A (3:29)

Okay, so to flip it on its head, what I'm really hearing you say is that just about anybody can develop workflows. And secondly, these workflows can be exceptionally creative, where in the past they probably were rigid without AI. Is that fair to say?

B (3:45)

A hundred percent, yeah. So I used Zapier the first time in 2015. So that was a few days ago. Right? It's been a while. It's a decade right now. And so tools like that existed for a very long time, but all these tools knew to do before is take data from one place and move it to a different place. And it was still very powerful and worth doing. But the ability now to actually know what the data is and act based on that data, whether it's qualitative data, and I can do math and make decisions from a budget perspective, from a traffic perspective, from whatever perspective you want, or whether it's qualitative data, which is could be anything. And now analyzing which category does it fall into or where do we stand from a feedback perspective or response for like, all these kind of things are now possible because AI can actually do this.

A (4:41)

I love it. And I would love to explore what is the benefit, what is the upside when you connect AI and custom GPTs into workflows? What does it make possible? We're going to talk about the how in a few minutes, but let's just talk about like the. The carrot, you know, that might incentivize people to give this a listen again.

B (5:00)

I think I touched on it. I'll just dive a little deeper in most of these processes. Okay, let's take something conceptual. You have a revenue coming in from one source, and you want to analyze it to decide what to do with that incident information. So something that I built for one of my clients, which is a great example, is they have a Shopify store. Okay, so in that Shopify store, they have orders coming in, but they are mostly a B2B company. So they have a very small B2C channel, but they're mostly a B2B company. And what they did is they had a person that would once a week, whenever they had time, which wasn't every week, as you can imagine, they would go and start looking at all the past orders and start looking at the companies or where these orders came from and start to investigate them manually and say, could this company be a B2B client for us? And then instead of generating $1,000, they could be generating $150,000, which makes a lot more sense. How do we invest the time in doing that? And old school automation couldn't do that. Old school automation can get you the information. Here is the order that came in from Shopify. I'm going to send you a Slack message every time something comes in. Awesome. Now I have a queue in a different place which maybe from a user operations perspective that's helpful, but it doesn't actually solve the problem. The problem is knowing whether that person is a potential client on the B2B side. Which means it's a combination of two things. The first thing is, are they the right company? Is it a company that will actually buy this amount of volume from me? And the other question, is he the right person? Is he a decision maker in the company? And there's a even reason to talk to that person. And so what this automation does right now with AI in it is instead of just moving the data to Slack or email, it actually gets the information from Shopify and then it does two steps. Steps number one is go find who this person is and what the company does and it creates a summary. So now instead of having a name and an email, you have a page. Half of it about the person, half of it about the company. The second step says, okay, if this is a person, this is the company. Do they fit into our ICP buckets that we have on the B2B side? In their case, it's anybody in the hospitality business, anybody. So restaurants, hotels and so on. Anybody who builds any kind of entertaining spaces, Anybody who has high end homes or home builder or designer. So they have multiple categories of potential clients. Does that person in this company fit into one of these buckets? This is again an analysis phase that historical flow only systems couldn't do. And now based on the outcome of both these steps, it will do one or two things. If it's not a potential client for the B2B side, it will send a Slack message, say, hey, we got another order. Here's some more information about the order that they didn't know before. So then that human person that did it before can still pick up the things that the AI missed, but with a lot more information. But if it is a potential actual client for the B2B side, you will create a draft email in the outbox of of the relevant salesperson based on the region and stuff like that, saying, hey, read this, if this looks fine, this is what they ordered, hit send. And these are the magical things that you can do now with AI. And this is just one example, there's a gazillion of those. But just think about every time you needed a human to evaluate the data, do some calculation in order for the rest of the automation to run. Now you don't need the human in those steps.

A (8:50)

Got it? Okay, so I'm going to try to tell you what I'm hearing you say, number one is for folks that are listening. Creating workflows that have AI in the middle of the workflow allows you to add a layer of intelligence into the workflow that historically needed to be done by a person and that person probably didn't have time to do it in the first place. Right. So it opens up new opportunities in the case of your client that you're talking about the Shopify customer who's looking for customers that are potentially bigger prospects. I would imagine other benefits of this are it's also extremely cost effective. Right. Because now it's not taking limited time from a very busy employee, it's actually being done for next to nothing by AI. Is that correct?

