179 | Beyond the Buzz: How to create AI Agents that provide Real Business value with Tom Winter

23:46So AI agents, my definition, are AI-powered problem-solving teams. So if you're thinking about the team, marketing team, let's say, it's everybody in the team has different skills. And if you connect them together, they can solve a problem together. But if you give an open-ended question to them, they will figure out first how they should handle it. So who takes what kind of a part of the task and how to, they will talk to each other to resolve the problem. But the problem is the path of resolving the problem is kind of undefined in the beginning. So they, as a team, they have to find a solution, how to resolve the problem.

24:20Okay, so the big thing, like what I would use mainly AI agents for, as an example, everything that is connected to analyzing data. Because if we want to see, for example, a specific trend in our data, we don't know in the beginning what we're looking for, as an example. So an agent would help us out to define what kind of things we can look for and then execute on that to find a specific trend. So this is an example. I will give you more examples here on. So the strengths, like the key strengths is select tools that you need. They can select tools that they need. So if you give them a set of tools that they can use, it doesn't mean that they will use all of them.

24:56They will use a specific tool for a specific task or a specific combination of tools like in steps to resolve a specific problem. So like, for example, if you have a data set somewhere and you want to find a trend, it will create a SQL query, ask the database, come back with an answer. On top of that, they will create another SQL query to come back like with additional answer, then use, for example, data manipulation in Python to check and combine these two data tables, then create a prediction model for the future to predict something in the future. But it will design what it needs, like what kind of steps it needs to resolve a specific problem.

25:34It simulates human-like process, which is a second thing. So it's kind of like it's similar to how we resolve problems. We gather tools that we need to solve a problem and we define the path. can handle uncertainty and make decisions, which is like, in my opinion, the biggest strength and weakness of AI agents because it will create a solution, but at the same time, it's unpredictable, fully unpredictable, what kind of a solution it is. We love them for creativity. We hate them for creativity. So focuses on achieving goals without rigid workflows. And this is like really important. From the developer's perspective, it was the biggest problem that I had in my mindset because my developer mind is an if mind.

26:18So basically everything that I'm, if I'm approaching any task, I'm already thinking about all the possible edge cases. So if this happens, then do that. If this happens, then do that. Like my code, like a couple of years ago was like full of ifs. Like, so there are edge cases that we had to think about to solve a specific problem. In AI agents, we're thinking about the goal, the end goal, and we're letting AI think about how to solve the problem and how to avoid hitting a specific edge case. So you have to change the mindset from if mindset to goal-oriented mindset, which was really extremely hard for me.

26:55So limitations of that, you can produce unpredictable results. So it might not be repeatable, which is a problem for me when AI writing, I don't like to use AI agents for everything. It's a lot more complex to build and deploy, especially there's not many solutions on the market that can truly create an AI agent. There's like a lot of wannabe AI agents, which maybe they are even, but it's really difficult to deploy. And it's very difficult to debug and refine because you don't know what's happening underneath. This is basically the process of thinking of AI in a black box. You don't know exactly what's happening.

27:36It's not so easy to debug it as workflows. So some sample use cases, AI sales agents, AI research assistants, AI content brief generator, AI onboarding specialist. I will tell you one thing that I heard from a customer, which was like for me, extremely funny. For them, I don't think so. So they created an AI agent for support, which, and they used it for e-commerce. So some, one customer ordered 80 pieces of something and they received 60, only 60. So they came back to chat with, was the support agent. And they asked and told them like, look, I have this order. This is the number of the order.

28:12I ordered 80 pieces of something. I got only 60. And the agent actually resolved the problem. So then the customer saw that he had 20 boxes of these pieces of things, like in each box, I think 100 pieces each, because it resolved the problem. So it created an order of 20 boxes instead of 20 pieces of something, and it sent it there. There was another one that told me that they created an agent for support and the charity started to talk with them, with the agent. And they asked like, can you pay to our charity£10 ,000? And the agent said like, sure, like, let's do it. So this is the creativity of the agent that you have to look out for.

28:55So sometimes it will try just to resolve a problem. It will find a way to resolve it. The way to resolve the ask about the charity funding, like was basically to pay them 10 000 pounds they didn't pay them in the end they agreed on paying half of it but since the agent said it out loud they had to follow through so there are some sample platforms to use so relay app and it and relevance bubble mind studio probably a lot more uh isar do you know any that you can recommend yeah i think relevance is probably the one i know most people that are using people who don't necessarily write code. Yeah, there's a bunch of them.

