216 | The 5 Levels of AI: What Every Business Leader Needs to Know, To Pick The Right AI Implementation with Isar Meitis

Feeling overwhelmed by all the AI tools out there?
You're not alone. From GPTs to agents to “vibe coding” (wait—what?), it's hard to know what’s actually useful… and what’s just another shiny object.

In this episode of Leveraging AI, Isar Meitis cuts through the AI noise with a practical, no-BS framework that maps out the 5 levels of AI implementation—so you can stop guessing and start using the right tech for your real-world business needs.

If you're a business leader trying to make smarter AI decisions, whether for productivity, automation, or innovation, this episode will change how you approach AI forever.

In this session, you'll discover:

• The 5 distinct levels of AI you can implement in your business today
• How to match your project’s complexity to the right AI solution
• Pros and cons of everything from basic LLMs to full-blown AI apps
• Why custom GPTs and agents are wildly underutilized by most teams
• How Isar uses automation, assistants, and even self-built tools to save hours every week
• The hidden questions that will help you decide between automations, agents, or full apps
• A sneak peek into a soon-to-launch AI-powered “Solution Advisor” built using all 5 levels

About Leveraging AI

• The Ultimate AI Course for Business People: https://multiplai.ai/ai-course/
• YouTube Full Episodes: https://www.youtube.com/@Multiplai_AI/
• Connect with Isar Meitis: https://www.linkedin.com/in/isarmeitis/
• Join our Live Sessions, AI Hangouts and newsletter: https://services.multiplai.ai/events

If you’ve enjoyed or benefited from some of the insights of this episode, leave us a five-star review on your favorite podcast platform, and let us know what you learned, found helpful, or liked most about this show!

Transcript

[00:00:00] GMT20250812-190929_Recording_avo_1280x720: Hello and welcome to the Leveraging AI Podcast, a podcast that shares practical, ethical ways to improve efficiency, grow your business, and advance your career. This is Isar Metis, your host, and I've got a really cool and different episode for you today. In this episode, we're going [00:00:15] to review how to pick the right level of AI implementation and the right AI tools for your use case based on a wide variety of different implementation options.

[00:00:26] Now this started with me having multiple conversations [00:00:30] when I speak on stages and when I talk to clients and potential clients and so on, and there's such a huge confusion right now when it comes to terminology and what are the options and what can I use and what should I use, and so on. And everybody wants agents, but they don't know what agents are.[00:00:45]

[00:00:45] So this episode is going to break the different implementation options into five different levels. Level number one will be large language models. Level number two is going to be custom Assistants. So custom GPTs, gems [00:01:00] in Gemini spaces, in perplexity, and so on.

[00:01:03] Level three is going to be workflow automation tools like Make and Zapier. 

[00:01:07] Level four is gonna be autonomous agents using tools like Relevance and Relay and N8n and level five is going to be vibe [00:01:15] coding. And then we're gonna talk about what are the pros and cons, what is the right implementation for each and every one of them?

[00:01:20] How far should you push the envelope, and so on and so forth. The goal of this is to actually help you make a decision for specific projects and specific needs that you have [00:01:30] to understand which level is the right level for your project, what are the pros and cons, and how to pick the right approach for your needs.

[00:01:37] So let's start with level one. Level one is just using large language models. This is what most people know, right? So if you are using [00:01:45] ChatGPT Gemini, Claude Perplexity, any one of these tools, you already have access and you know what a large language model is.

[00:01:53] What can you use A large language model four?

[00:01:55] Well, a huge variety of things. These tools are really good at drafting new [00:02:00] content. This could be emails. This could be things for social media. This could be blog posts. They're really, really good at summarizing information. So you have a lot of data and you wanna summarize it. They're good at editing content on helping you get better results from the content that you're generating.

[00:02:13] Again, regardless of what content it is, [00:02:15] they're very good at research and q and a. So if you wanna find information, whether online or with your internal information, most of these tools now connect and plug into your Google Drive and SharePoint and so on, and can research this entire database and give you answer and summarizations on more or less, [00:02:30] any topic that you want.

[00:02:31] Obviously, they're very good at doing this through the internet with deep research, and they're very good at generating reports out of all of that. They're also really good at brainstorming if you just wanna balance ideas and develop them in a better way than you can do on your own, or you can also [00:02:45] add them into a team environment, just as another team contributor.

[00:02:47] They're very good at that as well. So there's many, many different things. In general, they are very good at providing answers based on information.

[00:02:55] So whether it's structured information like Excel files and [00:03:00] so on, that they're very good at analyzing because they can write Python code all the way through other types of information. They're very good at. Interacting and engaging with our information and being very helpful in the process.

[00:03:11] So the biggest pros are it's the [00:03:15] easiest and fastest start. All you need is access to one of these tools. Most of them have free versions and all you need to know is English. Now, to be fair, if you know how to prompt and if you have good prompting skills, which you can learn by taking courses just like our course, but any other course out there, [00:03:30] or just look for prompting best practices from the labs themselves.

[00:03:33] You can learn how to write English or whatever language you're prompting in better, but that's all you need to get started. There's really no technical skills required. There's no setup required. You just get, log into one of these tools [00:03:45] and you can just start using it and it's accessible and easy to use for more or less, everyone.

[00:03:50] There's obviously disadvantages. Disadvantage number one is there's very limited automation that you can do, meaning every time you wanna do something, you have to re-prompt it. You can save your prompts to save yourself a little bit of [00:04:00] time, but you're still redoing the work every time you need to do a piece of work.

[00:04:03] And if it's something you do in a repetitive way in your company, in your business, in your daily life, then you will have to do it again and again and again. It has no memory meaning. If you'll want to share different [00:04:15] aspects and different steps across different things that you're doing, you will have to manually either prompt it or copy and paste information across multiple sessions.

