Loading summary
Transcript83 lines
- [00:01]
A
We recently announced our new conference, AI Business World. Here's a sneak peek at what you'll discover when you attend. Prompting Mastery Master proven strategies to get the quality results that you need every single time. AI Tool Selection. Learn exactly which AI tools to use for which marketing tasks. No more decision paralysis or guessing. AI Workflow Integration. Discover how to seamlessly integrate AI into your daily work, saving you hours every week. Now, there's two different ways to attend. First, if you already have a ticket to Social media Marketing World you're in, you don't need to do anything else because AI Business World is included at no added cost. But if you want to get tickets only to AI Business World, all you got to do is visit AI businessworld Live. Again, visit AI businessworld Live. I can't wait to see you there.
- [00:56]
B
Welcome to the AI Explored podcast, helping you put AI to work. And now, here's your host, Michael Stelzner.
- [01:07]
A
Hello, hello, hello. Thank you so much for joining me for the AI Explored podcast brought to you by Social Media Examiner. I'm your host, Michael Stelzner. This is the podcast for marketers, creators and business owners who want to know how to put AI to work. AI agents represent a shift in the way we do work. People say that chatbots are AI agents, but is that really true? In today's episode of the AI Explored Podcast, we're going to explore how to build AI agents. My special guest is an agentic AI practitioner who helps businesses deploy AI agents. She's the co founder of AI Build Lab where she helps businesses scale with AI. Her agency is Mavenly and her course is how to scale a business with AI and agentic workflows. Sarah Davison, welcome to the show. How you doing today?
- [01:57]
B
So great to be here. Thank you so much for having me. Michael, I'd love to hear a little.
- [02:00]
A
Bit of your story. How did you get into AI?
- [02:02]
B
I feel like it's a story that a lot of people have probably by now is that ChatGPT hit and I was busy minding my own business and started to get really interested in what was going on in the whole AI space until I was spending so much time in it that I was like, wait a minute, maybe I should make this a profession, maybe I should make this a thing. So I guess to answer your question, I don't come from like a traditional machine learning background or any of the big AI companies. My business partner and I are self taught and we taught ourselves this stuff. We went, we are from the ground AI practitioners, so we learned from Actually doing the work out in the real world. And yeah, that's been our journey.
- [02:44]
A
What's been your background? Like, tell us a little bit more about what you did before you got into AI.
- [02:49]
B
A lot of different things. I was in corporate for a while. I was in strategic alliances and business development for like big banks. I was brokering deals between the banks and like the airline relationships for the credit card points, that kind of thing.
- [03:03]
A
Okay.
- [03:03]
B
So that was good fun and I had a couple little startups as well and so I've definitely got the entrepreneurial spirit in my genes for sure.
- [03:11]
A
So when you and your partner decided to go all in, like when was that? When did you all start your business?
- [03:16]
B
Tyler is my co founder and business partner and we sort of actually randomly met on the TikTok FYP. It's so good. We were both geeking out on AI stuff and we I guess started connecting through comments and things like that. And then we randomly found ourselves in an AI community, started doing a fellowship and worked gelled really well. And then we just started doing projects. So we got known for being the people that solved the really hard problems in AI and people in our space, consultants and things like that, kept referring work to us. And yeah, we started building our practice as practitioners actually doing this stuff in the real world and essentially got to a point where we had done so many of these implementations, especially towards the AI agent front, that we figured, you know, we've got a system around things. We know how to actually make things work in the real world, not just in theory. And why don't we teach other people how to do it too, and companies how to do it too.
- [04:15]
A
So did you get started right after ChatGPT? Because we're recording this just about three years after ChatGPT was launched. So you got in pretty early then, basically.
- [04:23]
B
Yeah, yeah. It. It doesn't feel early.
- [04:25]
A
I know. It's changed a lot, hasn't it?
- [04:28]
B
Yeah, I mean, it feels. It feels sort of brand new and also like it's always been here.
- [04:31]
A
Yeah, 100%. Okay. So you've been working with a lot of different kinds of people over the last two to three years, specifically helping them with AI and AI agents. What do you believe is one of the biggest misconceptions people have when it comes to AI agents?
