes.

So if I want to create like a full workflow, I need to create an agent that will retrieve receipts for me.

Or you're going to create like a Google Drive where you drop all your receipts. Some applications say email your receipts to a location, and so you have one email inbox that contains emails that have all your receipts. So that is up to you, um, but you do need to create a mechanism to feed those receipts into the agent.

Yes.

So you don't want to— obviously you don't want to like do that one by one. So I usually would identify some kind of storage mechanism for those receipts.

Yeah. Okay. Thank you.

Yeah, you're welcome. And we're gonna— somebody asked about Lovable, uh, in the community, so we're gonna do a Lovable demo and maybe we'll do receipt tracking because that's really fun. Um, fun example for lovable development. Alright, are you ready? Let's go. Break, break. Okay, so have fun. Go build the rules. And again, some of you are gonna get excited. Feel free to start building in any tool you'd like. But tomorrow we're gonna focus on the tools. Thursday, we actually pick a tool, which I'm gonna give you. You don't have to use mine. I'm just gonna give you one. And then on Friday, we talk about optimization and making sure your AI doesn't doesn't hurt anybody. So have a great day. Thank you so much. And VIPs, I will see you in the link that you have on the calendar. Um, I'll see you there. If you have any trouble, just DM myself, Sherilyn, uh, Sky in school, and we will just ping you the link directly. Okay, thank you very much. Hopefully I'll see you all there. I tried to catch all of you.

Okay, see you soon.

All right, bye. Have fun corralling Write the instructions, build the rule sets. I'm trying to find my way out of here. I don't even know. Oh yeah, okay, I see it. Where's the exit? I can't get out. Okay, have fun everybody. See you later. I'm gonna grab the chat first before I leave. Save chat. Okay, bye.