[00:04:22] This is not a hundred percent true for some of the tools. Cha g PT now has long term memory. I actually use it a lot. I actually find that I use cha GPT [00:04:30] more than other tools, mostly because of memory, not necessarily because it's the best tool out there, but it's still not everything. It's still not every aspect, and there's still no, no handshake between step one or the process to step two necessarily unless you invest in creating the [00:04:45] flow of information from between these steps.

[00:04:46] It depends on having a very good prompt, meaning the people who are better prompters will get better results, and the people who are gonna ask simple, basic questions or guidance in English are gonna get sometimes mediocre, [00:05:00] sometimes less than mediocre results, which then leads to you not using them because it's not providing the level of results that are good enough for your business needs.

[00:05:08] Uh, another big limitation is that they cannot use tools. What I mean by tools is either tools that you can develop for them [00:05:15] or the tools that you use, the software tools that you use today that they can connect to your email, they cannot connect to your Google Drive, they cannot connect to your C-R-M-E-R-P financial system and so on.

[00:05:25] And another big problem is that you are not always getting consistent results, especially [00:05:30] when you're going across different users in your department or in your company that are trying to do the same thing. Each and every one of them is using it a little differently and they're gonna get different results, which from a business perspective, lack of consistency is a very big deal.

[00:05:42] So these are the pros and cons, but how do you get started? Well, getting [00:05:45] started with these tools is very easy. You just sign up for any of these tools. 

[00:05:49] Gemini, Claude Che, PT are the three. Leading ones. If you live in the Google universe and you're already paying for Google Workspace, you have access to the higher level of Gemini without paying [00:06:00] anything extra.

[00:06:00] So that would be my first recommendation. And Gemini is a fantastic tool. But ChatGPT and Claude are also great. And there's many other tools that you can go to, especially open source tools from multiple places around the world that you can try as well. Such as Mistral, such as Quinn, such as Deep Seek. [00:06:15] But if you wanna stay within the mainstream, ChatGPT, Claude and Gemini are all great options.

[00:06:19] And I mentioned prompting. You can use different prompt templates. There are many people who share them on LinkedIn, on their websites, and so on and so forth. You can take courses, as I mentioned, like the course that I am [00:06:30] teaching, or like any other courses that will teach you how to prompt better. So this is how you get started.

[00:06:35] Now, if you want to take it to the next level and do things that are a little more advanced, you can start connecting these tools to other platforms. Many of them provide [00:06:45] in-house connectors. So if you go to ChatGPT or you go to Claude, you can actually turn different connectors on. So as I mentioned, you can connect it to the Google Universe, you can connect it to SharePoint, you can connect it if you are on Claude to many, many other tools.[00:07:00]

[00:07:00] A slightly higher level above that, which is mostly available from a simplicity perspective is Claude. So in Claude you can very easily connect MCPs. MCPs are external tools that can easily plug into Claude. And you do [00:07:15] this just by looking whether a tool has an MCP server. If it does, it usually gives you a few lines of code, which is 5, 6, 7 lines of code that you can copy and paste into the config file of Claude, which brings a huge amount of capability into Claude itself.

[00:07:29] This [00:07:30] only works with the desktop version. It doesn't work with the web browser version, but it still allows you to connect a lot of other new tools that you have and bring them into Claude to do a lot more sophisticated stuff.

[00:07:40] So these are large language models and this is what most people are using [00:07:45] today. The vast majority are probably using ChatGPT, but a lot of people are using the other versions as well. If you want to go back and learn a little more about prompting. You can go back to episode 52 of this podcast.

[00:07:55] It was called The Prompt Playbook, proven Strategies for Crafting Killer AI Prompts That [00:08:00] Achieve Results. Again, episode 52 of this podcast. 

[00:08:03] And this leads us to level two, which are custom Assistants. So custom Assistants. Most people know custom gpt. Some people know gems from Gemini, or projects from Claude or spaces on [00:08:15] Perplexity.

[00:08:15] But what these are, are basically repeatable, reusable sets of instructions and reference information that you can use again and again and again. The way you build them in charge GPT is you go to Custom gpt, you click on [00:08:30] Create New, and you can give. The name and the description and then the instructions to the tool.

[00:08:35] What this does, and again, it's the same exact thing with gems, and it's the same exact thing with projects in Claude. Once you give it these instructions, which is basically a prompt, you can [00:08:45] rerun it every time without having to write the instructions again. You can also attach additional reference information such as style guides and PDF documents and Excel files, and additional types of instructions and references that the assistant can [00:09:00] use in order to effectively do the work that it's doing.

[00:09:03] I'll give you a simple example that I use. Several times a week, I have a custom GPT that writes proposals for me. So the input that it is expecting to receive is emails and transcriptions of [00:09:15] conversations that I have with prospects. All I have to do is drop these things in there. I don't write a single word and I click go and it writes the proposal for me.

[00:09:23] How does it write the proposal? Well, there's a long detailed set of instructions that tell it what to look for in the [00:09:30] conversations and in the email communication I have with the prospects, it has as attachments, a standardized template that I use. The standardized template that I attach is the one that has.

[00:09:39] All the different offerings that I have, all the different pricing that I have, but also the instructions to tell it, what to pick [00:09:45] in what scenarios, depending on the conversation with the clients. It also has a PDF that has an explanation of all the different services that I provide, so it can elaborate on all the different aspects of what I do, depending on the need and depending on what the client said.

[00:09:57] So the input that I have to give it is drop in a new [00:10:00] conversation or a new communication via email and it writes a proposal. I have probably 25 to 35 different kinds of these custom things that I'm doing right now that drive immense value and save me a lot of time. And you can build [00:10:15] them, as I mentioned, in any of the main platforms and they work as well in all of them.