- [04:49]
B
Yeah, I think because, you know, AI agents was this like big sci fi thing that was a little bit out of reach a couple of years ago. People were talking about it, but the technology just wasn't there. Like people were projecting what's going to happen. So it sounds kind of sci fi and there was a lot of confusion and now that it's kind of hitting the public discourse again because the technology is there and is actually being rolled out and it's actually happening. I think there's a lot of confusion around what AI agents actually are. Where people are confusing like agents with chatbots and agents with automations or AI automations. And they are quite different, is a spectrum of how you interact with AI and AI systems. But AI agents are quite different in terms of how you need to consider developing them, deploying them and ensuring, you know, you have the guardrails and safety around that.
- [05:40]
A
Love it. And talk to me a little bit about like autonomy because I know this is something that you have an opinion on when it comes to this.
- [05:46]
B
Yeah, so. And that's the definition, like when people ask us, tell us what's the difference between an AI assistant and an AI agent? And what I say is the word agency, it's the word, you know, it comes down to autonomy. So probably a really relatable experience to explain this is if you are interacting with ChatGPT and you're trying to write a blog article and you're like, hey chat, I've got this idea. And you know, you're sending this idea to chat and chat, sending you something back and you're in this constant like iterative cycle. And you know, honestly it's, it's great. AI assistants like ChatGPT feel like magic, but you are in the driver's seat. So nothing happens if you are not part of the process. AI agents are essentially AI tools with autonomy. They have the ability to reason, plan and execute and do tasks on your behalf. So they kind of read their environment, understand the task at hand and have access to tooling and to decision making processes, planning to actually do the task for you. So you go from this concept of AI assistance as this tool that you have to drive to, almost like you're delegating to an AI team member when you have an AI agent and when you have an agentic team or an agentic workflow that's having multiple team members, you and you know, work together to get a task done for you autonomously.
- [07:11]
A
I love that. So what I'm hearing you say is that an AI assistant is what a lot of people experience with custom GPT or cloud projects or any kind of project where they're just interacting with it and they're getting responses as a result of giving it a data or stimuli. But with an agent, it's a little bit more like a contractor, right? Where you task it to do something and it actually goes out and does it on your behalf and it says it's done. And that usually means you have to give it a certain amount of control, which we'll talk about that a little bit later. But what I'm curious about is when AI agents are done well and you've done them before, so you can speak to this, what are the benefits? What are the upsides? Just so everybody can understand, like, what this can unlock.
- [07:59]
B
You know, there's a lot of people out there who say, don't anthermoporphize AI. And we're going to be the people who do that. Because it really is like, the more that we've been doing this work in the AI agent space and the way that we are graduating our assistance to agents to agentic teams and agentic workflows, it's kind of like having a people team, honestly. You recruit a team member and you give them a job description and then you give them a task. And in the same way that if you were going to hire an intern, you wouldn't give them the keys to the castle right there, and then right there would be a process by which you build trust and reliability in the person in the contractor to give them that level of autonomy to go and do these tasks for you. So coming back to your question, what is the benefit? The benefit is like you are able to almost infinitely scale with these tools that are representative of team members or departments that you can have that are able to do the work for you in the way that you want the work done. And I think that that's another important distinction is that when you get in a world where everyone's got access to ChatGPT and Claude and all of these different large language models, like, what is the moat anymore, right? If everybody could access to that same power. And really what this whole agentic space opens up through different processes is this ability to really be able to customize the agentic experience or the, like, the layer that you build around your company in terms of codifying what we call the secret sauce. What makes your process, your workflow, your business so unique and capturing that in these agentic systems that can scale. And that's something that your competitors cannot copy. It's something that you know is generated from your own intellectual property.
- [09:55]
A
Somebody is going to want me to ask this question, so I want to ask this question. Okay, I have secret sauce, but I'm scared to let the AI have access to the secret sauce.
- [10:05]
B
Yes.
- [10:06]
A
Talk to me about that a little bit, because I know some people are like, I don't want to give AI my secret sauce.
- [10:12]
B
And that is a good point to talk about our whole entire philosophy of, from our experience of building AI agents and now teaching other people to do it, our whole philosophy that underpins us building and deploying AI agents is what we call evolution, not revolution. Because you are essentially giving these tools autonomy to go and reason, plan and execute on your behalf. Right. And exactly like you say, and probably rightly so, from different media discussions that have happened or things that have come out around AI agents Gone wild or something along those lines. And that's often because there isn't that gradual process of building trust and reliability in these agents to get to a point where you can then provide that autonomy. I think the big thing that people think with Agent Kit coming out from OpenAI the other day, the ability to build agents is so accessible, right? Anyone can now build an agent, but how do you build it properly like that is a completely different conversation. So the tools are incredibly available and accessible for people to be able to do that. But the methodology behind how you actually apply the framework to ensure that your agents behave the way that you want them to in the real world is an entirely different thing.