29:34Relay is another one. Mine's too. Yeah, you got all the ones that I know. And like you said, NA10 now has some agent capabilities built into the product. I'm guessing, I don't know that, but I'm guessing all the existing automation tools, so NA10 makes Zapier and all of those, will start building real agent capabilities into the process because it just makes sense. Yeah, I'm currently trying out Vertex AI. so directly in Google, Vertex AI agents. It sounds really, really interesting what they have there, but I don't have an opinion yet. Yeah, and I'll say something. The agent development world, not really, like any development world, I would guess, but the agent development world really divides into two separate universes with obviously a lot of blurry stuff in the middle.

30:19Some of it is for developers, people who are actually computer engineers who know what they're doing and can write and understand code. And some are no code or low code solutions where you don't need to know how to write code. I think the best outcome today is a combination of the two because there are things that you can do very efficiently with the no-code, low-code stuff, and there are things that you can do significantly faster if you know how to write code, especially now with more and more... Are you going to touch MCP at all or not really? No, no, no. Okay, so there's more and more infrastructure being built in the world to support agents that just makes it easier to develop without actually developing everything.

30:54I'll keep it simple as that. Yeah, I'm the one that is developing it on my own, like basically in code, like, because I know exactly what I want. And I tried many solutions, but I didn't find things that I wanted. Okay, so like top things that I learned, AI agents solve problems, not just automate tasks. This is the main thing. Strong technical skills are needed when you're developing it, like at least stronger than like when it comes to a workflows. And you have to shift your mindset, like from the finding edge cases to achieving goals. And I know it sounds simple, but it's extremely difficult.

31:28Like really, you will see that like once you do it, like once you start doing it, you will catch yourself about thinking about edge cases. And that's like really, really difficult to change. So what we do, I want to show you an example how we approach like AA workflows. And I will show you my AA workflow. This is a little bit simplified workflow. Sorry like that. It's so small, but I will go through every single step here. So don't worry if you don't see it. This is an AI workflow, simplified AI workflow, combined with AI agents and human interactions that we use at SEOwind to write articles. Like, this is how we approach it.

32:00For those of you who are listening, we're looking at a screen with way too many steps as a flowchart. But again, Tom is going to walk us through the steps. So you're not really missing anything. We're going to describe every single step on the screen. No, no, no. I will go through it. Yeah. Exactly. So, like, when it's green, that means, like, there's a human intervention. There's, like, agents inside that make specific decisions. So like it's a hybrid approach that takes all of the things that are really amazing from each of the world. But like for me, because I need predictability, it's still an AI workflow.

32:28There are places where I can add an AI agent to solve a specific problem, but then I'm going back to going with the workflow. OK, so let's start with the first thing. So as you can see right now, you can actually see. So when I'm writing an article, I'm starting with the focus keyword. So I'm going into Google and checking who ranks for there, there, and I'm collecting data. So like I'm collecting data from Google, Quora, Reddit, and other stuff. I'm going to search results, scraping content from my competitors, and gathering all the information about SEO and anything else. I'm using this to create a brief because I need data.

32:59Like always when I'm thinking about AI, I need data. I'm using AI as a logical engine, not a search engine. So I'm trying to give it every single time that it has to make a decision. I'm trying to give it as much data, contextual data, as possible. not as much because I don't want to pour in and overwhelm it with data. I want to specifically give contextual data to AI. So for example, if I'm creating an outline, I'm providing it with all the outlines of our competition that nailed the search intent because they're in top 10 search results so it knows what to focus on. I'm not asking it, write me an outline for this and that because unfortunately it will answer.

33:34It will answer with some BS total because it will hallucinate. It will try to resolve my problem, but it doesn't have the knowledge to resolve my problem. So basically, this was the step one. Then I'm creating the brief. So if I have all the data, as I said, I'm using SERP data. This is inside the system, but basically underneath, we're gathering all this data. So this is SERP data from like all the specific competitors. I have questions to answer from people. Also ask Quora, Reddit. I gather keywords from different places. What I do, for example, with keywords, I cluster them. Yeah, I'll pause you one thing.

34:08People don't know what SERP is. So SERP is an SEO term that stands for Search Engine Results Page. So basically, when you go to Google and you type something and you get the ranking of different websites, that's what SERP stands for. And people like Tom, who do this every day, thinks that everybody knows. Agreed, agreed. Sorry for that. But that's what it means. It's basically the ranking of the search results and what you see on the search page. Yeah, and then I'm getting keywords. Keywords means like, what are people putting in Google to find these pages? So we can actually, from SEO tools, we can find such keywords and we can get actually 200, 300 keywords that people are looking, like putting into Google to find all these results from our competitors.