[00:10:19] There are several different benefits to doing it in ChatGPT, it supports Canvas, it knows your long-term memory. It now, as of the launch of GPT five, has voice support [00:10:30] as well. It can create images within those custom GPTs. And there's several other things that custom GPT do better than gems and projects in Claude, 

[00:10:38] but all of them are really, really good and really helpful when it comes to creating repeatable, consistent work. [00:10:45] I also use them to do analysis of data. So if you have an export out of your CRM or your ERP or your financial system that is in the same structure every single time, you can write the instructions of what analysis you need to do.

[00:10:56] You drop the file in there. You don't write a single word, and you get the same output [00:11:00] every time. So we kind of talk, what are these good for? These are good for anything that is repetitive, anything that it requires knowledge-based answers, just like large language models. But that has to happen again and again and again.

[00:11:11] And it's just a tool that exists within your existing chatbots that [00:11:15] allows you to do some basic levels of automations. So the biggest benefits, the pros are consistency. Private knowledge integration, so you can again, connect them to your knowledge base and attach information to them that will help them do the work that they need to do.

[00:11:28] As an example, a lot of people use [00:11:30] them to write on behalf of their company, so you can use the style guides and the tone of voice and examples of good blog posts or good posts in social media that you've done, and you will know how to mimic that when it's doing its writing and there's no [00:11:45] coding or any setup required.

[00:11:46] It lives within your AI standard tools, so there's no extra cost and there's no extra setup. What are the cons? What are the disadvantages of using these tools? Well, it's limited with the process automation that it can do. It requires to [00:12:00] know how to write the instructions properly and it requires you knowing what data to attach in order to make it an effective tool that can actually do the results that you're looking for consistently. The way I'm doing this is I'm explaining to the large language model. 

[00:12:14] What am I [00:12:15] trying to do, what I'm trying to achieve? What is the process? What is the format that goes in, what is the format that goes out?

[00:12:20] I ask it to ask me questions to clarify the exact process, and then I ask the large language models to write the instructions for the assistant. And it does it and it does [00:12:30] it well. And then usually I use it as is. Every now and then. I need to make small tweaks on my own. Over time, by the way, I always make tweaks to these things as I use them and I see little things that I need to upgrade or that can provide me additional value.

[00:12:41] I go in and I make changes to the instructions. But you can literally have [00:12:45] a large language model, write the instructions for you after you explain exactly what you're trying to do Now. Another disadvantage is that it's not using tools. Again, you cannot connect it to your ERP, to your CRM and so on. This is slowly changing.

[00:12:58] You can actually do this [00:13:00] in custom gpt. On the bottom of custom gpt, there's a segment that allows you to write code. You can use that code to connect an MCP server or to connect to an API or to connect to a third party tool in a different way. So there's a way to do it. It's just not very user friendly, and it's definitely not to the average [00:13:15] user.

[00:13:15] And because of that, it does not automate the processes outside of GPT. So even if you are connecting it to pool data from, let's say your CRM, you still cannot automate things within the CRM itself. 

[00:13:27] You can just use data from the CRM to do the [00:13:30] things that you're trying to do. So how do you get started with these tools?

[00:13:33] As I mentioned, go to Claude Projects or to Gemini Gems, or to CHA GPT, or to Custom GPTs in ChatGPT. But start with a regular conversation. Expand what you're trying to do. Ask you to create the instructions, and [00:13:45] then copy the instructions into your custom GPT or GEM or project and just start using it.

[00:13:49] Don't forget to attach relevant information, and you can ask the actual conversation what information should attach to make it more accurate with its answers. You can upload your brand [00:14:00] guidelines, your policies, your FAQs, documents, whatever it is that you need as references as part of getting started.

[00:14:05] As I mentioned, start with a conversation to turn it into a automation, and then at the end you need to test and refine it initially until it's [00:14:15] working. Okay. And then refine it over time as you learn more about how it works or changes to your operation and stuff like that.

[00:14:21] How do you use it in a more advanced way? Well, one, we already mentioned in step one, in level one, which is MCPs, these tools can connect to MCPs, which means they [00:14:30] can engage and interact and work with other tools.

[00:14:33] I am now using an MCP server in Claude that is connected to N8n, which is an automation tool that allows Claude to build the automations for me. Why? Because it's connected through the MCP server straight into my [00:14:45] N8n automation platform, and it can talk back and forth. It can learn how it works, it can find information about it, and it can actually make changes to the environment.

[00:14:53] So this is advanced number one. Advanced number two is ChatGPT allows you to use voice and it allows you to use data [00:15:00] analysis, meaning it knows how to write and use Python code to analyze data. So if you have data coming in from whatever source in your company, whether it's financial data, marketing data, sales, data, customer support data, et cetera, if it's in a standard format, you can drop it [00:15:15] in there and it can write code and analyze that information for you.

[00:15:18] ChatGPT, as I mentioned, also knows how to create images. So if you need to create images that are standardized on a regular basis based on a standard input, you can build a GPT for that. And the last thing that [00:15:30] I will mention is that ChatGPT has what is called a ChatGPT Assistant versus a custom GPT.

[00:15:35] So what is the difference? The difference is that custom GPTs live within the chat environment. As you're using your regular ChatGPT interface, you can click on a specific custom GPT [00:15:45] that you've created or that somebody else created and just chat in there and use it. And assistant is the same exact thing.

[00:15:51] It has the same kind of instructions, the same kind of connections, the same kind of attachments that you can attach to it to give it reference information, but it works through the [00:16:00] API. Now, those of you who don't know what that means, it basically means that a third party software can connect and use the assistant versus you having to type and copy and paste information in and out.