- [11:37]
A
So here's what I want to implore everybody who's listening right now. Just listen to this entire conversation and then see how you feel by the end of it. Because I do believe that even though it's a big fear for a lot of people, given their intellectual property and procedures and all that stuff over to AI, I do feel that the benefits are worth it. Because I want to reiterate what Sarah said. First of all, you can customize these things to do it exactly the way you want to do it, and that can give you a competitive advantage and a moat. Because she set said that the whole world has access to chat, GPT and Claude and Google, Gemini and all the other ones, and they're all built on the same data set. But if you can customize something that is completely built on your set of data and your set of procedures that the rest of the world doesn't necessarily have access to, and there's a way that this data can be secure, then you actually might have a real advantage here. And that's what we're here to talk about today. And we didn't talk about this, but recently I had a guy named Carl Yeh on the show and he said that traditional businesses are going to be Disintermediated by startups that do not have the traditional restrictions that other businesses have. He said that Y Combinator, which is, you know, the company that started all these massive companies that we know today that we could just name off the top of our hand, these huge companies, they've been looking for people that want to disrupt traditional businesses. Right. And this is the thing, like traditional businesses in my mind is anybody that's been around for at least a decade, decade, and there's a lot of you listening, someone's going to come along with the tools that we're talking about today and they're going to potentially disrupt you. And it's to your advantage to understand how you can use these tools to allow you to scale faster, to accomplish more, to innovate more. And I feel like that's one of the big upsides to what we're about to talk about today.
- [13:28]
B
Absolutely.
- [13:28]
A
Do you agree?
- [13:29]
B
Yeah. And if I can actually add a little bit more spice to that disruption. Yes, Right there. I was lucky to be in Amsterdam at a conference speaking on AI and one of my people that I used to follow a lot, Daniel Priestley, in the whole entrepreneurship space, he said a line that really got to me and it just hit in the right way. And he said, because he, his audience are entrepreneurs, you've got 12 to 18 months max and your business, as it looks like, will not exist. And he said, and that could be both a good thing and like a bad thing or an opportunity for massive reinvention. And he gave the example of like when, you know, he had a business that was doing a lot of in person events and things like that, and then Covid hit and his whole business was built on that model or that infrastructure, you know, and when Covid hit, he had to pivot, he had to meet, you know, go online and do that whole piece in terms of his events business and his business massively scaled from there. So it really like hit for me because I sometimes struggle to explain what I'm seeing unfold in a way that does shed the light of what's actually happening, but also provides the opportunity and what is possible from this disruption. Because in disruptions there's always hope because things change. And it's like who's going to be there and take advantage of the changes and be in the lead. And I think that we're in this moment right now. You know, it's really exciting if you're paying attention right now. I don't want to come at it from a point I don't believe in fear mongering.
- [15:00]
A
Right.
- [15:01]
B
I believe in coming at it from a point of like, wow, like, this is the opportunity that that is possible for my business. That was never possible before.
- [15:09]
A
Fully agree. Okay, so we're here to talk about AI agents. Where do we begin? Let's just assume everybody listening here does not have one. Where do we start?
- [15:19]
B
Okay, actually, if I had a client that says, hey, Sara, we want to identify this process, I would not even touch an AI agent in the beginning. And like we were talking about before, there's a real methodology and process to how you build these agentic tools in a way that actually works. And that really comes down to deeply understanding the workflow as it is now. Because without understanding that, you're just essentially dragging and dropping nodes in a tool and, you know, saying, oh, this is an automation that works, that doesn't capture what actually happens in real life. So when we are engaged with organizations and they say, we want to go through this process of identifying this really, like, this is a really big block in our business, how do we identify that? We go deep and really understand the workflow end to end. Who's involved, why are they involved, and what is their job? And like, what happens when the process breaks? What makes Sally the best at doing this process? And what happens when things break? Like, what does the actual workaround look like? All of those things help inform us in the way that we design these agents, because then we're creating it from a place where we really understand where are the gaps, what are the opportunities? How do we really capture why Sarah is great and codify that into these systems that can actually scale that greatness out and really make the result of that workflow work at the end of the day.