B (9:40)

Yeah. And it's actually, you're touching on a very important point. It's cost effective from another perspective that most people who start building automations with AI miss. And I'll explain that most of the people who start automation with AI. So all of these automation tools, whether using make or Lindy or NA10 or Zapier, all of them, you can go and get an API key from ChatGPT. Claude, choose your poison, doesn't matter any of the large language models and start sending information. So let's take the example that I gave before. I get all the research that I've done. So now it's a whole page. I said that at the end of step one, it's a whole page. And I also have two and a half pages of instructions to find whether this is a good alignment with my icp. So the page I need to send anyway. But if you're just calling the large language model and sending all your instructions every single time, that means that every time you're doing the process, you're sending a lot more data that you need. If you go to the first step that we defined, I have a page of instructions and the data that goes is one line. It's somebody's name and email address and company name. That's it. So if you're doing it the basic way and you're just calling the AI and sending all your instructions with the data, you're now sending an entire page of data in the way we're going to show people how to implement it of using assistants that I'm sending the one line because all the instructions live within the backend side of the process. And so he doesn't need to consume tokens on the API.

A (11:19)

Okay, cool. So folks, if you're not techie and everything Isar said just went over your head, don't worry, we're going to break it all down. Okay, so first thing I want to do is I want you to kind of explain. A lot of people might be conceptually understanding what automation is, right? But I want you to kind of explain automation with AI at a very simple level so people can understand. Because what you did is you showcased a couple of examples were really one example where you have a series of instructions with AI in the middle of it. But what I want people to understand is kind of conceptually, what are the pieces of what we're talking about here, Right? Where does the custom GPT kind of fit into the automation? That's really what I think it's important for people to wrap their head around.

B (12:06)

Yeah, perfect. Let's break this into the big components and then dive in into each and every one of the components. There are two big components in what we're going to show people how to do today. One is workflow automation and we'll explain what that is in a minute. And the other is a custom GPT that I assume most people know right now. But let's start with the first one, then the second one, and then dive into each and every one of them. So workflow automation is a process that existed for a while. Again, there's multiple tools. And now with AI, there's even more tools that are out there that are available. But what it knows how to do is it knows how to connect to multiple standard tools that everybody use. So what you have in your tech stack, your marketing platform, your email platform, your Google account, your places where you post, so all your social media platforms, your email platforms, your email your basically all the tools that you use are already connected to this tool. So historically and even today, most of these tools have what's called APIs. And that for some of the listeners right now gives diarrhea. Like, I don't know what API is and I don't want to know what API is. And that's fine. And this is exactly why these tools exist. They did all the techie complex heavy lifting programming that needs to happen. So now you can say, I want to know every time an email comes to my Gmail account and I want to look at the emails and every time the email has a header that says invoice and I want to do something with it and then it solves all that problem. And then the same exact concept, it works with each and every one of the other tools. And with each and every one of the other tools, it knows how to pull data and it knows how to push data into the platform. So whether it's an email platform, a marketing platform, a CRM or an erp, whatever, the thing is it has the standard functionality that this tool knows how to do. It has as simple connector. Think about it as having a USB to anything you need. And then you don't care how the USB works, you just know that it works. You plug into it and then it works. So this is how those workflow automation tools work. And in them there are different levels of complexity and different levels of capability. And these two things usually contradict one another, right?

A (14:21)

Yeah. Do me a favor and just talk about the two that you tend to recommend the most and kind of maybe just explain the pros and cons of each.

B (14:27)

Yeah. So the two tools that I use all the time is make.com and n8n. So n8n is the letter n, the number 8 and then the letter n again. The easiest entry point of all these tools to the automation world is make.com they really have a extremely simple user interface. It's very graphical based, so you can actually see the flowchart as things are happening. And to connect things, they just requires a username and password. So you give it your username and password to Gmail and you're connected to Gmail. You give it your username and password to Salesforce or HubSpot or whatever it is that you're using as your CRM and it knows how to connect to that. And so it's very easy to use. It is also very powerful. Despite the fact it is easy to use. However, it does have limitations because it's easy to use. N8N, on the other hand, is way More techy and geeky kind of platform. And it's definitely not a good place to start because you may get really frustrated very early on and then not go down the path. So I tell everybody, start with Make. If you're good and it does everything you needed to do, just stay with Make. The only reason to switch, or the only two reasons to switch are the following. One, if there's stuff that is more sophisticated, complicated, that you're starting to get to and make won't do it and then there's a good chance that NA10 can. And two is cost and the reason for that. And that will give us One of the two biggest benefits of N8N. N8N is an open source platform, which means they have a solution that works exactly like make. You can sign up, you pay them per usage and you work exactly like make.com or zapier or all the other tools, but you can also just install it on a third party server. And that sounds really complicated.