[00:16:09] Why is that helpful? It's helpful because that means you can build an assistant that knows how to do a [00:16:15] consistent kind of work with reference information, with a clear set of instructions and connect it to. Other processes, which will lead us into step three of building workflow automations. But before we dive into workflow automations, if you wanna learn more about [00:16:30] how to create custom gpt and how to use them, two other episodes you can go and look up.

[00:16:34] One is episode 1 75, that was called Stop Wasting Time, automate Repetitive Tasks with Custom gpt. And the other one is episode 1 24. From Data to [00:16:45] Insights, how can AI transform your customer? Understanding with

[00:16:48] Can transform from data to insight. How can AI transform your customer understanding? And that again, is episode 1 24.

[00:16:55] But as I mentioned, the next step is level three, which is workflow [00:17:00] automation. Workflow automation actually existed way before we had generative ai. 

[00:17:04] I've used Zapier for the first time, probably about 10 years ago. And so being able to move data from one software to the other existed for a while.

[00:17:12] So the two leading tools are probably Zapier and [00:17:15] make the geekier version of both of these that is getting a lot of traction and a lot of buzz in the past six months is N8n the letter N, the number eight, and another letter N, so N8n.

[00:17:25]  The differences between these tools make.com is probably the easiest to get started [00:17:30] with. It has a very simple to understand user interface, and you can very easily connect your applications and build stuff with them. Zapier is a close second when it comes to ease of use, and it's connected to a lot more tools than Make does, even though both of them probably cover most of the main [00:17:45] tools that people are using in the world.

[00:17:46] And N8n is probably the hardest learning curve, but the most flexible and capable out of the three. So you need to decide. If you're a very technical person, probably go and learn N8n. If you're not, you'll probably be good enough with Make and or [00:18:00] Zapier.

[00:18:00] But what do these tools actually do, and how does that relate to ai? So what these tools know how to do is they know how to grab information from one software and move it to other pieces of software. Let's take a very simplistic example. You have people connecting with you on social media, [00:18:15] and you want to grab their contacts and put them on your CRM, and then you want to use.

[00:18:19] Some of it from your CRM and put it in a CSV file or an Excel file because you want to use it to track your something. And so these tools know how to do it very, very easily. They know how [00:18:30] to grab data from one place and put it in a different place without you having to know how to write any code. You literally just log into your regular tools and you build an automation that can take the data from one place and move it to the other.

[00:18:41] So what does that have to do with ai? Well, what I just described has nothing to do with ai, but what [00:18:45] you can do with AI is you can build AI steps into it that can then do everything that AI is very good at doing. It can analyze information, it can summarize information, it can extrapolate or extrapolate information.

[00:18:57] It can do all these things that AI knows how to do as [00:19:00] part. Of the automation, which makes the automation significantly more sophisticated. Simple example, you can have an AI that will read every one of your emails, understand what it's about, and categorize it and put it in a different folder in your Outlook [00:19:15] or Gmail accounts.

[00:19:16] That is actually very easy to set up. So the data comes in from one of your application. AI does the analysis of what bucket this email belongs to, and then it knows how to move it in the platform, in the actual email [00:19:30] platform to the relevant category. Either tag it or move it to a folder or whatever it is needs to do.

[00:19:35] Also, same thing you could do with prioritization. Also the same thing you can say, every time that I use an email that is. Asking for a proposal. I also want to get a Slack message because that will get my [00:19:45] attention. So you can set that up that way. So it's gonna look, the AI is gonna do the analysis whether there is a request for a proposal.

[00:19:51] If there is, it will send you a message in Slack. So these are the kind of things these tools know how to do.

[00:19:57] One of the benefits, the benefits is that you can completely [00:20:00] automate standardized, repetitive tasks across multiple tools and not just within data, within ChatGPT or Gemini or Claude. So what are the pros? Huge efficiency gains, like stuff that right now you do or people in your company do that is just tedious, [00:20:15] can literally disappear.

[00:20:16] The other benefit is that it works 24 7, 365. It never rests. It never takes vacations. it never has sick days. So it will do the thing that you set it up to do all the time.

[00:20:27] It dramatically reduces manual errors [00:20:30] because there's no human in the loop doing anything. It's just doing its thing and as I mentioned, now that AI is involved, it opens a lot more possibilities on what it can do because it can understand the information that it's moving around and hence make decision, analyze, [00:20:45] summarize, et cetera.

[00:20:45] All the things that these tools did not know how to do before.

[00:20:48] What are the disadvantages or the cons of these tools? Well, it requires planning and mapping workflows, meaning you need to understand what is the actual workflow you wanna build. You need to understand what is the process in real life that [00:21:00] people are doing that you want to replicate. You need to document that and you need to set it up.

[00:21:04] You also need subscription to these tools, meaning it's not within your ChatGPT. You need to pay for make or Zapier. To be fair, if you are using N8n, again, the geek here, more advanced tool, [00:21:15] you can host it yourself and then it still costs you money, but a lot less money. 'cause instead of paying the cost of automations, you are just paying for hosting.

[00:21:23] I'm paying $6 a month to run as many automations as I want, which is completely negligible. 

[00:21:28] But in general, you usually [00:21:30] need additional subscriptions. Also, if you're using AI as part of the process, then you're going to pay for the tokens through the AI API.

[00:21:38] So the way this works is you decide which model you want to use, whether it's from Claude, from chat, from Gemini, from Grok, from [00:21:45] whatever. You're gonna go to that tool and you're gonna request what is called an API key. And you get the API key, which just , gives you access to your ChatGPT, API.

[00:21:56] When I say your, you're gonna put in your credit card and it's gonna charge you for every [00:22:00] token that it's using. Basically any word that you're gonna send to it, and any word that comes out, or any image that comes out, or whatever the output is, you're gonna pay money for it. 