- [16:55]
A
Okay, real quick, couple questions on this. I know you have more to say here, but specifically on when you refer to the workflow, what do you mean by the workflow? Because obviously there's businesses that have lots of workflows, Right? So, like, talk to me a little bit about that.
- [17:08]
B
Yeah, and that could be a process that a team is involved, end to end. There could be many things, like putting together a marketing strategy where several peoples are involved, or it could be the customer service workflow of a particular department or a, you know, things along those lines. Or how we actually approach creating new products through product requirement documents where there's a process that people follow, or a workflow or a flow to how something gets done.
- [17:35]
A
Essentially you have an example, right, about documents or something like that that we, we had talked about. When we were prepping for this, maybe you can help like explain that, then people can wrap their brain around this.
- [17:43]
B
Yeah. The reason I think a lot about document, we call it AgentIC AI or DocGen because we've had so many requests for document generation using AI agents that we just call it DocGen now.
- [17:56]
A
Explain what that is like, what the heck is document generation?
- [17:58]
B
Yeah, so it's building an agentic team to essentially collapse what's usually weeks and weeks worth of work of the biggest domain experts in a field around a specific thing. Like you could think about a brand strategy document and the consultation that goes around that, or a tech roadmap that consultants go and have with different organizations and have to develop it. So that process would normally take the best in the field, people with real deep domain expertise, weeks upon weeks to like synthesize that information, to work together to come up with a document that's essentially like a strategy or a plan or even a product requirements document or a marketing plan. Like that involves a lot of people having unstructured data. So think interviews, think research from the Internet, think like things that you gather as part of that process so that you can synthesize it and generate a document on the end of that. And often these Doc gen processes without even AI are really manual, time consuming and you know, you have your best people on it, so you can't really scale beyond that. And what we have been able to achieve with these agentic Doc gen systems is essentially imagine like taking all that information, all of that unstructured data, all that data you gathered from interviews and emails and research and all that, and dropping it into an agentic workflow where you essentially have a team of experts, an agentic team, kind of like a department, kind of like how your team is structured around being able to do this workflow collapse. A time that it would take from say 11 weeks down to 11 minutes through this process and still get the deep domain expertise that you would at the back of that. So it sounds like AI could never do that. But properly, if you go through the proper processes like we talked to, to really go deep on how that process is made and what information is gathered and what's important, what's not, what makes Bob the best at writing these things and who's the best copywriter on the team and who's the best analyst and strategic expertise and combine them and, you know, join them as a team, you are able to do that very successfully.
- [20:12]
A
This blew my mind. New data from the US Chamber of Commerce shows that 85% of small businesses using AI saw an increase in their sales and in their profit. But you likely already understand the benefits of AI because you listen to this show. Now. Consider next leveling your skills by attending AI Business World. It's our first dedicated AI event. You'll learn from 20 of the world's leading experts, the very same folks you've been learning from on this podcast. Isn't it time to take your AI professional development seriously? Get your tickets today at AIbusinessworld live. I can't wait to meet you there. Okay, so what I heard you say is that at the very start of any kind of process where you want to develop an AI agent, the first thing that you need to understand is kind of what is the end to end process or workflow that are involved? Who are the people that are involved? Why do you use those people? What uniquely do they bring to the table? What are the areas where they're maybe not strong? What are the areas where you have opportunities to improve? You gather up all this information, and somehow this is going to be used in the next part of the process. Right, but the first part of the process is to deeply understand the workflow, because otherwise it's just going to create something. Right? And in most cases, we're coming into an existing business that already has procedures and processes, some more established than others. Right. Like, some businesses are just going to be like, it's sue, and it's all in her brain. So, okay, in that case, you got to interview sue, right?
- [21:44]
B
Yeah, that's right. And, like, what we found actually, you know, working and doing these is that most people think, oh, well, they've got an sop. Let's just, you know, map our agents to that. And actually, that is where most people stop, and that's where you get, like, really bland results. Because the SOP doesn't. Doesn't capture why sue is so great at her job and why Tom can intuitively look at an application and make a. Like a gut check. You know, people who are in that space like, that have this domain expertise, don't realize that they're doing it, but they're doing something very specific that actually makes everything work. Right? And so understanding this, we found that there's different layers of what we call work intelligence when something gets done. And there's the sop, or the standard layer, and then underneath that there's the operational reality. So the SOP is like, this is the perfect world of how everything happens. And then the operational reality is like, this is what actually happens. And then there's usually band aids across the situation, and people just intuitively know the way that they fix things isn't. Isn't necessarily on the ssp. It's happening in Slack messages or whatever, and they're like, solving it and moving on. And then there's like this other layer of what is the contextual intelligence that these people have? Like, how do we capture that? Exactly, like we said, what makes sue so, like, she just knows exactly what to do, end to end, and really defining and understanding that. So that when we build this out, we're able to capture Sue's magic, the domain expertise that she has acquired and enable her to do her work in the way that leads to expectations.