A (16:14)

It's like WordPress, right? It's almost like WordPress, isn't it?

B (16:17)

Yeah, yeah. So it's actually a lot easier than WordPress. They have multiple platforms. If you just Google right now, run host N810, you're going to get a list of 10 different companies that for five or six dollars a month will host any 10 for you. And all you need to do is go into the platform and say I want to install something. And said okay, what do you want to install? And they have 67 different things. And in the search bar you type n8n and you click go and you have an n8n install that you basically own. Which means then it doesn't matter whether you run one automation or 6,500 automations, you're going to pay $7 a month.

A (16:51)

Plus whatever fees to use the AI tools, of course, right?

B (16:55)

Yeah. So the AI tools are separate, but to run the workflow automation core, instead of paying per usage, every time you run an automation, you just pay for the hosting. So that's benefit number one. The other benefit of this is obviously from a data security perspective, it if you don't want to send your client's information, your financial information, whatever it is that you don't want to send to a third party. Now you're hosting this server and if.

A (17:20)

You'Re technical, you could host it inside your building. Right.

B (17:22)

100%. Like it's very easy to install it on a local machine if you really wanted to. But the reality is most of the third party platforms today that Host it probably have better security than the machine in your side you're building. But yes, you are right.

A (17:36)

And N8N and make probably have at least a thousand different tools that they interface with. Does that sound about right?

B (17:42)

Oh yeah.

A (17:43)

MAKE probably has more, I would imagine. Is that right?

B (17:45)

I think the number for Zapier is like over 7,000. I think the number for N8N is over 3,000. And on make, I don't even know. The reality is it doesn't matter. It does.

A (17:57)

All the big ones, right?

B (17:58)

Yeah. So it doesn't matter for two reasons. Reason number one, there's a 99% chance that all the tools you are using, unless you're using something that nobody else uses, is already connected. So that's reason number one. Like, think about it. As I said, 3,000 tools, okay? Most of us use one of four CRMs, one of three ERPs, one of 10 email marketing platforms, one of two either Microsoft or Google Office Suites. That's about it. So if you're connected to these, then you're good. So you don't need 7,000 or 3,000. It doesn't really matter. The other reason is that you can connect to almost any other tool doing an HTTP call, which is another diarrhea reason for many of the listeners. And I don't know how to write HTTP calls.

A (18:41)

Is that a webhook? Is that what they call a webhook or something like that?

B (18:44)

No. So there's two different things. So that's another thing. Right. So webhooks can trigger things back and forth, but transferring data can be done with an HTTP call. How does that work? I don't know. So how do I use HTTP calls? And that's a great trick to that all of your listeners are absolutely going to love. There's all these new agentic browsers, right? So we just got Atlas from ChatGPT, but before that there was Comet from Perplexity, which still exists. Yeah, it's awesome. I use it every single day exactly for the purpose I'm going to describe right now. So Comet or any of those agentic browsers have an LLM that can see what's happening in the browser. And so every time I need to do something that is more sophisticated than I know how to do, and I'm a monkey, so most stuff is more sophisticated than I know how to do, I will open the agent on the side and say, hey, look at this automation. I'm trying to do this. What should I do next? And saying, oh, you should probably add this node which is a step in the process. Like, oh, okay. And then I'm not just a monkey, I'm a lazy monkey. It's like, okay, add it. And then it actually goes in and adds the node. And they're like, okay, how do I now connect this to this other thing? Said, oh, you probably need an HTTP call. Like, I don't have a clue how to do that. Said, oh, I'll go and research. And then it comes back with the research, says, you should do this. I'm like, I don't want to do this. You go and do this. And then he goes and does the thing. It actually fills out the form and writes the code that it needs to write in order for the thing to work. And two out of four times it will work the first time. One out of those four times it will take you three rounds to ask you questions and once in a while you will bang your head against the wall and we'll actually have to go and do some YouTube research yourself. But it's one in every 10. The rest will probably get solved very, very quickly. And, and it sounds stupid, but I don't actually know what it does. I know that it works. I know that it gets the input that I gave it and provides me the outputs that I need. And from that perspective, I don't need to know anything more. And it's exactly as if you would have hired a person to do this. You know, you have to explain to them what you're trying to do. You have to explain what data that you have, you have to explain what the output that you need to get, but then the agent knows how to do that. And so this is my trick to going from something very basic to something more advanced, either inside of make or upgrade to N8N to get some additional benefits.