[00:22:06] The prices are very, very low and they vary between models, but even the expensive models are running around $20 [00:22:15] for a million words, which is still very, very little and significantly less than you would pay a person to do the same amount of work. But many of the models run in cents per millions of tokens, which means the price is per which means the price is practically negligible. [00:22:30] Now to be fair, these tools, even the basic one like make, require some technical skills.

[00:22:35] If you are afraid of anything technical, then maybe it might be a little too complicated for you. But the reality is make, and even Zapier are actually really easy to learn and you can very, very quickly, [00:22:45] especially if you're using templates that exist and there's many templates, you can get up and running fairly easy.

[00:22:49] And if you wanna build more advanced stuff, there is a learning curve, which will probably mean you're gonna start with a template, try to tweak it. It's going to break. You're gonna go back and forth. You're gonna watch a bunch of YouTube [00:23:00] videos, take a course, whatever it is that you're going to do. But in the long run, it will make sense because it's gonna save you a lot of time on repetitive tasks.

[00:23:06] So even if you spend a week learning how to build the automation, and now the automation does something that you do every single day, it makes perfect sense.

[00:23:12] How do you get started? Well, you identify the [00:23:15] repetitive tasks and processes that you have in the company, then you pick the right tool, either Zapier or Make or N8n or other tools.

[00:23:22] There are several different others in that category. As I mentioned, the easiest entry point, I think from an ease of use perspective is make.com.

[00:23:27] You start with one simple [00:23:30] workflow. So take something simple that you wanna learn how to do, go and look online for somebody who already built it.

[00:23:34] And if it's something simple with mainstream tools, many people are already sharing exactly how to do it, including the templates. And you can find YouTube videos and blog posts and repositories that, give you these examples and [00:23:45] give you access to them. And then you can get up and running very, very quickly.

[00:23:48] Learn from these and then you can start building your own automations.

[00:23:51] Now if you are looking for the next advanced version, it connects to how we ended up the previous level when I taught you about the difference between an assistant and a [00:24:00] custom GPT. So an assistant is basically a set of instructions that happen within ChatGPT in this particular case that will repetitively do the same thing again and again and again.

[00:24:09] Which means if you want to have AI be a part of your automation process [00:24:15] in Make or in Zapier or in N8n you can take the standard thing that you want to transfer into the AI and send it to an assistant. The benefits of doing that is that the instructions are the same every time, but I know what you're thinking.

[00:24:27] You're thinking, well, I can just call the AI and [00:24:30] send it the instructions every single time and still send the same exec instructions. And that is correct, but that means that you're paying for tokens for sending all these instructions. So if your instructions are long and detailed, and if you have several reference documents, A, the process is going to take longer [00:24:45] because every single time it has to send that information to the ai, including your attachments.

[00:24:49] And B, you're going to be paying tokens. You're gonna be paying money for every word in your instructions, and for every word in the attached documents that you're going to be uploading.

[00:24:56] However, if using an assistant, all that goes away, all that [00:25:00] information lives within the assistant that is already uploaded once to ChatGPT.

[00:25:04] And then you can use it again and again. Again, by only sending it the input information that it is expecting.

[00:25:09] So if we'll piggyback on the example I gave earlier.

[00:25:12] Let's say I want to build a proposal every [00:25:15] time a specific type of meeting comes in, I can have the automation intercept every time this kind of meeting happens on my calendar, or every time there's a transcription coming from Fathom, which is the tool that I use to take notes, and then it's going to.

[00:25:28] Take that information, [00:25:30] strip just the transcription, send it to an assistant that already has my template, that already has the information about all the different offerings that I have as far as descriptions, and that already has a clear set of instructions of what to look for in the conversation in order to know what to offer in the [00:25:45] proposal.

[00:25:45] And it's gonna send the proposal back as the output of the assistant and then back in the automation, I can decide what I wanna do with it. I can save it as a Google doc, I can save it on SharePoint. I can send it as an email. I can attach it to a Slack message. I can do whatever I want [00:26:00] in the automation once I have the output from the assistant.

[00:26:02] So the combination of these two things becomes extremely powerful.

[00:26:06] Before we move to level four, if you want to know a little more about these kind of tools, you can go and check. Episode 80 of this podcast is called AI No-Code [00:26:15] Business Efficiency Revolution with Reid Robinson, the head of AI product at Zapier.

[00:26:20] And episode one of three, AI Productivity Boost Automate Workflows with AI in Zapier Central with Valeria Kovich. So two great episodes if you wanna learn about Zapier, but if [00:26:30] you do learn about Zapier, it is very similar if you are using make or N8n. So these will give, you ideas of what you can actually do.

[00:26:37] And now to level four autonomous agents. So everybody's now building agents, or at least everybody is saying they're building agents. The [00:26:45] reality is most or a lot of what people call agents today are actually what we did in level three, meaning a step-by-step process that has some AI in it that will achieve a specific goal, which is perfectly fine.

[00:26:55] You can call it agents, it doesn't really matter. 

[00:26:57] But the biggest difference between this and what [00:27:00] real age Agentic work is, is that age agentic work has several different difference.

[00:27:05] So let's define for a minute, what's the difference between a process that uses AI and an agentic process? So, an agentic process has several unique [00:27:15] characteristics. First of all, it has autonomy, meaning it can make decision and it can chart its own steps to achieve a specific goal. You define the goal, you give it access to tools, which we're gonna talk about in a minute, and it will.

[00:27:27] Define for itself what [00:27:30] is the best path to achieve that goal. The second thing, as I already mentioned, it can use tools, it can have access to different tools and it can use these tools in order to achieve its goals. And the third thing, it can have access to memory. You can attach it to different databases where it can remember, [00:27:45] add information that it can pull afterwards for other steps to be used.