- [23:26]
A
Real quick question. For people that have Sues and Toms that work for them that might see this as a threat, how do you address that? Like, how do you persuade someone listening to. Persuade their staff to be open to this and not see it as a threat?
- [23:37]
B
Yeah, and that is a definite friction point that people have around the technology and its ability to really do a lot of things really well if you know how to make it work. And we try to bring them in the process alongside us. Like, we work with sue and we give her, like, the tool to say, like, as an assistant, to test it, to say, hey, sue, here's this little thing that we want to give you to help you do your job better. Can you go take it for a drive and really tell us, does it meet Sue's expectations? And then she finds that as she's using it, she's like, oh, my God, I could do things so much faster. I love this. It represents the way that I work. Or, hey, Sara and Tyler. This thing's got some gaps. This thing doesn't quite reflect. So as part of our process, we really try to, like, engage people who are involved in a very. Like, we work alongside them developing this stuff.
- [24:31]
A
Okay, so what comes next, you know, after you've gathered up all this information, all these insights, like, what's next in the process?
- [24:37]
B
At this stage, we're like, okay, let's go build the assistants that will then become agents. So we, because we've done such a thorough job of like, really understanding the end to end process to Sue's job, to Tom's job, to what happens when things break, to what is the domain knowledge that we need to capture? Because we know all of that now. We know what agents to build. So we know the different job descriptions that we need to provide are different assistants that will then become agents. So we go about essentially developing the team that will do this workflow, like the different tasks that they'll have and we develop them as assistants, similar to how I was describing and mentioning before, that we don't want to just identify the process, throw it out in the real world and then say, fingers crossed. We want to make sure that these assistants, we're developing them as assistants first, that they are meeting certain criteria, like, we want to have trust and reliability and quality in their performance before we start providing more autonomy, before we start giving them access to tools to go and act, or even putting them together in the agentic workflow to act as a team to get a task done.
- [25:50]
A
When we were prepping for this, you talked about how, like, you're doing some MVPs and you were going to mention kind of how you do that and some of the tools, just because that might be helpful for anyone who wants to try that themselves also.
- [26:00]
B
So this is something that, you know, if anyone's interested in getting into the whole AI space as a consultant, strategist, practitioner, it really changed the trajectory of our work. We call this thing show up with a mock up. Because what used to be the same with ChatGPT and things like that, where people are like, I don't understand what you mean by AI. I know it's important, but I don't know why it's important to me. And showing up with a mockup is like, because we've done the work. These tools are so easy to pull together now that we go and provide like a mini MVP and we go, hey, go test this out. Tell us what you think, what's missing? Because that becomes our feedback for do we actually need more agents because we need more specialists involved here, or are we overcooking this thing? What happens when it actually breaks? We found something missing in the process. So we would go through that process of showing up with a mock up and gather that feedback. And then that is a circular process that helps us define how we can make our agents better.
- [27:02]
A
Do you mind mentioning some of the tools that you use? Because I was fascinated when we were prepping.
- [27:05]
B
Okay. I want to say this because this is important for people who are looking to build. If you're a business owner and you want to integrate AI, like one of the things that we are very clear about with our clients, with our students, is to be large language model, tool and platform agnostic. And that's because if you can't tell, things are moving so fast, you don't want to put all your infrastructure and all Your development costs and all of your company in one on reliant on one tool or platform. So that's my disclaimer for saying we use certain tools that work well with us, but we're never tied to one or the other. Even with large language models, we make sure that we can flip between models. So if we have like, you know, OpenAI, just drop GPT5, we can flick into that with our builds, with our client builds. But for our day to day, like for MVP testing and things like that, we like to use a tool called Typing Mind which is a, it's connected to all of the large language. It's kind of like the playground for like OpenAI and how you have it in anthropic but just centralized in one place and then you have access to tools like mcps. So it's pretty easy to build agents like really quickly to have them to test with clients. And we've done work in a tool called Cassidy. We actually teach on Cassidy because it's super, super accessible and powerful at the same time. Yeah. And we're very much tool platform agnostic. What we say to people is also find what works well for your brain.