A (21:06)

Hey, before we continue, I wanted to tell you about an upcoming live training that could be a game changer if you've been experimenting with custom GPTs. What if custom GPTs could actually scale your business? I invited Wendy Breakstone, known as the Custom GPT Gal, to lead a special workshop for AI Business Society members. In this training you'll discover the three types of profit ready GPTs every scalable offer needs so you can stop guessing and start building with purpose. How to identify the exact friction points in your business where custom GPTs can deliver results faster while freeing your time and and which custom GPTs to build first. Because they're not all equal and some will move the needle a lot faster than others. For sure this training is happening live for AI Business Society members on December 11 and it will be recorded. If you're ready to scale your business with custom GPTs, head over to social mediaexaminer.com AI to join us. Trust me, this might be the strategic framework you've been missing. Head over to socialmediaexaminer.com AI to sign up now.

A (22:18)

So you set a great foundation on what these basic automation platforms are capable of doing. Well, they're not basic, they're sophisticated. Now, let's connect the dots into custom GPTs, because up to this point, we're talking about automation that is pretty rigid, right? Like you gotta follow a Rule, you got APIs, you're sending data back and forth, Right. But where it gets creative is what we're about to talk about next. So let's talk about custom GPTs and how they kind of work inside of tools like make and A N. Yeah, awesome.

B (22:50)

So first of all, let's define we said earlier, I promise that I'm going to explain what custom GPT is and how do they connect. So we explain the first half, which is automations. Now let's talk about custom GPTs. So custom GPTs live inside of ChatGPT. It's on the left side, right on, kind of like the second block of data on the left menu. If you haven't built custom GPTs, I highly recommend that I have probably 25 or 30 that run most of my business. But what they are is basically a repeatable set of instructions. If there's something that you do the same way every single time, you can turn it into a set of Instructions and have ChatGPT do it for you. So what does it have? It has some kind of an input. I'm going to give you this data and then I want you to do this analysis, process, manipulation, whatever it is, and I want you to create this output. Let's take two very unrelated examples. One of them is I want to compare my sales data to my marketing expense. So I have data from two different systems. One of it is my accounting and the other one is my marketing platform. But it has the same format every single time. I open a specific report and I hit export to a CSV and I have two CSV files. I can create a custom GPT that will know how to connect the dots between these two information. So the first time I have to understand how to do this and write the instructions. I actually do this with the help of AI. But then all I have to do is upload these two files and click go and it's going to give me the report in the format that I want it because it knows exactly how to do so. That's one example. Another example is writing hooks. We all know, all your listeners know that very, very well, that the most important thing in any content that you create is the hook. Whether it's the subject line in an email, the first line in a social media post, the first paragraph in a blog post. All of these are what's going to get people to actually read what you want to say. I have a custom GPT that generates hooks. It has over 750 examples of successful hooks that my team keeps on harvesting from different channels. And now I throw in whatever I want to write and it suggests three different hooks for me. And there's again, a gazillion other examples. I don't write my own proposals. AI writes my proposals. I give it a transcript from a transcription tool of the meetings with the client and the emails I exchanged, and it writes the proposal. So there's all these different options that you can do because these are repeatable instructions.

A (25:17)

But.

B (25:18)

And there's a but, by the way. First of all, it doesn't have to be custom GPTs in ChatGPT. If you're a Claude user, it's called Project. In Claud, if you're in perplexity, it's called Spaces. If you're like, they all have their own.

A (25:28)

Yeah, Google Gems. All these things. Exactly.

B (25:30)

Google Gems, yeah. Like, they exist everywhere. The problem is this thing lives inside of ChatGPT. There's no way to connect it to the outside world, which means every time you want an input and an output, you. You have to copy and paste information into it. And this is not what we're trying to do. I want to eliminate me copying and pasting, because then I'm the bottleneck or somebody in my company. And so OpenAI gave us a backend version, a server version of a custom GPT.

A (26:04)

The people behind ChatGPT, by the way, for those that might not make the connection. OpenAI is. Yeah, the company. Yep.

B (26:09)

And they call it an assistant. What is an assistant? It's a custom GPT that, instead of having a regular user interface, can talk to other pieces of software, which is exactly what I'm using in all these cases. So where does it live? How do you find these magical assistants? And so you got to go to platform.OpenAI.com and then when you get in, you log in with your regular ChatGPT account, so you don't need a separate account for this. And Then the only difference here is in the ChatGPT universe of OpenAI, you pay $20 a month or free or 100 or whatever level you are, and you can chat to it up to a certain limit. And on the backend side, on the API side, you basically pay for tokens. So you will have to give it your credit card and then as you use it more, you will pay more. Unless you are a giant and you're going to run I don't know how much. It's completely negligible because right now these models run on cents for every million tokens. So a million tokens is, let's say, 700,000 words that it is going to generate and that's going to cost you a dollar. So it's still very, very close to free. So people say, ah, I don't want to give up my credit card. I don't know how much it's going to consume. Well, you're going to tell it how much you're allowing it to consume. So I'm allowing you to use $10, that's it. And then you're telling it what's going to happen when it's going to run to $1, do you want to renew it automatically or not, and so on. So you have full control. But still the pricing is very, very low.