34:48What I can do from my perspective, I can, for example, cluster the keywords because LLMs are large language models. So instead of having a pile of 200 keywords, I can cluster them into different groups. So I understand how they work with each other. And then I have like 10 groups of keywords, which for me as an SEO expert, it's much, much easier than seeing a pile of keywords. So one thing that you have to remember from that, LMs are really good with clustering things. So if you have anything or like finding specific knowledge inside what you're looking through. So if you have 100 pages of a report, you can ask AI, you can use notebook LM, for example, and you can retrieve the data that you want because you want to focus on a specific part of what you're seeing.

35:30I don't want to overwhelm AI with too much knowledge. I want it to look at specific things that I want to focus it on. so as an example to whenever i'm like writing a prompt i'm starting with defining exactly the system message what it is so this is an example uh i'm defining when i'm asking for creating an outline for my article i'm defining who ai is so basically kind of role playing telling ai like what it what it has to do at the same time telling it who it is then i'm defining the task itself so exactly all the information in the task. And as I said, like earlier, I'm trying to get an answer in JSON file, which is a structured JSON.

36:14Really important, be precise. So think of AI as you would think about another human. What kind of information do you need to deliver to another human so they can resolve your problem? Like AI will not read your mind. This is really a problem. Just for the people who are listening and not watching, can you go back to slides for a second? And I just want to give people an example. So when we're saying one more, when we're saying give the AI a role is you are a senior SEO strategist and content articles. And there's like a whole paragraph of that defining the role, just like you would hire a person for that task.

36:47And then on the second step, if you go to the second slide, it's your task is to create a detailed article outline that aligns with the user intent and is optimized for SEO, follow these rules. And then there's like seven that we can see potentially more steps on how to do that, that tell it where to put information from all the different things Tom's mentioned earlier, what it needs to do step-by-step. So you're basically helping it follow a process. And as Tom mentioned, you need to know the process because you need to think about hiring an intern and telling them how to do this. This is what you need to do for AI as well.

37:20And remember that AI doesn't know where you took the data from. So you actually can tell it like, look, I took the data from this and this. This is how I found the data. This is how I structure it. The cool part that I like to use, also, I try to explain what will I use the output for. Kind of like telling it why I need it and how I will use it. Sometimes it helps actually to AI to understand exactly what to do. Because if I just put it in the middle of the task and don't give all the information that I would normally give to a human that will resolve the problem, then it has to actually be creative.

37:54And I don't like AI to be when it's creative too much, at least. So remember relevant context and research. Describe the output. So like, what do you want as an output? And maybe, for example, like what will you use the output for? So if you're using, if you're doing an AI workflow, I'm often telling like, look, I will use the output in JSON file. That's why I need to parse it. And then I will follow up with the next step in the process. That's why I need it in this form. Another thing that helps a lot when you do these is examples. so if you give it what good looks like and good could be here's the template i want you to use and then you know what template this is going to use or here's a good outcome based on this information so when i give you this kind of information that's the output that i want when i give this kind of information that's the output that i want when i give you this kind of information that's the output i don't want that will also help it a lot follow and be going back to what tom said multiple times and i agree consistent with it with its output okay so like here's like another step that I do, like, so once you hit and start writing the article, this is the step that we are gathering more data.

38:59So on one side, on the top, like we're gathering data about the project. So for example, about the customer that we're working for, because we want to understand exactly what kind of company we are working with. So AI knows who is it working for. What is the brand voice that they have to use? What is the data from Google Search Console? This is the data about SEO to define all the things for AI to know who is it working for and why it's important. When you're working with a human, the human gets better with every interaction. When you're working with AI, especially through API, unfortunately, every single discussion starts from scratch, or almost from scratch.

39:31There's ways to go around it, but still, you have to define all the things so AI knows it will not read your mind. Research agent. You can see when collecting the data, we're having an AI research agent. What it does, basically, we provide it with the title, description, and the outline of the article that we're writing. This agent has access to multiple tools to be able to gather the data that we need for the research. But we don't know what kind of data it needs. So basically, we are providing all the information and outline in the brief to the agent saying, look, agent, I need data that will make my article valuable.

40:06Can you help me out to find things? One tool that we have, since we scrape the content of our customers from the page, we make a vector database out of it. We can pull case studies, data, information about the product, anything from the website, but we don't know what we need for this specific outline. So the agent actually defines like, okay, I need to find case studies on our website and pull this data from it, then synthesize like only what are the results and I need it. It can also go with the finding statistics, trends from other websites, not only like internal websites. And basically based on the outline, it decides, okay, I need a case study here.