- [28:37]
A
Yeah. So just so I'm clear, Typing Mind and Cassidy AI, it sounds like they are tools that allow you on the back end to choose the large language model like Gemini or Claude or ChatGPT. And then it also sounds like they allow you very quickly to kind of mock something up that, that someone could effectively interface with without having to like have a chatbot account or a cloud account, that kind of thing. Is that what I'm hearing you say?
- [29:07]
B
That's right. Okay. Yeah, yeah.
- [29:09]
A
So once you've mocked up and tested internally, an assistant that effectively is doing some of the tasks, how do we get to the point where we actually want to turn this thing into an agent? That's what I'm really curious about.
- [29:23]
B
Yeah. And this is the process where we, it's an iterative process. It sounds like, oh, you just flick a switch and it becomes an agent. But it's actually a spectrum of how much autonomy that you, you give your agents throughout the process or your assistant to become an agent. And like, you know, it's characterized by several different things. So you want to give again. We're going to anthropoporphize here. Anthropomorph wise. I can never say that word. It's taken me like three years.
- [29:49]
A
I'm not even going to try. So I trust you.
- [29:51]
B
But you know, you want to give this tool Like a brain, right? So like the central source of context, which is a large language model, you want it to give it memories and context so you will give it access to things that are. The large language model is the baseline of the intelligence. And then you want to contextualize it to your business and your processes and your sops and all that layer that happens on top. And you want to give your agent the ability to be able to do multi step planning and reason, plan and execute. So it needs to be in an environment where it has like an input and then it could generate an output and go through the reasoning process. And so you would want to give it access to different team members as well, so they can kind of delegate work amongst themselves. And as you get quite advanced in these multi agent workflows, you actually go, oh my gosh, these team members need a boss. Because it's chaos, there's too many agents. And so you develop an orchestrator agent which essentially becomes like the program manager that divvies up the tasks amongst the team and keeps a check on quality control and ensuring that we're meeting there. But you'll have different types of agents. So you'll have domain expert agents because of the work that you develop. You'll have QA agents that check the quality of the work. You'll have a copywriting agent, say in that Doc Gen process which will actually review and write the copy. And then you want to ultimately give these agents tools to be able to go and perform and output the tasks that could look like having access to Gmail or being able to access the SharePoint site to access the information that comes in that then goes out, that kind of thing.
- [31:39]
A
Okay, so I would love to like unravel an example. It could be your Doc Gen example or it could be something else. But I just want kind of people to wrap their head around and maybe it's something you build internally for yourself, but help people understand, like let's just see where it goes, you know, and I'll have more questions. Okay.
- [31:56]
B
Okay. Yes, I've got a good example for you. Because obviously can't use client information, but we actually teach Doc Gen in our advanced course and we've actually modeled the DuckGen workflow, agentic workflow based on Tyler's family business. They allowed us to use them as an example, as what we would do. And they have a greenhouse supplies business that does like greenhouse manufacturing and sells it to, you know, different people and part of their audience, their market is what are called market gardeners. So these are People who want to start a market garden business. So like to sell at the farmer's market or, or to sell into like you know, cooperatives to grow their vegetables and whatever and commercialize it, but in a small way, like a small farm essentially. And so the storyline of this is that Roger, who's Tyler's father, he's the guy with all the domain expertise that used to provide individual coaching to these people who want to start the business. But of course that's not scalable. And so the challenge of the doc gen process is like okay, there's this Roger and his team that provide this one on one coaching that really understand what is the soil PH level in Virginia versus LA versus whatever. Like it's really deep and you need a lot of expertise. And they, they don't actually do this but like if they were to write a business plan or a market garden plan for each individual person that wants to start a business like this, what would it look like? And so that workflow that we build out, we take them through the process as if Sara and Tyler were engaged in this project with that company. So we take them through process of. Exactly. Like we said, we don't even touch agents in the first few weeks. We do the discovery work and the SOP mapping and all of that and then we get into actually developing. Okay, well based on what we understand of that business, what kind of agents do we need? We need an agent that is specialized in the domain of greenhouse businesses and market gardens and we need another agent that specialized in being able to develop business plans and another one that's more like on the technical side of growing a market garden. So the one that knows about PH levels in Virginia versus la. And so we start to string together, there's these different specialists that work as a team that to develop a market garden plan and then what the output looks like is that the grower's business, or Roger in this case would have a 90 minute call with someone who wants to start this kind of business, add that transcript and maybe some notes into the agentic workflow and that would generate a highly specialized 35 page document custom to that person for where they live, for their budget, for their branding, for all the things that actually go into writing a market garden plan and to give you like the kind of value that this kind of initiative can create, that business of growers, their competitor charges other people like six figures for that plan. And our agentic team can, because now it's not dependent on Roger's availability or who he's Hired on his team. It's independent like he's. We've essentially scaled Roger's expertise and his team's expertise infinitely to create this like new line of business.