A (27:40)

Question mark on that. You've been using these OpenAI assistants for a while now. What's the biggest bill you've ever had? Are we talking dollars a month? Less than 20 bucks?

B (27:50)

Yes.

A (27:51)

Okay, so it's pretty inexpensive. Okay, so when you think about an OpenAI assistant, clarify to people that is this like they've got a bunch you can choose from or do you have to create these assistants? Help everybody understand that?

B (28:05)

Yeah, so it is exactly like a custom GPT when you click. So we said go to platform.OpenAI.com and then you go into. They call it the dashboard. And inside the dashboard there's all these different options on the left and one of them is an assistant. You click on that, you're gonna have nothing. A create button.

A (28:21)

So it's a blank page.

B (28:23)

It's a blank page with a Create button. You click Create, you start creating these. But the interface inside the creation of it is exactly like a custom GPT.

A (28:31)

Oh, okay.

B (28:32)

It has a name and then it has a description. I actually don't think they have descriptions. I think they have a name and they have the actual instructions on how they actually work. You can pick the model and then you can give them data, which is the Knowledge base.

A (28:45)

Explain the model. Explain the model, by the way, just for people, because there's a lot of models, right? So like, do you just use ChatGPT5 generally, or do you let. How's that work on the model side?

B (28:56)

So, yeah, so you can pick and choose the model. What I tell people usually is if you're not running on a big volume, then just use GPT5. Right. Because it's going to be negligible in pricing. And that's it. If you're going to do a very high volume of queries, what is worth doing is testing it. And they actually have a testing environment inside of platform OpenAI where you can use it, like using a custom GPT. You can actually talk to it right there and then.

A (29:21)

Oh, and you can see if you like the output of it or whatever.

B (29:23)

Exactly. And then switch the model and see if it works. So if you can work on a model that cost you 10% of what GPT5 costs you, and it's still running perfectly, why pay for GPT5?

A (29:34)

Yeah. Like 4.0 is still a pretty good model, right? I mean, for a lot of people.

B (29:37)

Yeah, we all used it very happily for two years. Right. And it was awesome.

A (29:41)

Yeah.

B (29:42)

So you need to give it the instructions. The cool thing is you can actually build the instructions in a regular custom GPT and then just copy and paste them into your assistant, because it is literally the same exact thing with a very similar interface. The only difference, but that is where the cool and the effectiveness starts to be more impressive, is when it comes to the knowledge base. So let's explain for a second what's a knowledge base in a custom GPT and the assistant? And then we'll tell how to use it and why it's different. So when you write instructions, you write instructions. Okay, cool. I give it instructions. Let's take the example I gave before. I don't write my own proposals, it writes my proposals. So the instructions tell it exactly how to run, to write the proposals.

A (30:20)

But.

B (30:21)

But it has additional data files in its knowledge base that it can use to help it do the thing that you're trying to do.

A (30:28)

Because you've uploaded them. Right. That's the key.

B (30:30)

Okay, so in the custom GPT side, you just upload the files. Right. So what files do I have in there as an example? I have a very broad template of all my proposals, which is something that is stupid, that I will never use because it has every different option, which is kind of weird, but then the instructions tell it to pick and choose Just the relevant aspects that are relevant to the conversation that I had with the client. So it has a reference to work with. The other file that it has over there is a brochure that has a description and the benefits of each and every one of the training services that I provide. And so it knows how to minimize and pick just the stuff that it needs from the template and then it knows how to pour information into it from the brochure that has explanations. But the same thing, you can have pieces of code, the same thing you can have how to analyze your marketing data, the same thing you can have your brand guidelines and so on and so forth, right? Your style guide, like all these different things can be a part of your knowledge base that can be used by the AI to reference what it's trying to do. Examples. Here's a good example of our top best three performing posts on LinkedIn. Use these as a benchmark to write the way we write and to get people's attention, right? So all these things can be uploaded as knowledge base.

A (31:53)

But quick, where do they live? Do they live in the OpenAI platform or are they sent every time you query the assistant?