40:44I need some statistics. I need quotes on this topic and it executes. So this is something that we can't predict, basically, in the beginning because we write thousands of articles in our system. But the NAA agent can actually solve our problem because it decides what kind of data we need. And this is how we use it. So as an example, like to create a draft brand voice, you can also use AI. So you can grab a piece of content and ask AI, look, this is my content that I wrote using my hands. can you help me out to understand how do I sound? Like, can you help me out to write this specific brand voice?

41:21So like, this is an example, what you can do with AI. So if you don't know how to solve it, but you more or less know where to take it from, you can leave it to AI to help you out with that. Okay, next step, AI writing draft. So since we had everything, like all the data, we had the brief, we have the outline, we were able to write an initial article writing, like we wrote a draft, I would say. And this is very important because that brings us closer to writing the whole article. The problem with having a draft is the next step. We think that normally as a human, and I told you earlier that we went through this process as humans exactly the same way.

41:57I would look at the article and I would say like, okay, I want to edit it because now when I wrote the article, I finally see that I'm missing this or missing that. I want to improve this or that. So this is a second agent that we have. We call it AI evil and refine agent. And this is super important from your perspective. This was one of the things that we wrote in the webinar invitation. This is the critique stage that we use. So you can use critique stage with whatever you do with any kind of process, how it works within our system. So we get the draft and we ask AI agent to evaluate our draft and give us an output to see like, what are the strengths of the article?

42:38So list them out. What are the areas of improvement? And what are the suggestions, actionable suggestions, like how to fix them? And we put it into the system. So based on that, we're able to create fed on prompts, how to improve it. and we can rewrite the draft based on the things. So this is an example of critic stage prompt. So as you can see, you define again who you are, you start from that, and then you define what are the output that you're counting on. Of course, you can go deeper into that, you can have a JSON file out of that to be able to then switch it to turn it into a prompt. On top of that, what we're doing, because we finally see how all the data that we gathered, how it looks in the draft, and we can see like, okay, okay, I'm actually missing a case study here.

43:25I would like to get more information at the moment because I'm closer to the final product. I can see far more. It's not like at the level of the brief, we were predicting what we will see. At the level of writing a draft, we see far more. This is like a bigger picture. So we can actually see like, okay, I'm still missing this data. I need this or that. You can turn all these suggestions and you can also like define, for example, prompts for perplexity within AI and then execute it. you can do it actually within a workflow. As I said, turn all the suggestions into prompts and rewrite, and go back to the AI workflow.

44:01So we had an agent that made decisions, did all the fancy stuff with creative part, but then we're going back to the workflow to have more predictability. What we do on the end, we add internal links. We don't use AI for that, because in our opinion, AI is unpredictable. We prefer old-fashioned machine learning. and basically we have a final product as an article that we again use AI with a human touch to edit. Like we call, of course, we can use AI to rewrite certain things. We can add value because we finally see the article, but we are able to edit it there as a human and make it even better.

44:36So this is like where our human intervention. Key takeaways. It's not a one-stop option. I want to pause you just for one second, touching on the critique side. And I think it's very, very critical. And I think it's very easy to do. And I think most people don't do it. like even if you build basic automations with with custom gpt's or with make and zapier and and so on you can add these steps where one ai will critique the work of another ai it could be the same ai it doesn't have to be claude critiquing chat gpt like it could be another step with the same tool just with a different prompt looking for specific things and as tom said you can turn the output of that to the input of the next step right so then you're creating a quasi-dynamic system where one step actually leads to what is going to be the next thing that happens, whether it's additional research, whether it's finding better keywords, whether it is, and again, now let me generalize this for any task that you're doing because it's not just SEO tasks.

45:32You can have AI critique the previous steps that were done either by humans or by other AIs, giving suggestions of what might be the next steps. And the last thing that Tom did, which is let an actual human view the process and give their feedback and input and so on is very, very important, especially in the first X number of times you're running this. And then you can finesse the process and maybe have less. And it really depends on what the process is. Sometimes you don't need any AI supervision after you tested it X number of times and it's okay. And sometimes you still do depending on what's the use case.

46:02I will actually add a couple of things to the critique stage. Like you can also use critique stage in the design part. Like, so you don't have to actually create a workflow with a critique stage, but when you're designing something, you can actually ask AI, Like when you're creating a new prompt, for example, you can actually say the same thing. Like what is the best thing in my prompt? What are the things that I need to improve? What are the action points like that you see? Or what are the edge cases? So we can use the critique stage in the design process. And then you can actually use the critique stage as a normal part of your every single workflow, like which you turn the output to a prompt in the next step of the process.