- [35:33]
A
Now I know people are going to say, but I don't trust the output. We're going to get to quality control in a minute. So trust me, we're going to get there. And I know people are thinking about this right now, but I love this. So as I listen to this story about Roger's business, it's true that you could create a whole bunch of custom GPTs, right? Or cloud projects that are trained up with the data and that can do each of these individual tasks. But where this becomes agentic is where now they're interacting with each other without. Because you're setting it up that way, right? They're interacting with each other and they're actually probably drafting an email in someone's inbox with the actual report, right? So that literally you can just hit the send button or actually sending it, right? So talk to me about like the agentic side of all how all this stuff works together.
- [36:19]
B
I said they can reason, plan and execute and be involved in multi step process. So as an example and in that particular workflow that we build out, we have an orchestrator agent, which is the program manager that I described. So what actually happens in that workflow in that instance is like, like I mentioned, the input is a transcript of a 90 minute call and maybe some notes or whatever, like unstructured data and you know, that could go in all sorts of directions. And so the orchestrator agent knows all of the different team members on the team. So we've got the technical expert, we've got the business and marketing expert, we've got like all of these different team members. This is the information that we have and this is how we're going to divvy out this information for the different agents. Now the different agents, they have a very narrow domain of expertise. Like, like I mentioned, they're like really in charge and that's how you can get the depth that you wouldn't get with one agent trying to do this, this whole thing, right? So they both understand their very specific domain that they need to act on. Like hey, I'm, I'm just in charge of writing the part, the section in this report that's got to do with analyzing the ph soil level in Virginia or whatever. And there's someone else on the team that's in charge of assessing like the branding and the market, like how, how they should launch their plan. And so they have an awareness of. Of their own role plus the over, like the overall team and what they're here to do. And the Orchestrator agent is able to divvy up the role, the jobs. And when it doesn't have enough information because Roger didn't ask the right question or whatever, like, comes up with another plan.
- [38:06]
A
Oh, okay. And it waits for each of the other agents to do their role before it does what it's supposed to do. Right.
- [38:13]
B
And it kind of loops as well. So there's like, this quality control. So if, like, the Orchestrator agent's not happy with, like, if it doesn't meet the expert, you know, the. This is. We have to actually look through the process and find a better way of actually generating this. So this sounds really out there and bizarre, but just think about, like, how a team would do this. This is exactly the kind of communication that Sarah and Tom would be having that goes, hey, I didn't get, like, the information from Bob on this. I can't really do my job until I get that. But they all really know, like, this is where I am in the process. And this is, like, also what we're all there to achieve. And the end result is this very specific plan for someone who wants to start a market garden.
- [38:56]
A
Okay, so quality control. So a lot of people listening are like, AI hallucinates. We know it's getting better by the day as these models get smarter, but how do we ensure that the actual plan is quality? Right. Like, talk to me about that a little bit.
- [39:15]
B
Yeah, well, first of all, we are very strong in having what is actually called an AI, a human in the loop in the process. And so in the example that I just described, this would not be hitting auto send to the person at the end. This would be sending to the team, the grower's team, to actually QA it at the end, make some edits or whatever needs to be done before it.
- [39:36]
A
Gets sent on and does that feedback back into the system so they can learn.
- [39:39]
B
Exactly.
- [39:40]
A
Okay.
- [39:41]
B
So we do, like, what is called evaluations. This is actually a methodology that is applied in AI, where you essentially give. So you know how we said we hired these team members, we put them in a team, and they all do their job. Right. Evals are like the report card to say, hey, this is how you did. This is how you performed on your task. Like, on a scale of one to five, right, you scored a three in this particular area. And the reason why you scored a three in this particular area is because you used terminology that was overly corporate. You know, growers doesn't speak like that or whatever. It could be anything. And that evaluation criteria or it hallucinated. Right. So like all of these things. Right. That evaluation criteria helps become training data that you then feed back into the system to say, this is what good looks like. This is what a three out of five looks like. We want a five out of five. And the more that you're able to do these evaluations, the more that you're strengthening the system, the more that you're strengthening the feedback loop. So that you're creating a system that gets better and better versus a system that like, hallucinates and then keeps hallucinating and. Yeah, and then you're in trouble.