B (32:00)

So first of all, in the ChatGPT universe, the regular ChatGPT world, where it's custom GPTs, you just upload them to this specific custom GPT on the Assistant side, the backend API side, which is what we're going to use to build these automations, they live in one of two places. They either live inside of the assistant. You basically upload the file just like you do in a custom GPT, you drag and drop or select it from your stuff and then they live there. Or they live in what's called a vector store. And you can actually pick, when you come to say, do you want to upload a file or do you want to use a vector store? And said, okay, what the hell is a vector store?

A (32:38)

Yeah, that's a good question.

B (32:39)

I know it sounds really scary. It's like, I don't want to deal with vector stores. No. So it's basically a database. It's a container, it's a place for you to upload file. Think about it as a folder that it has access to and you have access to. So what the hell's the difference between a vector store and just uploading it to the assistance? The difference is the vector store has the ability to connect to third party processes, which is where now we're going to make the connection to what we talked about before. I can take any file or any information from my workflow automation, make.comzapier, n8N etc. And upload it into the vector store, which means now I can dynamically change the knowledge base and or the inputs that the assistant or the automation in the AI world is receiving. And this is the trick of everything, right? Because now I can change what the AI is looking at when it's running the instructions. I'll give you a simple example to make it very, very clear to people saying, okay, this sounds really confusing. I don't really understand what you mean. I'll give you two examples. One is that I actually use all the time because I teach AI courses. And the other that because most of your audience is in marketing and I think they will find it interesting. So the first one, I have a course. When people take the course, they need to take quizzes. I need to update the course every time that I run it, which is once a month, because the AI universe changes all the time. And so I update the course, which means now I need to update the questions. So what I have is I have an automation that does the following. Every time I drop a file, a new transcript of a lesson into a Google Drive, a specific folder, it triggers the automation. It grabs that information, uploads it to a vector store, and runs the assistant. What the assistant is saying. Now the other file that it has, by the way, in that vector store is constant. It doesn't change. The file that doesn't change is an example of the way I like writing questions for my quizzes. So now it knows how to write questions. And now it got a new lesson. And the instructions basically tells it, look at this new lesson that I just gave you. Look at the examples of high right questions and write new questions. And now this is the output from that step in the automation. So I'm going back to the automation universe, to the make universe. In make, the first step listens, looks into a folder. When there's a new file, it grabs the folder, it grabs the file, upload it to a vector store and triggers the assistant. The output goes back as a new set of questions and then the automation back in make saves it as a new document in a new folder, which is a quiz folder with the same name as the original file. That's something the automation knows how to do. What's the name of the file? Just add a dash quiz at the end of the file name. So all I have to do is upload the transcript and about three minutes later, sometimes one minute later, I get a new quiz with the relevant name that then has a separate automation that knows how to attach it to the people who sign up for the course. That's it. So the dynamic aspect of the data that I can upload makes it magical.

A (36:04)

Okay, we'll get to that second example for marketers in a minute. But I just want to wrap my head around the vector store concept. So when you're in OpenAI and you've created an assistant, you are connecting it to a vector store. And is the vector store a service that's provided by OpenAI or is it some third party service?

B (36:23)

Awesome question. It is in there under the OpenAI universe, under platform.OpenAI.com so where you created the assistant on the left side menu, just where the assistant exists, there's a vector store option. To make it even easier, when you build the assistant and you create a knowledge base and you click to add files, it asks you do you want to create a vector store? And you say yes. And you just create it straight from there. But you can still browse them from the left side menu and create different vector store for different use cases.

A (36:55)

And the vector store is kind of like the equivalent of a Google Drive, right? It's a place where.

B (37:01)

Yeah, it's a folder.

A (37:02)

It's a folder where files live.

B (37:05)

Yes.

A (37:05)

And the reason why you would use this and not update your system instructions knowledge base is. Help me understand that part of it. You know what I mean? Because I'm still. Because everything you just said I still think could be done with a knowledge base attached to an assistant. I'm trying to discern why the knowledge is living in the vector store. Maybe that's the part I'm just struggling with a little bit.

B (37:28)

Well, first of all, because that's the way they built it. Yeah. But second is because the vector store has. Again, I don't want to scare people, but it has an API behind the scenes that allows me to update files, delete files, upload files, and so on, versus I can put in static files inside the knowledge of the assistant.

A (37:48)

So it's dynamic. The vector store allows you to dynamically do things that would be very hard to do if you did not use the vector store. Is that effectively what you're saying?

B (37:57)

100%.