46:36Another thing that you can do with critique stage, you can actually create two, three, four versions of certain things. And you can use AI to critique and tell you like, okay, I like this the most because of this and that. And you can actually do it this way. So instead of critiquing one solution that you have and then optimizing this solution, you can give it four. And you can ask it to choose the best one and tell you why. So there's like multiple ways of using critique stage. so takeaways it's not a one-stop shop so like you have to actually if you want to eat an elephant you have to eat them in pieces like basically it's multi-step process you can combine them together depends on what you want to do and how you want to achieve start small grow it long bigger process is no different than the one that you would do as a human so like if you expect ai to do it like far, not faster, like in less steps than you would do as a human, I would try to avoid.

47:36I'm not saying that it will not do it. I would try to avoid it probably because in most cases it will not work. And context is the key. So like telling exactly AI how, what to do in specific things and what kind of things they can use. So in our writing process, we start with about 300 ,000 words to evaluate. and then we pull data together to give specific context to AI at specific stages so it knows exactly what to do that. For this heading, they can use this statistic. For this, they can use this case study and we're providing it with knowledge so it will not hallucinate all over the place. And that would be all.

48:16Fantastic stuff. I want to say a few things that are related to that. First of all, a quick summary. I think what you're structured is very, very important. You've got to start with understanding where you're going, right? Like what is it that you're trying to achieve? And you've got to address it as if you're going to give humans the work. And not just humans, humans who are new interns who know nothing about your industry, your business, your company, your processes, your past successes and failures, it knows nothing about your business. So if you bring an intern into your company and you would want to have them do something, you will spend two, three hours with them giving them all the relevant background information or they will most likely fail in their task.

48:52And it's the same exact thing here. Like you need to invest the time to figure out what it needs to know, context, and how it needs to get it as far as what type of data in order to do the task in the most effective way. And then I think the next thing that you said that is very, very important, you have to find the right combination between structured process and AI doing the thing that AI is good at, which is thinking, which is analyzing, which is categorizing, which is coming up with something new based on something existing. all these things AI is very, very good at. But if you let AI run freely across all of it, you may diverge very, very far from the original path you were planning.

49:35So this approach that Tom just shared with us is the right approach where you use a structured process to go through the process and you give AI the specific tasks where it's going to excel and it's going to do very, very well. And so I think it's a brilliant, brilliant approach. There was a question very early on in the beginning, and it actually is interesting because it came almost exactly at the same time on both Zoom and LinkedIn, which is about Manus, you know, the new Chinese tool and kind of like your thoughts about it. I know what my thoughts are, but what do you think about these dual kind of agents?

50:11I think they're good to play around with this. Like, in my opinion, it's not predictable enough. like my biggest problem, like with AI generally, like it doesn't have to be about manners. It's predictability of results. So if I have this, if I ask the same question and I normally ask it like 150 times the same question to AI to see like, what is the predictability? More or less, like when I'm talking to a human, I want to have a similar answer. So I don't want it to be all over the place because that means it does, the specialist that I'm talking to doesn't know what they're doing. So it's the same with agents.

50:44Like if we can't get the predictability, that means they're just like providing it with providing us with an answer that we want to hear and yeah no i i'm with you right and i think it goes back to what you said before there are cases where this is going to be great and there are cases this is going to be horrible and you just got to find the right use cases for it i'm not saying it's good or bad i'm saying like everything with ai you got to find the right use cases and find the right tools to apply to these use cases when you automate predictability is the key yeah yeah tom this was really really fantastic you shared a lot of thoughts and processes and concepts that people need to know in this stage of like okay we want to start implementing ai i want to start applying agents and so on i'm sure a lot of people find a lot value in that if people want to follow you work with you connect with you like what are the best ways to do that i basically live on linkedin this is my second home so if you want to find me and go to LinkedIn and just invite me to your friends, like direct message me.

51:41I love to talk to people. If you have some problems that you want to talk about, I really like to share knowledge and get also knowledge from others, like learn it. And I think we definitely can do a lot of that. So find me on LinkedIn if you want. And if you want to find like the product that I'm like doing is seowind.io and you can find it on the website. Awesome. Thanks everybody for joining us live. We had a lot of people on LinkedIn and a lot of people on Zoom. So I appreciate you spending this past hour with us. If you are not here and you're just listening to this to your car, also thank you for that.

52:11But come join us next Thursday. We do this every single Thursday at noon Eastern with somebody like Tom sharing a lot of knowledge about how to do things correctly with AI. And obviously, thank you, Tom. This was absolutely fantastic, well-prepared, very deep, easy to follow. I really, really appreciate you sharing all your knowledge with us. Thanks a lot and happy birthday. Thank you. Bye, everyone.

52:34No way to do that