- [40:52]
A
Okay, one other question for people that want to try this themselves. Are the Typing Mind and Cassidy AI tools capable of doing all this, or do you need kind of someone to come in and kind of use other kinds of tools in order to be able to pull off agents? I'm curious what your thoughts are on that.
- [41:09]
B
Honestly, it's not even about the tools. It's about the mental models, frameworks, approach that you have. And you understand the tools are just piping. You know, they're piping to make things work easier or whatever.
- [41:21]
A
So you could use N8 and you could use make, you could use Zapier, you could use all these tools to kind of pull this off. The key for it to be done well is really the operational overflow kind of stuff and the procedures that we've been talking about. Right. Because like N8N is really confusing. I mean, like, it's extremely technical. It's what a lot of people use. But MAKE also can accomplish the very same thing, right?
- [41:43]
B
Yeah. And like, the thing is that you, you can have access to those tools, but they won't do it for you if you don't know what the things are that you need to do with it. So that's where it becomes super important to actually have the awareness of not just like, I can now drag and drop and build an agent, but what does it actually take to build AI agents that work in the real world?
- [42:03]
A
You know, it's funny because there are like Operator or what the heck it's called by OpenAI, which kind of like just takes over your browser. We're not really talking about those kind of tools. We're talking about tools that are piped into your existing business infrastructure. They have access to your data, they have access to your Google Drive, they have access to your Microsoft whatever the Heck it is. Right. And, and that's kind of the key to all this stuff, right, Is to build all these like connection points and the human in the loop. I love the fact that you talked about the human in the loop because, like, I kind of feel like there probably should almost always be a human in the loop, at least for now. Do you agree?
- [42:39]
B
I agree on that. That's. I think the misconception is that because these tools are so easy that you can just drag and drop an agent, but there's actually a lot of work that goes into ensuring that you get these agentic systems that actually work. And I think it's even more important now that we're getting into the agentic era where we're giving these tools autonomy. That is a different game than like a chatgpt chat, right?
- [43:04]
A
Yeah. And they're going to require maintenance, right. I mean, there's going to be things that are going to change, right? Like models are going to change, procedures are going to change. You're going to need to be able to like have go in there and do updates. Is that true?
- [43:14]
B
That is true, yeah. So it's not a one and done thing for sure. There's things that happens called prompt drift, which is when, you know, the prompts start kind of sliding in terms of the quality and then there's, you know, like you talked about hallucination, large language, like, you know, different models drop. There's a lot of that actually goes into it. And we talked about evals. That's a whole entire maintenance, like part to it that I think in the hype, it gets lost, you know, in the hype of AI agents. And they are incredibly powerful tools, but there is work that needs to be done to make them work for the way that you want.
- [43:48]
A
Well, I love this Sara, I've been saying Sarah, but Sara Davison, thank you so much for sharing your insights. If people want to connect with you on the socials, do you have a preferred channel? And if they want to potentially work with you, where do you want to send them?
- [44:01]
B
Just connect with me on LinkedIn. That's my name, probably most active there. So I would say that that's the, the main channel. And if you're curious about like learning this stuff or wanting to know how to do this for your business or yourself as a person wanting to be a practitioner, then the Maven site with our program on there of how to scale a business with AI and agent workflows basically takes you through the whole process of what I talked about in four weeks.
- [44:30]
A
Awesome. And for those that are listening, it's Sara Davison. Thank you so much for coming on the show today.
- [44:36]
B
I appreciate it. Thank you, Michael.
- [44:38]
A
If you missed anything, we took all the notes for you over@socialmediaexaminer.com a78. Be sure to follow this show on your favorite podcasting app. And if you've been a listener, we would love a review and we'd love a share. Would you let your friends know about this show? You can tag me on Facebook, LinkedIn and or X. And do check out our other shows, 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.
- [45:12]
B
The AI Explored Podcast is a production of Social Media Examiner.
- [45:19]
A
Get your tickets to AI business world right now by visiting AIbusinessworld live.