[00:27:48] And you can even use agents together with other agents in order to build a team that will work together and they can use each their own memory or a unified memory for all of them together, where they can share information, which [00:28:00] makes it even more powerful.

[00:28:01] So what are some tools that you can use in this category or in this level?

[00:28:05] You can use tools like relevance, ai, like Relay, and even N8n that we mentioned in the previous steps has agents, has very powerful agent capabilities built into it. [00:28:15] That then allows you to mix and match between the old school workflow automation together with advanced AI agentic capabilities.

[00:28:21] So what are these agents best for? Well, they're best for managing tasks end to end. So basically more complex, sophisticated [00:28:30] processes or things that don't necessarily have a linear path to success. Anything that has a linear path to success, a step-by-step automation process will actually do better. Why?

[00:28:40] Because by definition, when you give it autonomy, it means it's not always going to do exactly the same [00:28:45] thing. That's the whole point in autonomy. And that means that every now and then it's gonna do a subpar job in one aspect of the process. So if you do have a clear step-by-step process, use a tool like Make or Zapier and N8n.

[00:28:57] But if you have something that requires more thinking and more [00:29:00] flexibility and more autonomy, build an agent now. The other thing it's really good at is handling multiple aspects of things together that needs to be evaluated. So you have data from three or four different sources, and you need to decide what are the implications or what's the right output based on all that information.

[00:29:14] [00:29:15] Plus some background information that is not a step by step process. That is a thinking process, an evaluation process, a decision making process that agents do very well. 

[00:29:23] And then the last one is complex workflows with minimal human inputs. Stuff where usually they would be human [00:29:30] inputs, such as picking the best content or making a decision about stuff in a proposal, and so on and so forth.

[00:29:37] In these cases, agents that are built for these particular tasks will do a better job. And as I mentioned earlier, you can even build a few of these together to make a more [00:29:45] advanced solution. So what are the benefits? What are the pros of this? 

[00:29:48] iT can run really large, complex processes.

[00:29:51] It frees up significant human time because it can replace more advanced tasks. It can work in multi-level. You can build several different agents. Each does a [00:30:00] step just like you would build a team in your company. So if we take marketing as an example, one can be the coordinator, one can be in charge of text, one can be in charge of images, one can be in charge of social media reading and responding.

[00:30:12] One can be in charge of checking what are the [00:30:15] current hottest trends and so on, and they can all work together to generate well performing content for you. And as we mentioned, they can connect to tools, so they can use a huge variety of tools that are either existing tools that you have right now, or tools that you build.

[00:30:28] So in most of these [00:30:30] platforms, you can build and define tools that then these agents can use, which allows them to do stuff in your real universe. Other than just providing written answers, like the first basic two levels that we talked about before.

[00:30:42] What are the cons? Well, there's a risk, right? [00:30:45] Because you're giving it autonomy, it may not get it right every single time.

[00:30:48] So there's a level of risk involved. You can obviously reduce it if you know what you're doing, but there's still a level of risk because you're giving it autonomy. How do you solve that risk or how do you reduce that risk? You set up a man in the loop [00:31:00] checkpoint or checkpoints in the process. So the agents do their thing, they provide an output, a human will review the output, approve the output, and only then the agent continue to do his work, which leads to the second con, which is it requires governance and monitoring, which means it requires [00:31:15] time and investment from the human perspective, it's done. When done correctly, it's obviously significantly lower investment than doing the thing with a human to begin with, but it still requires people to be involved in the process to review and evaluate.

[00:31:27] It requires way more advanced technical skills [00:31:30] in most cases. So there's a learning curve that you have to go either on your own or with your team or somebody you hire, and usually it takes longer to develop than just simple automations. But there are a lot of benefits as we mentioned earlier. So how do you get started?

[00:31:44] [00:31:45] First, you need to define the scope and the rules of what you're trying to achieve. Then you want to start with a human in the loop to test every single step in the beginning of what the agent is doing and slowly reduce the amount of steps and the amount of human evaluation that is required as you build [00:32:00] trust with those specific steps.

[00:32:01] And from tools perspective, as I mentioned, you can use relevance, you can use relay, you can use different things. You can use N8n, which is how I build many of my agents.

[00:32:10] If you wanna learn more about that, you can go to episode 1 [00:32:15] 96. How to Safely Run Powerful AI Agents like Manus and Spar With No Risk, or Episode 1 79, beyond the Buzz, how to create AI Agents that provide real business value. There are many other Agentic episodes that we've done recently, so [00:32:30] just scroll back.

[00:32:30] In the last 10 episodes, there are probably three or four about agents, so you can look for those as well.

[00:32:35] And the latest one that I wanna mention that actually got a huge buzz is episode 210, that is called Your Next hire is an AI agent, building agent teams that actually work. [00:32:45] So go check out these episodes. They will give you a much deeper sense of what is the difference between an agent and just an automation, 

[00:32:51] which takes us to level five AI infused applications using vibe coding tools.

[00:32:57] So what are these tools? These tools allow you to write [00:33:00] your requirements in simple English and some of them in other languages as well. And they will write code and in some of these tools also get the databases and deploy the code and do everything you need in order to have a working application.

[00:33:12] What are they good for? Where you can create any [00:33:15] application you want, including AI part applications, you can launch new products, whether internal products for things that you need in your company or external products that can make you money. You can integrate it with your existing solutions because these tools are very good at building integrations [00:33:30] through your APIs or MCP servers and so on.

[00:33:32] So you can build integrations to whatever tools that you want, and you can have a fully functional software customized to your unique needs. So you heard me say on this podcast many times before, I think the concept of a app store [00:33:45] will die in the next few years, and the concept of SaaS will be challenged because right now most of the applications that we have on our phones or on our desktops do 57 different things when we need them to do three.