A (37:58)

So hypothetically, if you changed the way your assistant works, much better to just say all your knowledge is in the vector store. Just go in there. And then if you update the vector store and you've got a bunch of assistants that are referencing the vector store, they just automatically are all smarter. Now, as a Result of it is that effectively what I'm hearing you say.

B (38:19)

And this is exactly what I do, other than in one case. And again, I don't want to confuse everybody, but that's not my fault. It's OpenAI. That's the way they did it.

A (38:26)

Yeah.

B (38:28)

One of the tools that assistants know how to do and ChatGPT knows how to do is to analyze quantitative data. You can give it an Excel file and it knows how to create code. To analyze the data, it writes Python code. Again, you don't need to know that it's doing that, but that's what it's actually doing. When you're giving ChatGPT an Excel file and ask it to analyze it, it's actually writing Python code in the background. And if you click on the little when it's saying thinking, and if you click down the little carrot next to the thinking thing, you'll see that it's actually writing code before it's giving you the answer. If you want an assistant to analyze data in this way, writing code, the only way to do this is through a separate process inside there that's called code interpreter that works inside the assistant and not through the vector store. So that's the only exception to what you just said.

A (39:24)

Okay. Can you say that your instruction set, your system instruction lives in the vector store, so you can do it all that way, or is that not recommended?

B (39:31)

No, no, no, no. So. So the assistant is where the instructions run and where the operation actually happens.

A (39:36)

I see.

B (39:36)

And the vector store, just like we said, it's a folder pool files. Okay.

A (39:40)

You were going to give another example for marketers. Maybe you could give that one.

B (39:43)

Yeah. So every marketer out there, or hopefully every market arter, repurposes the content. Right. So you have content in one format. You and me have a podcast and YouTube channels, but other people have blog posts. And you want to use that long form content to create short form content. Post on social media as an example, or the other way around. Some people do it the other way around, right? It doesn't matter. You have one piece of content and you want to create another piece of content. You can do that with this mix of the two things that we said. So you're going to drop the content by the way, it could be wherever you're dropping it either way. So on WordPress, I just dropped a new blog post on WordPress, your make.com automation knows how to listen to this and see. Oh, there's a new blog post. It knows how to grab the information from the blog post. And now I can upload it to a vector store of an assistant and the instructions can be, here's how we write LinkedIn post, here's how we write tweets, here's how we write.

B (40:45)

A script for a video that's going to go on Instagram. It doesn't matter what it is, right? And then. And you will have examples in the vector store on how to do it correctly. But the source of data, the actual file of the original piece of content will be dynamically changing inside the vector store, okay. And then generating the output on the other side. And now the output can automatically go wherever you want. It can go straight to be connected to your platform that manages your social media or directly to be posted on a social media platform or, and this is another great trick of these systems, I said in the beginning, you don't need the human in the process, but you want the human in the process. You don't want the system just to post stuff on your behalf. You want somebody to take a look and say, yes, this is awesome, or I don't know, or I'm so glad we didn't post this thing. So what I tell people is add a step that connects it to your regular task management system, Whether using asana or Monday.com or ClickUp or Notion doesn't matter. All of these have like tasks that you assign to people. You can connect make.com to that and it can open a task just like the other stuff opens a task and say, hey, here's a task for you. Review this post that was created by the AI and fix it and only then post it. And then one of two things can happen. Either the human can do the posting or you can trigger the next step of the automation based on what's happening on Notion or Monday or Asana. So when the task moves to complete or moves to whatever step, it will pick up the updated version and then continue the automation that could be posting it, that could be sending it to the graphic designer, to the. It doesn't matter whatever the next step of the, of the process is. So you can actually build a human in the loop steps to this otherwise automated process.

A (42:36)

First of all, I'm starting to make the dots connect in my brain. I do have a couple of questions when you're adding these. What we started at the top calling custom GPTs, but we're really talking about is OpenAI assistance. When you add these assistants in with Vector Store into any automation workflow, are they smart enough to be able to send Back variables. For example, I think you mentioned earlier analyzing Shopify orders to try to see whether or not somebody is actually part of a much bigger brand that is just kind of secret shopping your product and are they intelligent enough to like just send back a variable like say, okay, OpenAI. Your job is to look for this and go out and do some research and come back with all this information and then signal to the system, yes, they are. And here's all the data. Are they smart enough to do that or does it have to be done in multiple steps?