[00:33:56] And they don't always do those three in the optimal way for us. Well, now you'll be able to [00:34:00] create those applications on your own without having the need for a development team or paying money for that or, or anything like that. And so for 20 bucks a month, you can build an application and then use it every time you need to use it, and it's gonna be tailored to exactly your needs.

[00:34:12] So that's the first pro. Let's talk about some additional [00:34:15] pros. It can be a competitive differentiator. Again, whether internal or external tools can allow you to do things that your competitors are not doing right now, or offer different services or capabilities that your competitors are not offering right now because they do not know that they can create applications.

[00:34:29] [00:34:30] It can use databases, so none of the other tools knows how to use databases, at least not effectively. So if you want to save information to a database that can be retrieved afterwards to be used either by the agents themselves or by other software, or by other components [00:34:45] that needs to connect to it, well, this is your best option.

[00:34:47] It can include login and permissions and payment capabilities. All these things usually do not exist in the previous levels that we talked about. So if you're trying to build anything that is a little more robust, that has a database, that has a login, that has permission [00:35:00] levels, that has a payment mechanism attached to it, all these things only come if you use Vibe coding tools.

[00:35:06] And the biggest pro, as I said before, it can generate. Revenue. It can generate leads, it can generate things for your company that right now you're doing [00:35:15] manually, most likely, if you're doing at all the cons, it's the highest cost and complexity. So these tools cost you money. The deployment cost you money.

[00:35:22] There's a steeper learning curve, obviously. 

[00:35:24] The cost. There's the cost of the Vibe coding platform itself. There's the cost of [00:35:30] tokens to whatever AI tool it is using in the backend. And there's the cost of whatever other thing you're trying to do. So if you're using a third party database, you will have to pay for the database and obviously the data that you're storing there and so on.

[00:35:41] So from a pricing perspective, it's more expensive than the [00:35:45] previous options and levels that we talked about. How do you get started? Well, first of all, you need to understand a business problem that an application can solve. Then you wanna choose the right builder. Uh, the tools that are probably the most common today are lovable and bolt and rep and base [00:36:00] 44.

[00:36:00] So these are the tools that you can use. Or you can go to the more advanced development tools like Cursor and GitHub copilot that are probably a little more advanced and are gonna be more helpful to you if you actually know how to write code and understand code.

[00:36:13] Then to get started, you wanna [00:36:15] consult with a large language models.

[00:36:16] Going back to level one, explain your problem, explain what you're trying to build, and ask it to write the initial requirements document for you. If you wanna get even fancier, you can build a custom GPT or a cloud project that already has best practices on how to write [00:36:30] software requirements that you can download from the internet, and you can even have the AI research for you, and then it will write better requirements for you To do that.

[00:36:37] Start with a simple MVP and then build more and more functionality as you go on. 

[00:36:44] If [00:36:45] you want to learn more about Vibe coding tools, then go and check out episode 2 0 2 of this podcast. It's called The Ultimate AI Showdown, comparing Vibe Coding Tools with Mark Khe, and it's a great episode for you to understand what these tools can do.

[00:36:59] So that covers [00:37:00] the five levels of AI tools you can use to implement new solutions. So quick recap. 

[00:37:04] Level one is using just a large language model like ChatGPT, or Claude or Gemini and so on.

[00:37:10] Level two is creating custom Assistants with custom GPTs in [00:37:15] ChatGPT project in Claude, Gemini Gems, et cetera. Level three is workflow automation with tools like Make Zapier or N eight. N. Level four is autonomous agents where you can use tools like relevance or relay, or N8n again, and [00:37:30] level five is vibe coding of writing and creating complete applications.

[00:37:34] Now the trick is how to find what is the right level for your implementation. Well, there's a bunch of questions that you need to ask yourself in order to figure it out. The first one is, what is your primary goal? [00:37:45] Is it something as simple as getting answers from data? Then one of the first two levels is good enough.

[00:37:50] The second one is how urgent you need it. Do you need it right now? Well, then you're not going to develop a process or an automation, or definitely not an application to do that, and then you are stuck with the [00:38:00] first two levels.

[00:38:00] And then you should use one of the first two levels, which are all good.

[00:38:03] Like there's nothing wrong with it. It's the lowest level of entry and the least amount of effort.

[00:38:07] The third question you need to ask yourself is, what is your or your team's technical skills? Like, can you even use a vibe, coding tools? Can [00:38:15] you use a tool like entertainment at this point?

[00:38:16] If not, then how much are you willing to invest from a resources time and money perspective to teach yourself and or your team all these other tools so you can get the additional benefits? The next question needed to ask is what tools or platform you already have and that you're paying for. [00:38:30] Maybe you have a Google license and so you already have the first two levels.

[00:38:34] Maybe somebody in your organization already has a license to make.com and then you can use make.com and so on and so forth. So figuring it out, uh, is a big deal on the same topic 

[00:38:43] which things you want to connect to. [00:38:45] So what third party platform you want to connect to, and which of the tools.

[00:38:48] Have an easy connection to the platforms that you're already using. So look at your ERM, look at your CRM, your ERP, your marketing platform, and so on, and see if they connect to Zapier or make or N8n [00:39:00] natively. And if they do, that will allow you to guide the conversation on where to go next.

[00:39:03] The next question is, how sensitive is your data? Is it data that you can freely share with anywhere? Is it data that is internal to your company? Is it private information that you cannot share? Are you in a [00:39:15] regulated industry like finance or legal and so on? So this will also define what kind of tools you can and cannot use and how you can use them.

[00:39:22] That's another, by the way, big benefit of N eight N over the other automation tools. You can self-host N8n because it's an open source [00:39:30] platform, and if you do that, then the data stays within the box in which you're hosting the platform, which allows you to run more secure data than you can run, let's say, on Zapier or make, because then the data goes to their servers.

[00:39:43] Question number six is how much autonomy [00:39:45] you need, and also how much autonomy are you comfortable with, right? So you may need auto autonomy, but you're saying, well, I want the same exact outcome every single time. Well, then you cannot have autonomy. So you need to define that balance for each particular use case, and then [00:40:00] make the decision based on that.

[00:40:01] That will allow you to decide between level three and level four, basically, between workflow automation and autonomous agents. The next question you need to ask is what is your budget? And either way, most of these tools are relatively easy, [00:40:15] but as you start building more and more of these capabilities.

[00:40:18] You will need more other tools to connect to it. So as an example, if you're using Vibe coding, there's a decent chance you will also use automation in the backend one way or another, and you will use tokens from at least one [00:40:30] large language model provider. So all these things starts adding up. So you need to identify your budget.

[00:40:34] And again, the budget also need to take into consideration if you're developing an application, where are you going to deploy it? What kind of databases is it going to use? Where are you going to host the application? All these kind of things. And from a training perspective, [00:40:45] how much is it going to cost you both in time and money in training yourself and or your team to use the tools that you need to learn in order to do these things.

[00:40:53] Question number eight is, do you need quick wins or long-term transformation? And that will also define which of the [00:41:00] options is probably most relevant for your particular use case. Then the subset question of this is, can you start small in scale from there? Meaning, can you start with an MVP with a simple version of the solution?

[00:41:10] Test it out with just a large language model, or even just a simple automation in a [00:41:15] custom GPT, and then if it works, figure out all the bells and whistles that you really wanna see in order to deploy it across the entire organization or for your clients and so on. And then the final thing is more of an HR question.

[00:41:25] Who is going to monitor and maintain that solution? As things [00:41:30] evolve over time with the answers to all these questions, it will allow you to guide yourself in the right direction. Now, as I was going and writing these questions and thinking about this process, I actually thought it could be a good idea to give you a tool that will do it for you, that will ask you the questions and that will be able [00:41:45] to help you then make a decision on what is the right solution for you based on the use case, based on your experience, and based on all these things.

[00:41:53] And I literally followed steps. And what I did that made it really, really fun and interesting and relevant for this episode is I [00:42:00] followed all five levels. I actually started with a simple chat with ChatGPT to identify what questions should I ask and what the tool should do. And then I built a custom GPT that actually answers these questions, and then I decided to go to the next level because it was limited with the [00:42:15] things it could do.

[00:42:15] And then I've actually built a simple automation, and I skipped the agent side because there's no much agentic work here to do. But then I went all the way to build an actual application that does it. So those of you who are watching this episode on YouTube, [00:42:30] I will share the screen with you and we'll show how it looks like.

[00:42:34] So it looks like this. It's called the AI Solution Advisor, and you see an output of what it was doing. So you can see for the latest input I gave to it, it told me that the right setup is AI agents. It shows me the five levels [00:42:45] and the level four. In this particular case that was selected, why was this chosen for you?

[00:42:49] The user has advanced experience with large language models and custom gpt, mid-level experience with workflow automation, blah, blah, blah, blah, blah. Um, and the goal involves [00:43:00] automating a complex multi-step process involving data retrieval analysis and personalized communication, which aligns well with AI agents.

[00:43:06] So how does all this thing happen? And then it gives me different options of the tools that I use and an implementation plan, and even gives me the basic prompts [00:43:15] that I can use in order to set up all the different aspects that I need to use for it. So how does this thing work if I. Refresh it, you'll be able to see how it looks like.

[00:43:24] It literally is a bunch of questions with a progress bar. And the first question is, what is your role? Are you a founder at [00:43:30] operations in marketing? Are you a technical person? So you're gonna answer that. How big is your organization? So it kind of knows what the scale of this thing needs to support, and then what is your primary goal, and so on and so forth.

[00:43:39] So it's gonna ask me a bunch of questions, takes about two minutes to fill out everything, including some open-ended questions [00:43:45] that allows me to understand your project and what you're trying to do, as well as your experience in a more detailed way. And then it makes recommendations, and even as I mentioned, gives you the initial prompts that you need to use, recommendation about the specific tools you need to use and so on.

[00:43:58] Now, I'm not really seeing this tool yet [00:44:00] because it's. 85% ready. But what I'm going to do, I want to record a separate episode that will walk you through the process of how I developed this, how I started with a simple prompt on Gemini, how I moved it to ChatGPT, how I asked ChatGPT to create the [00:44:15] initial version of this, how I moved it to Claude to develop it further from there, and how I took it from Claude to Replic to build an actual application that I can deploy and give you access to.

[00:44:23] And by the time I record that additional episode, which will probably be in two weeks, this application will be ready and then I'll be able to share it to [00:44:30] you and you'll be able to use it from that moment on. Two, put in your inputs and get an answer of which of the five levels or maybe a combination of them you should probably use for your project.

[00:44:40] That's it for today. I hope you found this episode helpful. I know it helped me a lot [00:44:45] thinking through the process, so hopefully it enables you to think through what AI solution you need in a much more clear and concise way. If you are still confused or if you have any comments or any questions about this, please feel free to reach out to me on LinkedIn or just come and [00:45:00] join our Friday AI hangouts, which we do every Friday at 1:00 PM Eastern, and it's completely free.

[00:45:05] It's open to the public, and we just get together and talk about AI and we can talk about this or about anything else that you're interested in.

[00:45:11] Have an amazing rest of your day and your week. [00:45:15] Keep on exploring ai, keep sharing it with other people because the joint benefit of this is that we're all gonna learn from one another and have an awesome rest of your day. 

[00:45:23] ​