B (43:31)

The answer is yes. And when we said in the beginning it makes your automation a lot more flexible is because of what you said, right now I can qualify anything that I want based on a set of criteria, right? It's going to follow the thing to the T. But if you say this is the criteria, tell me if this is a B2B or B2C. Tell me if this is a big enough order or not a big enough order. Tell me if this is a good post or a bad post based on our brand guidelines. Tell me whatever it is that you can clearly define, then you can send a mix of different outputs. So the output could be a yes or no, could be a score between one and five, could be whatever it is you tell it to provide as the output which can then guide the next step of your automation. Because in all of these automation tools there's an if function and the if function basically splits it up to 1, 2, 3, up to as many options as you want. And so it could be a yes, no question. And then there's two paths continuing, but it could be a score between 1 and 5 and then there's five exit strategies out of that if that it's gonna go to or going back to the example I gave before, now that I know that this person is from Jamaica, the salesperson is the Central America salesperson. So I want you, I want part of your output to be this email address that is off that person and then it's going to go to that person. And so you can do a lot of really cool things that going back to what we said in the beginning makes your flow of automation significantly more dynamic because you can make decisions based on the data itself and not based on just moving data around.

A (45:17)

How prompt compliant are these things? Because this is one of the things that a lot of people are probably concerned about. Because if you're setting up an automation where you're depending on it sending back a certain variable in a certain way, and for whatever reason, it decides to do it differently, that's going to break your automation, obviously. Right. So have you had that problem?

B (45:34)

So the answer is yes, obviously. And there are two important workarounds. Number one is what I said before. You put a human in the loop, check step in every critical step of the process. So we already covered that. Number two is being very specific with your prompting and testing it up front to see how it's going to behave in different scenarios. And the reality is, number three, every now and then you'll have to fine tune and find things.

A (46:03)

Couldn't you have a quality assurance AI go over everything too, if you really wanted to?

B (46:07)

So you can, Right. So your QA step can definitely be another AI checking the previous AI to verify the information is correct. I think anything that's going to get sent out, such as a proposal or a post or whatever, you should probably have a person in the loop verifying one way or another, anything that's internal, you're saying, okay, worst case scenario, somebody will catch it and we should be fine. There's a lot of art and science to good prompting and as you get better at it, you'll get better and better in writing these instructions. And I said that before, I never write the instructions myself. I actually run a very long conversation in the tool itself. So ChatGPT or Claude or whatever, explaining what I'm trying to do, having it do it together with me, and then only once we reach a solid output, I ask it to write the instructions for me and then he writes two and a half pages of highly detailed, amazing instructions that I can never ever write and that becomes my version one. And then I keep iterating from there.

A (47:08)

Isar Matis this has been a really fascinating exploration. If people want to connect with you online or potentially work with you, where do you want to send them?

B (47:17)

So the first thing is the podcast you said you saw. People here are now a podcast player if they're listening to this. So just open whatever it is that you're listening on, look for Leveraging AI. That will be the first place, by the way, if you're on your podcast player, stay on the podcast you're on right now. So AI explored and give Mike a five star rating and give him some kind of a good review. That helps a lot and you don't have. That's your part of the equation, right? If you want to help, just give Mike a five star review and write a comment. Thank you. So the first thing is my podcast, Leveraging AI. I actually have courses that teach AI and These courses, some of them are open to the public. And there are two courses right now. One is a basic course that's called AI Business Transformation Course that will teach you the entry level, how to prompt, how to create images, how to create videos, how to do the basic stuff. And then we have a course that teaches exactly this, what we talked about today, but in very high level of detail, step by step, including templates that you can copy straight from the course and start using. And so both these courses are available. Mike is going to be gracious and give you a link in the show notes to get to these courses courses. And if you do that, you can use a promo code called AI explored, all uppercase, to get 100 hours off the courses. And then the last thing, if you just want to find me or work with me or whatever, the easiest way is on LinkedIn. And I'm Isar Matis. So I S A R M E I T I s and it's a mouthful and you never heard that name before, but that makes me the only isar Matis on LinkedIn. So after years of having to spell and explain my name, I'm very grateful.

A (48:51)

To my parents for Isar, thank you so much for coming on the show today.

B (48:56)

Thank you. This was awesome. As always. It was a real pleasure being here.

A (49:01)

Hey, if you missed anything, we took all the notes for you over@socialmediaexaminer.com A83. Be sure to follow this show on your favorite podcasting app. And if you've been a longtime listener, would you give us a review on whatever platform you're listening on. And do let your friends know about this show. Also, 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 am 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.

B (49:33)

The AI Explored Podcast is a production of Social Media Examiner.

A (49:40)

Get your tickets to AI business world right now by visiting AIbusinessworld live.